The mechanics worked surprisingly well. The agents could research in parallel, challenge assumptions, compare conflicting evidence and carry a strategy through into operating plans and actual management documents. What took more work was deciding how that reasoning should behave once it started interacting with decisions already made.

We ran into that problem fairly early.

I had given the system a strategic mandate that contained a mixture of decisions, hypotheses and open questions. I had already chosen a direction for the business. I still wanted the agents to challenge whether we could execute it, whether the economics held up, what risks I might be underestimating and what evidence would cause us to change the way we approached it.

During one run, the analysis gradually wandered outside those boundaries.

Nothing obviously broke. The agents found legitimate evidence in the existing business. Current customers, current revenue, current capabilities and current constraints were all easier to substantiate than a future strategy that had not yet been executed. As those facts accumulated, they started carrying more weight in the analysis. One specialist reinforced another, the thesis shifted, later agents inherited the shifted thesis, and eventually the system produced a recommendation that was quite defensible on its own terms.

It was also answering a question I had not asked.

That was useful because we could trace how it happened. The original mandate did not make the status of every major assumption clear enough. Current-state evidence influenced something I had intended to treat as a strategic commitment. Once that happened, every subsequent agent was reasoning from a slightly different starting point.

We corrected the architecture rather than simply correcting the recommendation.

Every important statement in the engagement began carrying a type. A decision was marked DECIDED. A hard boundary became a CONSTRAINT. Something we believed but needed to test became a HYPOTHESIS. We also separated PREFERENCES from UNKNOWNs.

The labels mattered because they changed what an agent could do.

A hypothesis should be challenged. An unknown should be investigated. A preference can be weighed against evidence. A decision can also encounter contradictory evidence, but the contradiction has to come back to the person who owns the decision. The system can make the conflict visible and argue strongly that the decision deserves another look. It cannot quietly behave as if the decision disappeared.

Once we started doing that, I noticed how much of management operates through distinctions like these without writing them down.

Someone in a leadership meeting knows that one comment from the CEO is a suggestion while another is a decision. An experienced executive can hear new evidence and understand that it affects the implementation plan without necessarily changing the underlying strategy. People who have been around the company for years know which customer request is an interesting data point and which one is evidence of a broader market pattern.

AI agents do not inherit any of that simply because they have access to the documents.

We had another version of the problem with context. One recent customer conversation became disproportionately influential because it was specific, current and easy for the system to reason about. We eventually added a relevance test around context: does this information materially affect the strategic question, or is it receiving attention because it happens to be recent and vivid?

A vendor created a similar issue. The system started treating the company supplying a particular capability as strategically important. The strategic dependence was actually on the capability. The vendor was just one way to obtain it. That distinction became another rule in the system.

The process accumulated rules this way.

We stopped relying on the chat transcript to tell us the current strategy and created a persistent state containing the thesis, assumptions, evidence, contradictions, decisions, and unanswered questions. We had to write material changes into that state. Evidence was tracked back to its underlying source rather than to whichever agent happened to retrieve it. When two sources disagreed, the contradiction stayed visible until it was resolved.

The specialist agents also began challenging one another. Market work could question customer evidence. Finance could challenge the hiring assumptions. Product could ask whether the commercial promise could actually be delivered. The operating model had to account for the resources implied by the strategy.

That interaction proved more useful than simply asking more agents for opinions. Fifteen independent reports can still produce fifteen versions of the same assumption. Forcing the assumptions into contact with one another exposed where different parts of the strategy depended on incompatible beliefs.

The sponsor gates added another layer. At several stages, the system stopped and showed me what it believed, what evidence had changed, where confidence remained weak and which decision it needed from me before proceeding.

I originally designed those gates to keep authority with the human. They also created a useful record of my own reasoning.

Founders supply a lot of context to a strategy exercise, and we are not objective sources. We have history with the business, preferences about where it should go and a tendency to speak confidently about things we believe to be true. In this system, something I told the agents could be recorded as sponsor evidence without automatically becoming an established fact. That feels like a small distinction until the system begins comparing your conviction against documents, customer evidence and actual operating results.

At times, I corrected the agents. Other times, the structure forced me to be clearer about what I was actually claiming.

By the end of the process, the same discipline had moved into execution. We had tripwires tied to dates and owners. We translated strategy into questions for each department. We produced different documents for the board, management team, and people responsible for execution.

Then we had to deal with the fact that AI makes changing all of those documents extremely easy.

A strategy can be approved at noon and rewritten into twenty slightly different versions by dinner. We eventually stopped allowing approved decisions to be casually regenerated. A subsequent change became an amendment with the original position, the revised position, the reason and the documents affected by it. The historical decision stayed in the record.

That is a practical problem I had not thought much about before doing this. When producing another version becomes almost free, version discipline matters more.

The technology underneath these systems will keep changing. I do not know which agent framework or model we will be using a few years from now, and I would be hesitant to design much of the company around any one of them.

The operating questions feel more durable.

You need to know which information is evidence and which is assumption. You need a record of the decisions the organization has actually made. You need to know who can reopen those decisions, what happens when contradictory evidence appears and how a change moves through the rest of the organization. You need a way to stop recent context from becoming strategy simply because it is the easiest information for a model to see.

None of those are especially new management problems.

The difference is that an agent system can now move through them with enough speed and autonomy that informal rules are no longer sufficient. What used to live in the heads of a few experienced people has to become explicit enough for the system to follow.

That is where I ended up spending most of the three days: writing down rules I had never realized we were relying on.

Two questions, drawn as a tree. Is it your core differentiator? Build it. If not, is it critical when it fails? Buy it. If it isn’t core and it isn’t critical, go ahead and build it; have fun. The line underneath it: AI made it possible to build anything. That doesn’t mean you should. And this was the theme of the week. Both in the upcoming episode of my podcast (the Signal - check it out every week on Tuesday at 9 am.) - ok, no more plugs.

I should disclose something before I go further. I sell software to lenders. Many of you already know this, but if you're a first-time reader, that is my business. So when a framework says “buy the critical things you don’t differentiate on,” I am exactly the guy who benefits from you believing it, and you should read the rest of this with that in your pocket. Hold me to the tree, not to the conclusion.

Because here is what happened when I stopped nodding. I ran the tree against the last twenty years of my actual working life, most of it inside institutions with regulators attached, and it kept giving me answers I couldn’t defend. Not wrong answers. Incomplete ones.

The tree is a startup tree. It is correct as a startup tree. In a bank, it stops one node too early.

THE DIAGNOSIS

Start with what the comment section proved, because it is the most interesting thing about that post and the author didn’t write it.

I had AI count: twenty-four separate people, out of roughly sixty who commented, independently wrote the same sentence in different words. One said it cleanest: AI has lowered the cost of building; it hasn’t lowered the cost of ownership. Another said it sharpest: AI made writing code cheaper; it didn’t make understanding old code any cheaper. A third: the cost curve shifted, the ownership curve didn’t.

That is not a hot take. That is a field observation with a sample size, and it arrived in a comment section instead of a research paper, which is where most of the useful data about this era is currently sitting.

Why this matters is that every IT team and software vendor worries about the same thing. Let’s play this out with a person named Samantha. She’s been at your company for a decade. In fact, she built half the system, then left, and nobody knows how it works.

What left with Samantha was not the code. The code is still sitting right there, readable, running. What left was the ability to explain why it works that way. Nobody wrote that down.

Hold that thought, because it is the whole article.

Ownership isn’t who wrote it. It’s who can explain it.

Now put that sentence inside a federally regulated institution and watch what happens to the tree.

THE MISSING BRANCH

Three things are true in a Canadian FI that the tree doesn’t account for. None of them are my opinion. OSFI has already written all three down.

One. Buying doesn’t move the risk. It moves the code.

Guideline B-10 covers third-party arrangements, all of them, not just the ones you call outsourcing. The transition ended May 1, 2024. The operative sentence is one line long:

The FRFI is ultimately accountable for managing the risks arising from all types of third-party arrangements.

So the “Buy” leaf on that tree is not an exit door. You still own the risk. You have relocated the source code and kept the accountability, which is a real trade with real advantages, and it is not the same trade as making the problem go away. In some ways the decision is closer to build vs rent. Rent has its own risk profile, its own pricing power problem, its own roadmap dependency. I sell rent. I’d still rather you knew that’s what you’re buying.

Two. “Critical if it fails” is not a gut call in a bank. It’s a filled-out form.

E-21 requires institutions to identify their critical operations and set tolerances for how long each one can be disrupted. Sections 1 and 2 have been in force since August 2024. Full adherence and operationalization is expected September 1, 2026, which is two and a half weeks from now.

That means the second node of that tree already has an answer sheet inside every FRFI in the country, and if the E-21 work was done properly, it is already filled in. Most operators I talk to have not connected those two things. The resilience binder and the build-versus-buy conversation happen in different rooms, run by different people, on different floors. Same problem as last week’s letter, different pillar.

Three. If it’s model-shaped, it’s in scope no matter who built it.

E-23 applies to all types of models regardless of the technology used, explicitly including AI and machine learning. Released September 2025, effective May 2027 after an eighteen-month transition. A weekend build does not escape it. Neither does a vendor’s. Nobody gets to argue their thing isn’t really a model because it was assembled from a chat window.

Put those three together, and the tree needs a third question. Here it is.

Is it your core differentiator? Build.

Not core, critical when it fails? Buy.

Not core, not critical? Can a second person explain how this works to an examiner in 2029?

Yes: build it.

No: you are not choosing between build and buy. You are choosing whether to create an undocumented dependency inside a regulated institution.

I’d call that last one the examiner test, and I picked it deliberately over the two obvious alternatives. Back to the LinkedIn post, one commenter wanted a node about complexity and maintenance load, which is reasonable. Somebody else wanted one about security exposure. Both are downstream of the same thing. If you can explain a system to a hostile third party five years from now, you already have the documentation, the boundaries, and the second owner. If you can’t, you don’t have a system. You have a person, and people leave.

Samatha leaves. The examiner still shows up.

THE COUNTERARGUMENT, GIVEN ITS BEST SHOT

The strongest pushback in that thread came from two people who build for a living, and they are not wrong, so I’m not going to pretend they are.

Their argument: why does everyone assume the cost of operating isn’t falling on the same curve as the cost of building? AI compresses maintenance too. It reads the code, writes the tests, generates the docs, handles the dependency bumps. For a non-core process where a vendor charges per seat plus an AI premium on top, a small internal team can now genuinely beat a multi-year commitment. The break-even moved.

That’s true. It has moved, and it will keep moving, and a blanket “buy your non-core critical systems” is lazier than the framework it claims to defend. One more version deserves a line: at small enough scale the right answer is building something small enough to throw away, so the Samatha problem can never reach you in the first place.

Here is where I’d push back.

In a regulated institution, the binding constraint was never maintenance labour. AI does compress that, genuinely and permanently. The binding constraint is accountability and evidence, and B-10 does not get cheaper because your agent writes better code. The examiner still needs a human who can explain the decision. E-23 still applies to the thing you generated. Nobody has automated the person who has to sit in the room and answer for it, and I don’t think anybody is going to.

Compliance is the shape of the product, not an obstacle in front of it. That is the part of the cost curve that hasn’t moved.

THE NUMBER

Twenty-four of roughly sixty commenters, independently, arriving at one sentence: building got cheaper, owning didn’t. Without putting two and two together until I wrote this article @ and I talked about it, so it’s a soft number and I’ll call it that. But the count isn’t really the point.

Here is the point, until that framework and discussions points to regulated industries the entire conversation that’s happening only talks about the cost of ownership as engineering time. In banking, insurance, defence and healthcare it isn’t. It’s the second cost, the one that shows up years later when somebody who doesn’t work for you asks how the thing works and won’t accept “Samatha built it” as an answer.

That gap between how the tech industry reasons about ownership and how a regulated institution has to is, more or less, why this letter exists.

MONDAY MORNING

Take one internally built system, the one nobody wants to touch. Not the whole inventory. One.

Ask three questions about it, which I’ve adapted from that comment thread in that post last week:

Can we state the boundary of this system in one paragraph? Do we know the failing case that will send us back into it? Is there a second person who could change it next quarter?

Two no’s and you don’t have a build-versus-buy decision to make. You already made it, quietly, some time ago, and you’re now maintaining a system whose real documentation is one person’s memory. Whether you replace it or write it down is a smaller question than knowing which one you’re in.

Then go find whoever owns your E-21 critical operations list and read it next to your application inventory. Those two documents are supposed to describe the same institution. In most places, nobody has ever put them on the same table.

That’s the week.

Credit where it belongs: the tree isn’t mine, and neither is the best line in this letter. I only added the branch.

Hit reply and tell me about the system nobody wants to touch. Names off, pattern on. The best replies become episodes.

— Chris

For some, AI offers a shortcut. For others, it is a tool to enhance their productivity and thinking significantly.

Much of what I observe involves using AI to rewrite or summarize content, essentially serving as a faster keyboard. Now, you can speak, and it generates output independently. However, it often lacks your context, reasoning, and the critical thinking needed to achieve substantial benefits as a thought partner.

If you use AI thoughtfully by asking targeted questions and encouraging it to challenge you, you achieve better results and enhance your own thinking. Ultimately, the decision must remain yours. Once you delegate that responsibility, outcomes become unpredictable.

Subscribe now

Speed changes what you are willing to test

The reality is that AI increases our speed. While it can be used as a shortcut, it also enables faster decision-making.

I have found that I can experiment more, as outcomes that once took three weeks now take only ten minutes. This efficiency encourages me to try more ideas simultaneously.

The website project is the clearest example.

The first website I ever built for one of my businesses was back in 2003. We had a very talented designer. We sat in her office for about eight hours with a whiteboard and all these ideas, and then we were given homework, pages of it, building content and everything else. That was a $30,000 or $40,000 website.

At FundMore, our website had not been updated since the company’s inception. We rebuilt it in a week, though the actual work took only about eight hours, including content and review. Previously, this process involved multiple handoffs and revisions, making it a project that spanned several weeks.

I provided Claude with our requirements, reference sites, and brand guidelines. Within ten minutes, it generated three mockups, allowing me to select a direction. After some iteration, my team and I developed the content and integrated it into the site.

The site is now dynamic, featuring interactive tools, calculators, infographics, and animated videos—elements that previously required weeks of design work.

The site went live yesterday, fundmore.ai, I welcome feedback, as rapid iteration is now possible.

The mastermind

Last March, my co-host on the Signal podcast (every Tuesday morning at 9 am) and I discussed Cowork’s capabilities and arrived at the same idea: Could we build a team of specialized agents to support decision-making?

Napoleon Hill’s Think and Grow Rich was one of the first business books I read. While it outlines thirteen principles, the concept of the mastermind stands out: surrounding yourself with intelligent people leads to better collective decisions, regardless of individual intelligence. He called the Mastermind the coordination of knowledge and effort, in a spirit of harmony, between two or more people, for the attainment of a definite purpose. Coordination, harmony, and a definite purpose. Not a group chat.

This idea stayed with me. Two months ago, I set out to implement it. Since Cowork was not ready, I used Claude Code, marking my first experience with coding.

I began by defining a core executive team: finance, people, product, marketing, and sales leaders, among others, creating nine distinct roles.

I reviewed lists of business books in my Audible account, extracted key ideas, and organized them in Notion. Each role is thus grounded in the leading thinking of its discipline, based on books, papers, and strategy rather than job titles.

I then created a researcher agent. It searches the web when the core team lacks an answer and accesses all of FundMore’s knowledge, now stored in Notion.

How a decision runs through it

For major decisions, I prepare a brief outlining the decision at hand, my current perspective, and the primary challenge or key question under consideration.

That goes to a strategy agent, who reviews it and distributes it to the nine roles for analysis and feedback. The strategy agent decides which three or four seats this really impacts, and it will sideline an advisor if it does not believe that seat will add value. Then it red teams it. You end up with a red team and a green team debating the answer.

All of that gets packaged up and passed to a coach who reviews the feedback from the strategy agent. This process takes fifteen to twenty minutes and is reserved for significant, often irreversible decisions, rather than routine ones. I wanted it to take all the knowledge and the context and surface it back to me as a question. Did you forget about this? Think about the decision this way. Here is some advice around it.

The goal was not to create an echo chamber, but rather a system that identifies gaps in my reasoning and prompts me to evaluate them.

If you remember one point, let it be this: the ‘mirror problem,’ or sycophancy, is well-documented. Anthropic’s 2023 research found that leading AI assistants often prioritize agreeable responses over accurate ones, reflecting human preferences. Unless you intentionally design against this, AI will default to mirroring your views.

While we all have advisory and peer groups, they are not always accessible when urgent decisions arise.

Where to start

Begin with a clear prompt, structured in four simple parts.

State your objective, request that it challenge your thinking, ask for the strongest counterargument, and inquire about any gaps in your reasoning. Specify that it should only disagree when supported by evidence.

Apply these four components to a current decision. The AI will respond with questions or new perspectives you may not have considered.

You have to be the thinker.

Ultimately, you must remain the decision-maker. Use AI to challenge your thinking, not to make decisions for you.

These tools yield very different outcomes depending on how you use them, and true leverage comes from thoughtful engagement.

Preliminary research from MIT Media Lab found that students using AI assistants showed lower brain engagement. While the study focused on essay writing, not executive decisions, it highlights a key risk: the danger lies not in the tool, but in passive use.

Regardless of your experience level, success begins with understanding your context and objectives. Do not rely on AI to make the final judgment; that distinction is crucial.

Last Thursday, Questrade announced Equity Engine: “Automated set-up and management of the Smith Manoeuvre strategy, letting clients turn home equity into a tax-advantaged investment portfolio without tedious management oversight.”

The strategy was never mine. Fraser Smith built it in the early 1980s, and I was one of a handful of brokers explaining it to clients who’d never heard of it. What I didn’t expect was that eighteen years later, the first firm to actually productize it wouldn’t be a bank.

And the part that got my attention isn’t the tax strategy. It’s what had to happen inside that company for the feature to exist at all.

THE DIAGNOSIS

For decades, Canadian banking was built on four pillars: banks took deposits and made loans, including mortgages; trust companies administered trusts and estates while also competing in deposits and mortgages; insurance companies underwrote life, health, and property risks; and securities firms raised capital, advised investors, and traded stocks and bonds. They all played in their own lane for the protection of Canadians; in fact, it was the law. Until Ontario’s amendments took effect June 1, 1987, and within a year Scotiabank had bought McLeod Young Weir. The legal wall came down thirty-nine years ago. BUT… the org chart wall never did.

The law generally prevented a company in one category from owning a company in another. That is why they were called the “four pillars.

Ontario changed its rules on June 1, 1987, allowing those businesses to begin combining. Soon afterward, Scotiabank bought the investment dealer McLeod Young Weir—an early example of a bank crossing into the securities business.

Walk into any Canadian FI today, and you’ll find the modern version of the same four rooms: Wealth & Asset Management, Retail, Investment Banking, Commercial & Corporate. Separate P&Ls, separate systems, separate leadership, separate incentive plans. Nobody is legally stopped from crossing. Everybody is structurally discouraged from it.

I’m not reading that off an org chart. I worked at three of the five, and the thing that stays with me isn’t that the divisions competed; it’s how normal it felt that they didn’t talk. A client sitting in front of me had a mortgage on one side of the house and investments on the other, and the two sides had no idea the other existed. Referring across wasn’t forbidden. It just wasn’t anybody’s job, and nobody’s number moved when you did it.

The Smith Manoeuvre is the cleanest possible test of that wall, because it cannot be executed by one room.

Here is the mechanic, and it hasn’t changed since I wrote it up in 2008. You need a readvanceable mortgage, a split facility where the credit line grows dollar for dollar as you pay the mortgage principal down. Each month you re-borrow exactly the principal you just paid, invest it, and deduct that interest under Income Tax Act s.20(1)(c). Your bad debt converts to good debt, one payment at a time, and the refund gets reinvested.

Read the requirement list again. Lending owns the readvanceable facility. Wealth owns the investment account. Tax reporting sits somewhere else entirely. No single P&L owner can ship it, so no single P&L owner proposes it, so it doesn’t get built. It isn’t a technology gap. It’s an ownership gap.

And here’s what makes the forty-four years indefensible: the plumbing was never the problem. The HELOC showed up in Canada in the late 1970s. Manulife One launched in 1999 as the country’s first all-in-one account. Scotia made STEP automatically readvanceable in 2009, meaning pay down a hundred dollars of principal and the available credit goes up by a hundred dollars, no phone call. Seventeen years ago, the lending half of this strategy became fully automatic.

Then look at the scale. By 2016, the FCAC counted roughly three million HELOC accounts in Canada, about 80% of them held under readvanceable mortgages. Call it 2.4 million households already holding the exact facility this strategy requires, most of them with no idea what it can do. The product was built, sold, and distributed at national scale. Nobody built the connecting tissue, or the marketing strategy was always about fixing up your house, buying a new car, buying a vacation property, debt consolidation, or simply having access to emergency funds. Nobody talked about leveraging it for tax advantages.

Now here’s the part that matters more than the tax angle, and it’s the thing I’d want on a whiteboard if I ran product at a Big Five.

Equity Engine is not an investing product. It’s an audit-trail product wearing an investing product’s clothes.

The Smith Manoeuvre doesn’t usually fail on market risk; it fails because when CRA audits interest deductions, the classic kill shot is a contaminated line: put one personal expense on the investment HELOC and the deduction on the whole thing is at risk. Executing this by hand, correctly, every month, for twenty years, with records that survive a review a decade later, is the actual difficulty. That’s why the strategy stayed the property of disciplined DIY investors and a small network of certified advisors instead of becoming a product.

Which means the thing Questrade automated is not “invest my equity.” It’s “re-borrow only the principal paydown, never touch the line for anything else, and generate a record I can hand to an auditor.”

That is the identical claim we make about agentic lending operations, and it’s the one most people get backwards. The automation is not valuable because it’s faster. It’s valuable because it makes the paper defensible. Speed is what you sell. Defensibility and compliance are what you’re actually buying.

One honest caveat, because credibility down here is the only currency. The wire release says clients can “instruct AI platforms to initiate an order on their behalf.” The Globe, reporting on the same launch, describes it differently: investors “can also draft trades through the AI assistant before approving them through the Questrade app.” Draft, then approve. That’s a human gate, and it’s materially narrower than the marketing. I don’t read that as a gotcha. I read it as the correct answer. It’s also, almost word for word, what OSFI’s July 10 agentic bulletin asked for: control agent autonomy, put approval checkpoints on high-risk actions, treat AI output as input to a decision rather than the decision. And the broader operational resilience guideline all of that sits inside, E-21, hits the date OSFI set for full adherence and operationalization on September 1, 2026. That’s three and a half weeks from now. The version that ships in regulated finance is always the narrower one, and the firms that pretend otherwise in a press release get to explain the gap later.

Subscribe now

THE NUMBER

44.

Fraser Smith built this in the early 1980s. His son Robinson took over the practice in 2006 and still runs it. Early 1980s to 2026 is roughly forty-four years.

Forty-four years of a legal, CRA-recognized, entirely mainstream strategy. Six banks with every one of the required pieces already in inventory: the readvanceable mortgage, the brokerage account, the tax slips, the client. Thirty-nine of those years with no legal barrier between the divisions that own them.

Nobody shipped it. Not because it was hard. Because it belonged to two rooms and lived in neither.

If you want a single sentence for why incumbents lose to smaller firms in AI, it isn’t model access, data, or talent. It’s that the smaller firm has one room.

THE PLAYBOOK

(If you read my article in 2008, it’s effectively the same but with AI)

Questrade is arguably the second Canadian firm to organize around the customer rather than the pillar, Wealthsimple being the other. Neither is a Big Five. That’s not a coincidence, and it’s also not destiny. Here’s what this specific feature actually requires, which is the useful part if you’re sitting inside an incumbent:

One owner across two P&Ls. Not a committee, not a working group. A single accountable owner with budget authority spanning lending and wealth. If you can’t name that person, you can’t build this, and every subsequent step is theatre.

Agents that read state, not just text. The monthly cycle is: payment posts, principal component identified, available credit readvanced by exactly that amount, transfer to the non-registered account, invest per mandate, log everything. That’s an agent reading ledger state across two systems and taking a bounded action, not a chatbot answering questions. Most FI “AI roadmaps” today are still the second thing.

The human gate on the money-moving step. Draft, then approve. Not because the model can’t do it, but because that’s the design OSFI has now told you it expects, and because the first mis-traced dollar is a client’s whole deduction.

The audit file as a first-class output. Every readvance, every transfer, every purchase, timestamped and attributable, exportable in a form that survives a CRA review in 2036. If your automation can’t produce that, you’ve built a convenience feature and priced it like one.

And the open question yet to be answered. Whose readvanceable mortgage is this running on? From my research, Community Trust, Questrade’s lending subsidiary, does conventional firsts and seconds and does not offer a readvanceable product or a HELOC at this time. So either Equity Engine automates only the investing leg while sitting on top of a competitor’s mortgage, in which case Questrade is monetizing investment flow off Scotia’s or Manulife’s balance sheet, or Questbank is about to launch a readvanceable mortgage to feed it.

The second reading is a much bigger story than a feature bullet. It’s a new bank designing a mortgage product whose primary purpose is to feed an agentic acquisition engine. If you originate, that’s the version to plan for.

Either way, here’s the part the lending side keeps missing: consumer agents create lender-side volume. An engine re-borrowing monthly, forever, is a recurring credit event. An agent scanning for “opportunities to switch mortgage or insurance providers,” which is literally what Questrade’s Personal Balance Sheet is described as doing, is a switch-request generator. Somebody’s back office absorbs all of that. Automation on the consumer side is a demand shock aimed directly at operations that never automated.

Four Canadian FIs have moved on agents in ten weeks. The cadence is the story, and I don’t think most operators have clocked it yet.

If you’re inside a bank and you’ve watched a good cross-pillar idea die in the ownership gap, hit reply and tell me how it died.

— Chris

P.S. I still have the 2008 piece, MS Publisher layout, stock photo of a family on a lawn, 7% rates. The rates changed. The mechanic didn’t. That’s the tell that this was never a rate story or a market story. It was an execution story the whole time, and execution is exactly what just got cheap.

THE DIAGNOSIS

Most people use AI reactively. Rewrite this email. Build this deck. Summarize this document. All fine, and your inbox is still managing you. You wake up to thirty messages and spend the day reacting. The tool saves you an hour, and the hour disappears into more email and more meetings.

Saving an hour is a refund. Leverage is reinvesting it.

Reuven found his version of this the hard way. He connected an LLM to his email and calendar and asked it to be ruthlessly honest about how his time was being spent. The answer: ten to twelve hours a week on customer escalations, FYI meetings, and decisions that didn’t need him in the room. He wasn’t blind because he’s careless; he was blind because that pattern is invisible from inside the calendar. The audit made it visible. Then he built the triage plan and delegated the pattern, not the tasks.

Mine took longer. Eighteen months ago, I tried the scheduling fix everyone recommends: push all meetings to the end of the day and protect the morning. The day filled up anyway. Scheduling rules don’t fix a leverage problem. What fixed it was systemic: seven months ago we decided to become an actual AI-first company. Every week, every leader answers seven questions: what worked, what didn’t, how they used AI with a pre-built prompt so it takes minutes, not an evening. All of it, plus our call transcripts, email, CRM, and coding tools, feeds one knowledge base. The agents we build stand on that. The compounding showed up around month three. By month six it was the operating system.

The receipt: in January my calendar was wall-to-wall. Next week I have maybe five meetings, none longer than an hour. Same company. More customers than January. I used to call the old way management by walking around. Now it’s management by walking around, without the walking. The morning brief reads every system before I’m out of bed.

One warning, and Reuven said it best on the show: don’t outsource the thinking. The brain is a muscle, and if you hand the strategic work to the machine, it atrophies. Give it the menial layer. Keep the judgment.

THE NUMBER

40%.

Last October we hired our first AI engineer, and my big idea was to screen-record my own work, ship him the recordings, and have him figure out which parts could become agents. We spent weeks on it. The output was about 40% accurate. No lift. By any normal measure, a failed project.

It was the most useful benchmark we ran all year, because it told us exactly what to automate then and what to park. And here’s the part that matters if you’re sitting on your own failed pilot: the capability catches up faster than your roadmap. That was October. This week, the same job, turning a recording of my work into a working agent, took one evening. No engineer.

ONE THING WE SHIPPED

Claude released a screen-recorder skill: hit record, do the task, narrate if you want, and it turns what you did into a reusable skill. No engineering.

I pointed it at the task I was doing anyway, pulling together board materials. It watched me work, mapped the repetitive parts, and offered to build the skill. Monday night, after our actual board meeting, I ran it. The full board share-out came from a recording of what I actually do, the thing that took an engineer and me weeks to attempt at 40% last fall.

Reuven’s version is smaller and maybe better: he recorded himself fixing the fonts on every deck Claude produces, the tweaks he made by hand every single time, and turned it into a skill. Now the decks come out right the first time. That’s the shape to look for. Not the moonshot. The thing you repeat every morning: one instrument, not an orchestra.

That’s the week. If you missed the live, we’re back Tuesday at 9 a.m. Bring your skepticism; it’s the dress code down here.

Hit reply and tell me the workflow that’s costing you the most. The best replies become episodes.

— Chris

P.S. On the episode, we mentioned a Claude skill an executive-coach friend built. It runs your decisions through an Elon Musk lens and hands you a one-page “here’s what you should have considered.” I run most major company decisions through it now. Reuven says it punched him in the face. Reply if you want it, and we’ll send it over.

If you looked at my calendar in January, it was wall-to-wall. If you look at it today — next week I have maybe five meetings, and they’re all less than an hour. There’s a ton of empty time, which allows you to think, grow, and spend time on the business rather than in it.

Same company. Same job. Here’s how that happened.

The scheduling fix that didn’t work.

It started with a podcast, Reuven, my co-host on The Signal, shared with me about eighteen months ago. The theory: push all your meeting invites to the end of the day, and you create a morning where you have the most energy and creativity; you spend the first five or six hours just working on the business. I remember listening and thinking, that sounds awesome. So I went out and told everyone, “You want to book a meeting with me?” It’s from four to six.

It didn’t matter. The day got filled up anyway. It’s the common things, the one-on-ones, people needing information, wanting a decision made as they move through the day. The calendar was never the problem.

The decision that did

Last October, I made a real, conscious decision: to understand what’s actually operating this business.

If you’re a lender, you know that happens on the front lines. Think about a decision an underwriter makes, declining a deal. Sure, there’s policy. But much of that is inherent knowledge, and unfortunately, that knowledge dies with the decision. It doesn’t get captured anywhere. Then the next decision gets made.

So we made a real conscious effort to capture that information across the company so that it could become the knowledge base. When the AI agents we eventually built had that knowledge stack, they could actually answer the questions.

Because think about the first time you engaged with ChatGPT. You asked it to write an email; it was half right, you gave it more context, it came back, you iterated three or four times, boom, magic output. And then you started over the next time. If you don’t record it and let the system understand how you work, you’re starting over every single time.

Becoming AI-first — the unglamorous part

About seven months ago, we decided we would be a true AI-first business. It started with this: at the end of every week, I asked every leader in my company to tell me what they did. What worked, what didn’t work — I even asked how they were using AI. About seven core questions. And I gave them a prompt, so our connected tools drafted the answer for them, and they added to it.

This was not about micromanaging. I probably read one in three. It was about understanding how the company operated.

All of it went into a Notion database. Coupled with the connected tools we use every day — Granola transcripts, our emails, HubSpot, and our coding tools everything gets turned into context, and the agents can access it. That became the foundation. We started to see the compounding effect by about month three. By month six, I really think we’re an AI-first company now.

What the mornings look like now

I wake up at 6:30 every single day, make an espresso, load up my phone, and there’s this report diarizing the entire business. Not just what’s happening in my world, but my team’s world: delivery, product, finance, sales pipeline, support tickets. I get it two ways: either as a report or as an artifact with fancy graphs, charts, and outputs that I can glance at throughout the day. I don’t do a single thing to produce it.

I think back to when I worked in an office. The way a manager understood the business was the tour of the cubicles — pop your head in, “what are you working on?” — gathering information for the weekly report to their boss. This is management by walking around, without walking around.

The other big leverage point: what should I actually be working on? We’ve all read the books: Eat That Frog, tackle the biggest piece first, tackle the low-hanging fruit first. Every one of those theories required you to spend an hour, the night before or on Sunday, writing down everything you had to get done. You don’t have to do that anymore. The system creates the list for the week. You choose which one to tackle. That alone is a couple of hours a week.

Three receipts

The proposal. We were on a discovery call with a customer a couple of days ago. The call ended, and the draft proposal existed within a couple of minutes — follow-up email drafted too. Because it understands our business, there was very little refinement—a couple of tweaks — not the tweaks of six months ago.

The NDA agent. I took every NDA we’ve ever received — the redlined versions and the green-checkmark versions — and had it understand them. Now every NDA that comes in gets flagged and redlined first. There’s still a cursory read before we sign anything, but not the thirty minutes it would have taken someone on my team.

The steerco. About a week and a half ago, a product update call with a customer. The two teams agreed they needed an output and that, with both teams contributing context, it would take about a week to a week and a half. I was sitting on the sidelines. I opened Claude in real time, copied the live transcript — the call wasn’t even done — and asked it to build the output. Before the call ended, I had the whole matrix of everything both teams had to execute, because we had all the context from the months we’d been working together. Probably ten people were going to contribute to that. Instead, two people from my team reviewed it, and two or three from theirs.

It didn’t all work

Last October, we hired our first AI engineer, and my whole thought was: how do I get leverage from him building internal agents? I spent the first few weeks screen-recording what I was doing and shipping it over so he could analyze which of those jobs could actually become agents. The output was about 40% accuracy. We weren’t getting lift.

Eight months later, a screen-recorder skill comes out of Claude. Last week I recorded myself gathering board material, and it handed back the repeatable task and the option to build the skill. I ran it the night of our board meeting, and the output was incredible.

Apparently that’s everything in AI now: if you can think about it, just wait. Someone will release it.

The human side is the hard part.

AI is still scary for a lot of people. There’s a large unknown: everyone logged into ChatGPT at some point and asked it to write an email, and then they saw the transformation: “holy cow, it can do 80% of my job.” You can’t force this on people. You have to bring them along.

We worked top-down. We didn’t hand out licenses to everyone on day one; the management team got them first. At the beginning, I did show-and-tells on my own weekly management call: here’s how I’m using it. Then it evolved into a weekly rotation in which someone would demo what they built using AI.

Most people have a bit of competitiveness in them. I go back to my coaching days: Anson Dorrance, the UNC and US women’s national team coach, always said people thrive in competitive environments; you have to put them up against their peers, make it transparent. And honestly, those demos got better and better every week. After a couple of rounds, I stopped running it. I didn’t need to anymore.

One thing I still do is ask a question when I feel like we did something manually. Could AI have done that? Did we ask AI to do that? Not as a gotcha, but because I wanted people to understand it’s not scary, and it makes the job easier. The message to my team is simple: I want you on high-value work. I don’t want you running nine-to-five and then thinking you have to work three more hours at night. I know I can get 10x the productivity out of you in less time, and maybe one day we even get to that work-life balance everyone dreams of in startups.

The point

Capture the knowledge before it dies with the decision. Connect the tools. Let the system learn how the business actually operates. It took us seven months, and the compounding effect is the whole game.

The empty calendar isn’t the goal. It’s the proof.

Then they ran a previously approved file through it, and the model flagged the deal as an exception.

There was just one problem: it wasn’t an exception, and everyone in the room knew it. The file was structured differently from what the written policy anticipated, but an experienced underwriter understood immediately why it should be approved and how it was likely to perform. The model had read the policy. It had never met the business.

Nobody in that room questioned the intelligence. The intelligence was fine. The intelligence was spectacular. What it lacked was the institutional context that had never been written down because until now, it had lived entirely inside someone’s head.

That gap has a name now. Everyone found it at the same time: CONTEXT.

Everyone discovered context at once

Andrej Karpathy, founding member of OpenAI kicked off the renaming: “+1 for ‘context engineering’ over ‘prompt engineering,’” he wrote, calling it “the delicate art and science of filling the context window with just the right information for the next step.” Shopify CEO Tobi Lütke agreed: “the art of providing all the context for the task to be plausibly solvable by the LLM.” And this week Reuven Gorsht, my co-host on The Signal podcast, published the strategic version of the argument, and it’s the best one yet: your moat isn’t intelligence, it’s context. His line is the one worth taping to the wall: “You can rent intelligence by the token. You can’t rent context at any price.”

They’re all right. And almost all of the commentary that’s followed makes the same mistake.

It treats context as a technique. Something you do to a model. A bigger window, a better retrieval pipeline, a cleverer assembly of tokens at query time. That’s the engineer’s view, and for engineers it’s the correct one.

But if you run a bank, a lender, or an insurer, the technique is not your problem. Your problem is that nobody has told you where your context actually is. Because it isn’t in a window. It’s an asset you already own, and you’ve been accumulating it for decades.

It lives in exactly three places.

First, the reframe: the model is your core banking

Before the three places, one mental swap that makes everything else fall into line.

Think about your core banking system. Or your LOS, or your policy admin platform. There was a time when that choice was the strategic decision, the thing boards debated and careers were staked on. Today it’s plumbing. Mission-critical plumbing, stable and load-bearing and absolutely necessary. And nobody cares how it runs. They care that it does. Every ounce of differentiation your institution has lives in the layers built on top of it.

Model intelligence just made the same journey. It took core banking forty years to become invisible infrastructure. It took frontier models about three. The intelligence is stunning, the intelligence is table stakes, and the intelligence is available to your smallest competitor for the same monthly rate you pay. An advantage you have to re-buy every quarter isn’t an advantage. It’s a subscription.

So the question stops being which model? the same way it stopped being which core? The question becomes what you build on top. And what you build on top is made of context.

Here’s where yours is.

Place one: your data

Not the data in your data warehouse. That’s the easy part, and it’s the smallest part.

I mean the company’s actual knowledge. The email discussions where your head of credit explained, better than any policy document ever has, why you don’t touch a certain asset class. The SharePoint drive with nine hundred SOPs, six of which are load-bearing. The closing notes. The exception memos. The onboarding deck someone made in 2019 that still explains your niche better than your website does. Industry estimates put roughly 80% of enterprise data in this unstructured pile, and in a financial institution I’d bet the real knowledge skews higher than that.

The philosopher Michael Polanyi had a phrase for the deeper layer underneath even that: “We can know more than we can tell.” He called it tacit knowledge. Your institution is drowning in it. The underwriter who can feel a bad file in the first ninety seconds. The service rep who knows which customers mean it when they threaten to leave. None of it was ever written down anywhere, because until about eighteen months ago, nothing could read it anyway.

That’s the first job of your context layer: making what the institution knows reachable by the systems doing the work. Not a chatbot bolted onto a document dump. A deliberate answer to the question: where does our knowledge actually sit, and what can see it?

Place two: your embedded workflows

The second place your context lives is in motion.

Every institution has two versions of every process. The documented version, the one in the SOP with the flowchart. And the actual version, the one with the workaround for the system that’s been “being replaced” for four years, the unwritten rule about which exceptions go to which person, the step everyone does out of order because the order was wrong to begin with. The documented version is what a model can read. The lived version is what actually drives your business forward.

I’ve written before about the Integration Tax, about ops teams holding the stack together with their bare hands. This is the same territory from a different angle: all of that scar tissue, every workaround and escalation path and sequence your people carry in their heads, is context. It’s the difference between an agent that can process a loan and an agent that can process your loans. A generic model dropped into your operation knows the industry-average workflow, which is to say it knows nobody’s workflow.

The institutions getting real results from AI right now are not the ones with the smartest models. They’re the ones that did the unglamorous work of making their real processes explicit enough for a machine to run them.

Place three: your trust

The third place is the one that doesn’t feel like context at all, which is exactly why it’s the most valuable.

Your relationships. The broker who sends you the file first, before your competitors ever see it, because of how you handled a mess in 2021. The customer who renews without shopping the market. The team that ships because they trust the direction. The regulator who picks up your call. This is why people come back to you, and none of it is stored anywhere except in the accumulated history between your institution and the people around it.

You cannot scrape trust. You cannot fine-tune on it. A competitor could steal your entire document drive and copy your workflows step for step, and they still couldn’t take this layer, because it doesn’t live in your systems. It lives in the pattern of your conduct over time.

But here’s what most institutions miss: trust still produces data, and almost nobody treats it that way. Every renewal, every referral, every escalation that got resolved and turned a furious customer into a loyal one that history is context your AI should have. The relationship itself can’t be automated. The memory of the relationship absolutely can be, and when it is, every interaction your systems touch starts from the full history instead of from zero.

What to do Monday morning

Call it a context audit: three questions, one for each place.

Where does our knowledge actually sit, and can anything reach it? Map the emails, drives, SOPs, and documents where the real knowledge lives. Then ask the harder question: what fraction of it could any system, AI or otherwise, actually access today? The gap between those two numbers is your Context Debt, and it’s the most expensive number you’ve never measured.

Which of those workflows exist nowhere but in people’s heads? Take your three most important processes. Compare the SOP to what actually happens. Every divergence you find is context that walks out the door at 5 p.m.

Who holds our relationships — and what do they know that no system does? Your top producers and your longest-tenured ops people are carrying institutional context worth more than most line items on your balance sheet. Start recording the history while it’s still in the building.

Then stop evaluating models. Seriously. The eval matrix comparing four frontier models on reasoning benchmarks is the 2026 version of the core banking bake-off: a year of effort to choose between options that are converging while you deliberate. The intelligence is the least differentiated thing you will buy this decade. The context layer that feeds it, built from your data, your workflows, your trust, is the only part of the stack a competitor can’t order from the same menu.

Intelligence is table stakes. Your context is the game.

Stop shopping for intelligence. Start owning your context.

The week before that, Salesforce shipped Agentforce Operations. Back office agents, generally available. The launch leads with supply chain work, but one of the examples in Salesforce’s own announcement is a loan application moving through underwriting end-to-end. Extracting data from tax returns. Chasing signatures. Validating against compliance rules across systems.

A few months ago, I wrote that the LOS was reaching the end of life. That the architecture replacing it would be headless. The workflow is invisible, the database the command center, the AI the interface. I expected to spend two years defending that call.

Instead, two of the biggest platform vendors in the category shipped it.

Here’s what struck me when I read the announcements. The vendors are not announcing that they built the future. They are announcing that they will no longer be the ones defining it.

For two decades, the deal worked one way. The tech vendor wrote the workflow. The lender bought the workflow, then bent their operation around it. Every implementation project, every configuration workshop, every “that’s not how the system works” conversation was downstream of one assumption: the workflow ships with the platform.

I’ve started calling that assumption a vendor-authored workflow. It is baked so deeply into lending technology that most operators have never imagined the alternative. And as of this spring, it’s broken.

A few months ago, the question was whether the tech would ship. The tech shipped. The question now is who gets to define what the lending platform actually does.

Look at what actually changed, because it’s more than two press releases.

MCP, the protocol underneath all of this, went from roughly two million monthly SDK downloads at launch to 97 million by March. Sixteen months. Kubernetes took nearly four years to reach comparable deployment density. This is not an experiment developers are kicking around. It is infrastructure, already laid.

Meanwhile, the cost of running the old model keeps climbing. The MBA put production cost at $11,898 per loan in Q1. The long-term average is near $7,900. Margins are a sliver. The status quo isn’t a safe harbor anymore. It’s the expensive option.

The technical part of MCP is boring. The economic part is the story.

An LOS vendor historically controlled three things. The data model. The workflow. The interface. That bundle is what you were really paying for, and the workflow was the lock. Once your operation ran on the vendor’s stages and screens, leaving meant an eleven-month implementation and a seven-figure invoice.

When the platform accepts instructions from outside agents, two of those three things come unbundled. The workflow no longer has to be the vendor’s workflow. The interface no longer has to be the vendor’s screens. What the vendor keeps is the substrate. The data model. The system of record. The compliance rails, the audit trail, and the permissioning that decides which agent may touch which file. That is not a smaller business. It is a different one. In a regulated industry, it may be the more defensible one.

Clayton Christensen described this trap in The Innovator’s Dilemma. Incumbents respond rationally to disruption, and the rational response accelerates it. Blend, adding MCP is the right move. Not adding it would be worse. But the move that keeps them relevant is the same move that turns their workflow layer into a commodity. The vendors who win from here won’t be the ones selling workflow. They’ll be the ones selling the certainty underneath it.

Now flip to the lender’s side, because this is where it gets interesting.

Today, your underwriting workflow lives inside the LOS. Step by step, stage by stage. Configurable, not redesignable. Your operation is an annotation on the vendor’s flowchart.

With MCP, the workflow becomes something you author. An agent calls into the LOS for the file state. Into the bureau for the pull. Into the income verifier for the document. Into the AUS for the decision. The sequence, the exceptions, the escalation rules, the compliance posture: yours. Written by your team, changed by your team, in days instead of quarters.

The lender owns the orchestration. The vendors own the substrate.

And the substrate is not a consolation prize. An agent can author a workflow. It cannot manufacture what a regulator needs underneath it. Somebody has to guarantee where borrower data lives, keep the audit trail immutable, enforce which agent is allowed to decide what, and hold the orchestration rails steady while the agents on top change weekly. That is why it is still worth working with the FundMores, the Blends, and the Salesforces of the stack. You are no longer buying their workflow. You are buying the foundation your workflow stands on.

Carlota Perez, in Technological Revolutions and Financial Capital, splits every technology revolution into two phases. Installation, when the infrastructure gets built, and the money chases the technology itself. Deployment, when the infrastructure is assumed, and the value moves to whoever uses it best. Ninety-seven million downloads and two platform announcements in a single spring is what the end of installation looks like. Deployment is where operators win or lose.

Most lenders will miss this, and the reason is the frame.

The reflex when a vendor announces MCP support is to file it as another integration. Something for the IT roadmap, somewhere below the CRM migration. That frame is wrong. The platform just announced it will accept instructions from outside the platform. Which means the platform is no longer the operating model. You are.

Acting on that requires a competency most lending organizations never built. Workflow authorship. Agent orchestration. The discipline of designing your own operation instead of inheriting one. That is a different muscle than vendor management, and it doesn’t live in procurement.

The lenders who build it will ship operational changes in days, hold their underwriting logic as proprietary advantage, and stop paying rent on a workflow that was never really theirs. The lenders who don’t will keep adapting their operation to someone else’s flowchart, in the year the flowchart’s authors admitted it’s up for grabs.

A few months ago, I made a prediction about architecture. The LOS was reaching the end of its life, and something headless would replace it. The prediction held. The timeline didn’t. What I thought would take two years shipped in months. And the part I underweighted wasn’t the technology at all. It was the power shift.

MCP isn’t an integration story. It’s a sovereignty story. The agent is the workflow now. The platform is the substrate. And the lender, for the first time in two decades of mortgage technology, gets to write the operating system instead of buying it.

That is not a reason to walk away from your platform vendors. It is a reason to change what you buy from them. The workflow and the front end are yours now, adaptable in days. The compliance, the data integrity, and the orchestration rails are theirs to keep solid underneath you.

Stop renting your workflow. Keep paying for the rails. The difference between those two line items is the next decade of lending technology.

Chris Grimes is the founder of FundMore, an AI native loan origination platform. FundMore builds agentic mortgage and lending infrastructure for institutional clients across Canada and the US.

The week they ratified it, it was already wrong.

Two of the models named in it had been replaced by their makers. One rule about context limits no longer described any product on the market. And three tools that would have passed every test they wrote weren’t on the list at all, because they hadn’t existed when the drafting started.

They hadn’t governed the technology. They had photographed a snapshot of it and called the snapshot a policy. The ink dried slower than the landscape moved.

Every leadership team I talk to is trying to do the same thing. Draw a fixed boundary around a moving object. An approved tools list. A model allowlist. A committee that meets monthly to rule on a market that changes weekly. It feels like the responsible thing. It is the org chart equivalent of nailing the weather to the wall.

There is a name for what this costs you. I have started calling it Control Debt. The gap between how fast the technology evolves and how fast your rules about it can be rewritten. You pay it down in drag, in stale decisions, and in cycles you never got to run.

While you spend a quarter writing the rulebook, the operator who built for adaptation runs three cycles. The cost isn’t a fine. It isn’t a breach. It’s compounding irrelevance. The stack freezes at the exact moment you tried to control it.

The instinct to control the AI stack feels like the responsible choice. It is quietly the most expensive bet on the table

.It is worth being honest about why control is the reflex. It comes from somewhere real.

In lending, control is survival. Audit defence, model risk, data handling, the regulator who will ask you to show your work. For two decades the safe move was to standardize, lock the vendor list, and run a quarterly review. That worked because the environment was stable. The tools you approved in January were the same tools in December.

That world is gone. The reflex didn’t notice.

A control system built for a stable environment, run at constant speed inside an environment that stopped being stable, does not produce safety. It produces lag. The committee still meets. The list still gets maintained. And the gap between the list and reality widens every month nobody is willing to admit.

There is a law that explains exactly why this fails, and it is not a metaphor.

W. Ross Ashby, working in cybernetics in the 1950s, gave us the Law of Requisite Variety. His exact words: only variety can destroy variety. The version everyone uses now is softer, only variety can absorb variety, and the meaning holds either way. For a controller to actually regulate a system, it has to have at least as many possible responses as the system has possible states. Less famously and more bluntly: if the thing you are trying to control can do more than you can respond to, you are not controlling it. You just think you are.

Now hold your governance committee up against the AI landscape.

The committee has low variety. It meets twelve times a year. It produces a list. The landscape has enormous and accelerating variety. New models, new capabilities, new failure modes, every week. By Ashby’s own math, a fixed monthly framework cannot regulate a system moving that fast. You are not being careful. You are being outmatched, and the scoreboard is hidden because nothing has broken yet.

So what does the alternative look like? It is not chaos, and it is not no governance. It is a different shape of control.

Stanley McChrystal learned this the hard way. In Team of Teams he describes how a hierarchical, command and control task force kept losing to a decentralized enemy, Al-Qaeda in Iraq, that adapted faster than the chain of command could issue orders. His answer was not to tighten control. It was to replace it with what he called shared consciousness and empowered execution. Everyone sees the same picture. The decision gets made at the edge, by the people closest to the problem, fast.

That is the model. Govern principles, not inventories. Stop maintaining a list of approved tools and start maintaining a clear standard for what good looks like, then let the person doing the work evaluate this week’s option against it. Build infrastructure that assumes the tools will change. Abstraction layers instead of hard vendor bets. Swappable model providers instead of one name welded into the architecture. A standing habit of evaluation instead of a standing approval queue.

You are not loosening the grip. You are moving it to the layer that doesn’t move.

Here is where the regulated reader gets nervous, and they are right to.

None of this means you abandon real controls. It means you govern the invariants instead of the implementation. Where does borrower data go. Who is accountable for an output that reaches a customer. How do we verify a decision before we act on it. Those questions are stable. The answers don’t change when a new model ships. Lock them down hard.

Which specific model, which vendor, which prompt, which tool. Those are implementation. They change constantly, and the moment you write them into policy you have started accruing Control Debt again.

Lock the invariants. Let the implementation move. That single distinction is the whole game.

The leaders who freeze the stack to feel safe are going to be the ones who fell behind safely. Their policy will be airtight and two years out of date. Their committee will still be meeting.

The competitive edge does not go to the company with the tightest guardrails. It goes to the company that built guardrails as direction instead of as walls, and can pick up next quarter’s tooling without convening anyone.

Govern the direction, not the inventory. Build for the next version, not this one.

Stop trying to control the evolution. Build to harness it.

Chris Grimes is the founder of FundMore, an AI native loan origination platform. FundMore builds agentic mortgage and lending infrastructure for institutional clients across Canada and the US.

POS. LOS. Credit. Income verification. Asset verification. Fraud. AUS. Flood cert. Appraisal portal. MI portal. Investor portal. Title. Esign. Closing platform. Doc storage. The CRM where she logged the call this morning. A Slack channel where ops is asking her where it is. An email thread where the borrower is asking the same thing.

She isn’t behind. She isn’t bad at her job. This is what normal looks like for this file to live through.

None of those tools is broken. Each one is, in fact, the best version of itself the market has produced. She isn’t fighting bad software. She’s fighting the space between good software.

That space has a name nobody at the company can find on a P&L. It’s where work goes to wait. Documents living in three places. Conditions cleared in one system, still showing open in another. Status meetings whose entire purpose is to figure out which tool the file is currently sitting in. Borrowers ghosted not because anyone forgot, but because every tool was sure someone else was handling it.

You don’t have a tool problem. You have an architecture problem. And nobody on your team owns the architecture.

Open any mortgage operation and count the systems a single file touches between application and funding. You’ll get to twenty without trying. POS to CRM to credit pull to income and employment verification to asset verification to fraud check to LOS to underwriting workflow to AUS portal to flood cert to appraisal vendor to MI portal to investor portal to compliance check to title to esignature to closing platform to doc storage to email and Slack carrying state across all of it.

Twenty is the low end. Some lenders are at thirty.

Each of those tools has a champion who fought for it. Each was the right purchase in isolation. Each one produces metrics that look good on a vendor scorecard. None of them know about the file as a whole.

This is what the industry calls “best of breed.” In practice it became accumulating. Not integrating.

Clayton Christensen spent a career studying what happens when industries assume they’ve reached the modular stage of their lifecycle. His law of conservation of modularity is blunt. When end to end performance is not yet good enough on the dimensions customers care about, integrated architectures win, because the performance gains live in the integration. Once components become good enough, value migrates outward to modular providers competing on cost.

Mortgage tech assumed it was at the modular stage. It isn’t.

Borrower cycle time is not good enough. Borrower certainty is not good enough. The unit cost of producing a file is not good enough. None of the dimensions a borrower can feel, or a CFO can sleep with, are good enough.

So the math doesn’t work. Buying the best individual modules and snapping them together doesn’t produce a better mortgage. It produces a fragmented mortgage with the seams visible to the borrower.

The MBA’s Q1 2026 Quarterly Performance Report puts the average IMB production cost at $11,898 per loan, against a long term average near $7,900 and a current pretax profit of $727 per loan. That’s nearly $11,200 of stack burden riding on a sliver of margin.

Some of that is labor. Some is compliance. A real chunk is what I’ve started calling the Integration Tax. The cost that lives in the seams between tools.

Rekeying the same data four times.

Reconciling status across three systems before anyone can answer “where is this file?”

A processor spending forty minutes a day chasing a document that was already uploaded, just to the wrong tool.

The Integration Tax doesn’t appear on a vendor invoice. It hides inside payroll and cycle time. Which is exactly why nobody fixes it. The cost is real. It’s just nobody’s job.

Donella Meadows put it cleanly in Thinking in Systems: the structure of a system is the source of its behavior. If your behavior is an $11,898 production cost and a 42 day cycle that hasn’t materially moved in years, that isn’t a labor problem, and it isn’t a tool problem. It’s a structural problem dressed up as a thousand small inefficiencies.

Meadows ranked twelve leverage points for changing a system. The weakest interventions, the ones with the lowest payoff, are parameters. Buy a faster credit tool. Hire two more processors. Tighten an SLA. The strongest leverage points are the goals, the paradigm, and the structure of the system itself.

Reading the average mortgage tech roadmap is reading a document that lives almost entirely at the bottom of that hierarchy.

So what does “Kill the Stack” actually mean here?

Not rip and replace. Not build it all yourself. The industry has tried both versions of that. The in house custom LOS graveyard. The giant rip and replace transformations that cost millions and end with a different fragmented stack.

Kill the Stack means stop treating the accumulated tools as the architecture.

The architecture is the file moving end to end. The tools are inputs to it. Right now, in most lending operations, no one owns that file’s journey. The CIO owns infrastructure. The COO owns process. The CTO owns engineering. The vendors own their tools. The file owns no one.

This is where AI and agentic systems are actually useful. Not as another product to bolt onto the stack, but as the connective tissue the file has been missing. Holding state across the seams. Moving data between tools the way a coordinator does today, except continuously and without forgetting. Clearing conditions where they’re already provably clear. Telling the borrower what’s happening without the borrower having to ask.

That’s a different shape than “buy an AI feature.” It’s an architectural commitment that the file, not the tool, is the unit of design.

The industry has been buying tools for fifteen years. Cycle times haven’t materially improved. Costs are higher, not lower. Pull-through has declined every year for the past four (MBA, October 2025).

The next decade isn’t about buying better tools. It’s about deciding that someone, a person, a team, an architecture, owns the file end to end. Until that decision gets made, the Integration Tax keeps compounding. Quietly. Automatically. Every file.

Twenty systems isn’t the problem because twenty is a big number. It’s a problem because nobody designed the connection between any two of them.

That’s the architecture nobody owns.

That’s the stack worth killing.

Chris Grimes is the founder of FundMore, an AI native loan origination platform. FundMore builds agentic mortgage and lending infrastructure for institutional clients across Canada and the US.

Three pages. Built overnight by an agent. Yesterday’s customer escalations with the threads attached and triage recommendations queued. Pipeline movements with the deals that stalled flagged. The two regulatory items from Canada and the US that actually mattered for our roadmap, summarized in two paragraphs each.

Six months ago, that document required two analysts roughly five hours every morning to produce. Now nobody produces it. It just arrives.

That was the morning I stopped reading the debate over AI and jobs.

The agent didn’t take anyone’s job. It killed a workflow. The two analysts are still here. They’re spending Tuesday morning doing the work that actually requires their judgment, not assembling a status document that one piece of software can build while they sleep. The headlines say jobs are getting replaced. What’s actually getting replaced is something else.

Call it friction labor. The cumulative human attention spent on tasks that exist because no system surfaces the answer automatically. Not the work itself. The maintenance contract for the work.

And almost every AI conversation in business right now is missing it.

The public conversation defaults to a frame at the job level. “X” percent of roles will be automated by 2030. “Y” million workers displaced. McKinsey produces the chart. Goldman publishes the estimate. The numbers feel rigorous, and the debate feels important.

It’s the wrong unit.

David Autor at MIT has spent more than 25 years on this question, and his core finding is uncomfortable for both sides of the debate. Automation rarely eliminates jobs wholesale. It eliminates tasks. The remaining jobs are reorganized around tasks the technology can’t handle.

The bookkeeper example is the canonical proof. Between 1980 and 2018, computers took over most of the routine work bookkeepers did. Employment in the role fell by about a third. Real hourly wages for the people still doing it rose nearly 40 percent. The job didn’t get replaced. The task mix did. The bookkeepers who survived ended up doing more interesting work and getting paid more for it.

That pattern doesn’t fit neatly into either narrative. It isn’t “AI is coming for everyone.” It also isn’t “AI changes nothing.” It’s: the unit of disruption is the task, not the job, and we’ve been measuring the wrong thing.

A loan officer’s role contains around 30 distinct tasks, going by the official task taxonomy. A processor’s looks similar in scale. Most of those are friction labour: pulling data, formatting documents, checking statuses, chasing missing pieces. A small fraction is judgment work. Saying “AI is replacing the loan officer” obscures the actual shift. The role survives. The task mix changes.

Walk through what friction labour looks like in any company, and the inventory is always the same.

Rekeying data between two systems that don’t talk to each other. Summarizing a thread of emails into a status update that somebody needs in a different format. Checking whether an upstream thing happened before triggering the downstream thing. Monitoring a queue. Reformatting an output for the third audience this week. Writing the recap nobody attended the meeting to read.

None of that is knowledge work in any meaningful sense. It’s plumbing.

Companies hire knowledge workers and load them with plumbing. The smartest people in your building spend most of their week moving water between buckets, because no system was ever built to handle it and nobody had the budget to staff a real one.

Ajay Agrawal, Joshua Gans and Avi Goldfarb wrote about this directly in Power and Prediction. Their argument: AI is a steep drop in the cost of prediction, and the real economic value isn’t in “point solutions” that speed up individual tasks. It’s in “system solutions,” workflows redesigned around the assumption that prediction is now cheap.

The implication is the part most companies miss. Every friction task in your organization exists because somebody had to predict, monitor, or reconcile something at a cost the system couldn’t absorb. The friction is a fossil. It’s the shape of what you couldn’t automate the last time you reorganized.

When the cost of prediction collapses, that fossil layer becomes optional.

I’ll tell you what the morning brief actually does, because the abstraction makes it sound bigger than it is.

It reads our Slack channels. It reads my inbox. It checks the CRM for movement on the deals I care about. It scans two regulator news feeds. Then it writes three pages and emails me a single document at 6am. That’s it.

The first version took me about four hours to build. The current version runs on a connected Cowork stack and gets better every time I tell it what to cut. It cost less than a single analyst’s monthly seat.

It replaced about eight hours of weekly analyst work spread across two people. They didn’t lose their jobs. They got their Tuesdays back. They’re now doing the work they were hired to do in the first place: the qualitative interview synthesis, the deep review of deals that stalled, the customer relationships that need actual human attention.

What got killed wasn’t a job. It was a maintenance contract on a workflow that existed only because we couldn’t afford to keep paying humans to do it manually but also couldn’t afford to live without it.

The value didn’t show up as a headcount reduction. It showed up as a clock speed gain.

That’s the part companies are mostly not measuring, and it’s where the gap is opening.

Counting jobs is easy. Counting friction is hard.

Friction labor is invisible because it’s distributed. Five minutes here, twenty minutes there, three hours on a Friday afternoon. It never appears as a line item on a P&L. Companies measure headcount and revenue per employee. Neither of those tells you anything about how fast a question gets answered, how fast a decision routes through the building, how fast a deal moves from interest to close.

That last number is the one about to bifurcate every industry.

Two companies, same headcount, same tools, same product. One operates at three times the clock speed because they’ve systematically replaced friction labor with agents over the last 18 months. The other is still in a planning offsite debating which percentage of headcount to cut.

By the time the second company finishes its plan, the first one is already in the next market.

The debate over job replacement is going to keep going. It’s the easy debate. Politicians will hold hearings. Consultancies will sell decks. Every quarter someone will publish a new number and a new chart.

The companies that win the next decade aren’t asking how many people they can cut. They’re asking what friction they can kill.

The two questions sound similar. They produce completely different companies.

Your competitors aren’t replacing people.

They’re replacing the drag.

Stop counting jobs. Start counting friction.

Chris Grimes is the founder of FundMore, an AI native loan origination platform. FundMore builds agentic mortgage and lending infrastructure for institutional clients across Canada and the US.

Wednesday afternoon I sat down with the Bank of Canada’s April Monetary Policy Report and watched Tiff Macklem do something central bankers don’t usually do in public: lay out two opposite outcomes as live, equally weighted scenarios. If oil stays elevated, consecutive rate hikes are on the table. If the U.S. tightens trade restrictions in the CUSMA review, deeper cuts become necessary. Same press conference. Both directions.

Twenty-four hours earlier, Jerome Powell had chaired an FOMC vote that split 8-4. It was the most dissents at a single Fed meeting since October 1992. One governor wanted a cut. Three regional presidents wanted the “easing bias” stripped from the statement and replaced with symmetric language. They wanted the Fed to say, on the record, that the next move could go either way.

Two central banks, in twenty-four hours, said the same thing in different vocabularies: we don’t have a base case anymore.

That changes everything underneath. Almost every lender’s pricing engine, retention script, hedging book, lock workflow, and renewal communication assumes a direction. Down rates are good for refi. Up rates are good for retention. Every motion the system makes is wired to a forecast.

When the central bank itself says “either is plausible,” that wiring becomes a liability.

Here’s what the week showed.

United States

Rates back up. Borrowed motion just ended.

The 30-year fixed rose to 6.30% as of April 30 (Freddie Mac PMMS), up from 6.23% the prior week. That’s the first increase after three consecutive declines. The 15-year averaged 5.64%. The 10-year Treasury closed near 4.40% Thursday and held at 4.39% Friday.

The MBA applications survey for the week ending April 24 confirmed the turn: total applications down 1.6%, refis down 4%, refi share back to 42.5% from 44.2% the week before. Purchases up 1% on the week, still up 21% year-over-year. The marginal weekly print rolled negative the moment yields lifted.

My take: Last week’s edition named borrowed motion: pipeline activity wired to bond yields rather than generated by the lender’s own architecture. This week is the back-test. Yields lifted on the Iran-conflict premium and a hot core PCE print, refi share softened, applications softened, and three weeks of cosmetic recovery evaporated in five trading days. Same wiring, same input, opposite output. A pipeline that lights up and dims with a single external twitch is not generating flow. It’s mirroring weather. The lenders treating this as a setback are still reading the rate. The ones treating it as data are reading the architecture underneath.

The 8-4 vote isn’t political theater. It’s an architectural signal.

The Fed’s dissents broke in both directions. Stephen Miran wanted a 25 bp cut. Beth Hammack, Neel Kashkari, and Lorie Logan didn’t object to the hold. They objected to the easing-bias language and wanted the statement to say explicitly that the next move could go either way. They lost the vote on language. They didn’t lose the underlying point.

Core PCE for March, released Thursday, came in at 3.2% year-over-year, up from 3.0% the month before. The hawks have a case. The dove has a case. Powell exits as chair on May 15 and stays on as governor; Kevin Warsh’s nomination cleared the Senate Banking Committee 13-11 the same Wednesday the Fed was voting.

My take: Read this as a forecast lock signal, not palace intrigue. When the central bank itself can’t agree on direction, every lender whose system requires a directional consensus to function is exposed. Pricing engines tuned for an easing cycle. Retention workflows assuming refi opportunity. Hedging desks running short-duration bets on cuts. None are wrong yet. That’s the trap. They’re wrong only when the move comes the other way and you find out from the P&L. Architecture that holds both paths and switches without human intervention is the difference between a thesis being right and a quarter being whole.

Fannie Mae just told the industry to govern AI. Most aren’t ready.

Lender Letter LL-2026-04, issued April 8 and taking effect August 8, requires every Fannie Mae seller or servicer using AI or ML in origination or servicing to operate under a documented governance program. That means written policies covering the full life cycle of any AI/ML system, annual review, a designated owner, communication to relevant staff, and vendor-risk management on subcontractors using the same tools. It is the first sector-specific AI governance mandate in U.S. mortgage. Coverage and analysis caught up to it this week.

It lands alongside AD Mortgage’s 2026 Broker Survey, released Thursday, which reported that 55% of brokers now use AI daily or regularly (35% daily, 20% regular). 54% have not yet decided which technologies to adopt.

My take: Two facts on a collision course. Most brokers and lenders bought tools before they architected anything. They have AI-as-feature, not AI-as-substitute. The governance mandate isn’t asking whether you use AI. It’s asking whether you can govern it. Full life cycle, owner, vendor diligence, documented at the policy level, by August 8. The lenders who treated AI as a tool drawer will spend the summer writing governance documents about software they don’t actually understand. The lenders who chose an architecture have one document to update. Tools accumulate. Architecture chooses. The mandate is going to make that distinction visible to the GSEs in writing.

Canada

BoC said the quiet part out loud: there is no base case.

Macklem held at 2.25% as expected on April 29 and used most of the press conference to walk through two scenarios. The base case has inflation peaking near 3% in April 2026 and easing back to the 2% target in early 2027, but only if oil recedes as expected. In the alternative scenario where oil stays around US$100 a barrel, the Bank projects inflation peaking at 3.1% in Q1 2027 and lingering near 3% for a year, opening the door to consecutive rate hikes. If instead the U.S. sharpens trade restrictions in the CUSMA review (the formal Joint Review opens July 1), cuts become necessary to support the economy. The Bank’s growth path of 1.2% in 2026, 1.6% in 2027, and 1.7% in 2028 depends on which scenario lands.

The 5-year Government of Canada bond yield closed near 3.22% Thursday. Best 5-year fixed through brokers sits around 4.04%; RBC’s posted 5-year at the bank tier is 4.29%. The spread between the policy rate and what borrowers actually pay continues to widen.

My take: The Bank of Canada has now joined the Fed in publicly declining to commit to a direction. For Canadian lenders this changes how every renewal conversation in 2026 has to sound. “Rates will probably move lower” is now wrong in both directions. The conversations that retain borrowers will be scenario-based. Here’s your payment under oil-up. Here’s your payment under trade-down. Not directional. Servicers running single-rate-narrative scripts will lose borrowers to brokers who can model both within an active call. That’s not a comms problem. It’s an infrastructure problem.

CUSMA prep just opened its first meeting

Carney’s 24-member Advisory Committee on Canada-U.S. Economic Relations met for the first time on April 27, prepping for the formal CUSMA Joint Review starting July 1. Roughly C$33 billion of Canadian construction inputs, about 8% of the total, comes from the U.S. The active tariff layer on those inputs (50% on steel, 50% on aluminum, 25% on autos, lumber duties layered in) compounds directly into homebuilding cost. Carney named those tariffs as the live irritants and ruled out concessions before formal talks.

My take: The renewal wave is already the largest simultaneous payment shock in Canadian lending history. Layer a CUSMA outcome on top of it (bilateral, sector-specific, multi-quarter), and every renewal conversation in 2026 has a trade-policy variable baked in. Reinertsen’s principle applies: when variability rises, the value of fast feedback rises with it. Renewal infrastructure that surfaces scenarios in seconds wins. The infrastructure that surfaces them after a loan officer manually pulls a rate sheet doesn’t.

The bottom line

Two central banks made rate decisions in 24 hours. Both held. Both, in different vocabularies, said the next move could go either direction. The Fed had its biggest dissent in 33 years. The Bank of Canada laid out opposite scenarios as live possibilities. Mortgage rates ticked back up. Refis softened. Core PCE stayed hot. Fannie Mae’s first sector-specific AI governance mandate is now 90 days from going live. The CUSMA prep cycle opened.

The base case died this week. Quietly, on a Tuesday and a Wednesday, in two press conferences nobody covered the same way.

Forecast lock is the cost lenders pay when pricing, communications, retention, renewal, and hedging infrastructure assume a direction the central banks themselves no longer assume. Most lenders won’t see the bill until the next move comes the wrong way and the workflow doesn’t bend.

The lenders with infrastructure that holds both paths and switches without intervention spent this week doing nothing dramatic. The ones still running directional pipelines spent the week unaware they’re carrying a position.

You don’t get to know which side of the fork you were on until the road bends.

See you next Monday.

Chris Grimes is the founder of FundMore, an AI-native loan origination platform. FundMore builds agentic mortgage and lending infrastructure for institutional clients across Canada and the US.

Last month, I reviewed our quarterly numbers. Revenue was up. Headcount was up. But our margins kept shrinking, not by a lot at once, but steadily, like a slow leak you can hear but can’t find.

So I traced the bleed. Not to any single cost center. Not to a bad hire or a bloated tool budget. The cost was hiding in the gaps between things. In the handoffs that took two days because two systems didn’t talk to each other. In the status meetings that existed because no dashboard surfaced the information automatically. In the hours my ops team spent re-keying data that already existed somewhere else in the stack, just in the wrong format.

The most significant expense in my company was not listed on any line item. It was the friction between people, tools, and decisions — the drag that doesn’t show up in any report but determines whether a company scales or stalls.

I call it organizational drag. And if you’re running a company right now, it’s almost certainly the biggest cost you’re not measuring.

Engineers have a term for one version of this: technical debt. Ward Cunningham coined the metaphor in 1992 to describe what happens when you ship code that works but isn’t clean — you’ve borrowed against the future, and eventually you have to pay it back with interest.

What started as a metaphor became a line item. McKinsey estimates that 20 to 40 percent of the value of a company’s entire technology estate is consumed by tech debt. Other studies put it worse — 40 to 50 percent of development time goes to unplanned rework, bug fixes, and maintenance on systems that should have been rebuilt years ago.

Tech debt is the most visible form of invisible cost, because engineers have language for it. They can point at a codebase and say: this is broken, this is the cost, this is what it would take to fix.

But tech debt is just the version that got a name. The same pattern runs much deeper.

Every company carries what I think of as process debt — the accumulated weight of workflows that were never designed, just inherited.

The approval chain that requires three signatures because someone made a bad call in 2019 and the response was to add a checkpoint instead of fixing the decision-making. The weekly report that nobody reads but everyone fills out because no one remembers who asked for it. The onboarding flow that routes through four departments because the original one was designed when the company had twelve people and it just… grew.

Ronald Coase won a Nobel Prize for a deceptively simple insight. In his 1937 paper “The Nature of the Firm,” he argued that companies exist because markets have transaction costs — the friction of finding, negotiating, and enforcing every exchange. Firms reduce that friction by bringing coordination inside.

But here’s what Coase’s insight implies and most operators miss: firms themselves accumulate internal transaction costs over time. Every workaround. Every undocumented approval. Every process that exists because removing it would require understanding why it was created, and nobody has time for that archaeology.

Process debt is invisible because it looks like work. People are busy. Tasks are moving. The system appears functional. But underneath, a significant percentage of that activity is just friction — the organizational equivalent of spinning wheels on ice.

Nobody audits this. No one walks through the company and asks: how much of what we do today exists because it was the right thing to do, and how much exists because it was easier to add a step than to fix the system?

The third layer is the most expensive and the hardest to see: attention debt.

Cal Newport, in A World Without Email, describes what he calls the “hyperactive hive mind” — the default operating mode of most knowledge companies, where coordination happens through constant, unstructured, back-and-forth messaging. Slack threads, email chains, quick syncs, “got a minute?” interruptions. It feels like work. It feels productive. It is, in most cases, the single biggest drain on the only resource that actually generates value in a knowledge company: focused human attention.

The math is brutal. Every context switch carries a cognitive recovery cost. Your brain doesn’t toggle between tasks — it drags residue from the last one into the next. The more unresolved threads you’re carrying, the worse the drag gets. Newport calls this the “attention capital” problem: the raw material of knowledge work is human cognition, and most organizations treat it as if it were infinite and free.

It isn’t. Every status meeting could have been a dashboard. Every approval chain that pulls someone out of deep work. Every Slack ping resets a train of thought. These aren’t minor interruptions. They’re withdrawals from a finite account, and most companies are overdrawn by Tuesday.

Here’s what makes organizational drag so dangerous: these three debts don’t just coexist. They compound.

Tech debt creates workarounds. Those workarounds become process debt — undocumented, unquestioned, just the way things are done. Process debt requires coordination to manage, because when systems don’t flow, humans have to step in and push. That coordination becomes meetings, check-ins, status updates — attention debt. And attention debt means nobody has the unbroken time to step back, diagnose the underlying systems, and fix them. Which generates more tech debt.

The spiral is self-reinforcing. And because each layer is invisible on its own, the compounding effect is nearly impossible to see from the inside. What you see is the symptom: things feel slow. Headcount keeps growing but output doesn’t scale proportionally. Margins thin and nobody can point to a single cause.

The cause isn’t single. It’s systemic. It’s the drag between everything.

I’ve started running a different kind of audit. Not financial — operational. For every recurring process, meeting, and workflow, I ask three questions. What is this compensating for? What would make it unnecessary? And can we replace it with AI?

The answers are uncomfortable. Most of the time, the thing that would make the process unnecessary is a system that should have been built two years ago but wasn’t, because everyone was too busy running the workaround.

That’s the trap. The drag consumes the very capacity you’d need to eliminate it.

Breaking the cycle requires treating invisible costs with the same rigor you’d apply to visible ones. Measure handoff time, not just task time. Track rework rates. Count the meetings that exist because a system doesn’t. Put a number on the context switches your team absorbs in a week and do the math on what that costs in lost output.

You can’t cut what you can’t see. And right now, the most expensive thing in your company is almost certainly something you’ve never measured.

The question isn’t “where is the money going?”

It’s “where is the time going?”

And the answer, in almost every company I’ve seen, is the same: into the space between your systems, your processes, and your people’s attention. Into the drag.

Stop optimizing the visible costs. Start measuring the invisible ones.

Chris Grimes is the founder of FundMore, an AI-native loan origination platform. FundMore builds agentic mortgage and lending infrastructure for institutional clients across Canada and the US.

Everyone in the room got excited. I didn’t. Not because the numbers were wrong — they were real. Because of why they were happening. The rate fell because bond yields moved on a single news headline about Iran ceasefire chatter. The pipeline didn’t widen. The funnel didn’t improve. A geopolitical headline pulled the lever, and the lender got a heartbeat.

That’s the pattern across the industry right now. Apps go up when rates dip. Prepayments surge when they dip a little more. The whole system is wired to one external input. When the input moves, the system moves. When it doesn’t, the system doesn’t.

I have a name for this. Borrowed motion. The activity in your pipeline isn’t generated by your stack — it’s borrowed from whatever the bond market does on a given Tuesday afternoon. You don’t have flow. You have weather.

That’s a problem because this Tuesday and Wednesday, two central banks decide on rates within 24 hours of each other and both are expected to do exactly nothing. The borrowed motion stops. The volatility doesn’t. And the lenders who depend on that one input start to look very exposed.

Here’s what I saw this week, and what I think it actually means.

United States

Rates fall a third week. Applications and refis wake up.

The 30-year fixed averaged 6.23% as of April 23 (Freddie Mac PMMS), down from 6.30% the prior week — the third consecutive decline and the lowest level in three spring homebuying seasons. Mortgage News Daily showed top-tier rates briefly at 5.99% on April 24, a level seen only once before this year, on January 9. The 10-year Treasury closed at 4.31%, a touch lower after the DOJ dropped its probe into Jerome Powell.

The MBA’s applications survey for the week ending April 17 confirmed the bounce: up 7.9% week-over-week, the sharpest increase since late February. Purchases up 10%. Refis up 6%. Refi share at 44.2% of total volume.

The refi share jumped to 44.2% of total volume — the highest reading in months. Borrowers who locked in north of 7% during last year’s spike are jumping at any meaningful drop. This was a refi window, not a purchase one.

Underneath the headlines, serious delinquencies are still trending higher. FHA and VA delinquencies are well above pre-2020 norms, per Rob Chrisman’s Friday commentary. The borrowers stretched thinnest by last year’s rate spike are still rolling into distress, even as the headline rate gives the broader market a cosmetic lift.

My take: This week is the cleanest example I’ve seen of borrowed motion. Every “good” data point — applications, refi share, the rate move itself — was downstream of a single ceasefire headline and one DOJ news cycle. None of it is a signal that the housing market or the lender’s stack got better. It’s a signal that one external input twitched. The lenders treating this week as proof of recovery are the ones whose pipelines depend on rate stability to function. The ones treating it as noise are the ones with infrastructure that doesn’t need a rate gift to generate flow.

Two AI launches that tell two different stories

Better launched a conversational credit decision engine inside ChatGPT on Monday — a direct integration that lets lending teams hit Better’s Tinman AI underwriting from inside the chat surface. The same day, iLeads launched a machine-learning lead classification model that scores leads on property, mortgage lien, and borrower attributes, claiming up to 3x net revenue uplift.

Two AI announcements. Two completely different architectural choices. Better is moving its underwriting brain to where the operator already lives. iLeads is making the existing lead funnel smarter at one node.

My take: This is the AI bifurcation in plainer view than usual. iLeads is AI bolted onto the existing stack — same funnel, same flow, smarter scoring at one step. Better is AI replacing the surface entirely. The former saves time at one node. The latter changes who does the work and where. Most lenders this year will buy the iLeads version because it doesn’t disrupt anything. Most lenders this year will end up wondering why their AI investment didn’t move the needle. The reason is right there in the architecture diagram — one is a feature, the other is a substitution. Goldratt would call the first one optimizing a non-bottleneck. The whole system doesn’t get faster because one node got smarter.

The Powell-Warsh week told you nothing about rates and a lot about volatility

The DOJ dropped its criminal probe into Jerome Powell on Thursday. Kevin Warsh’s confirmation hearing the prior Monday had him pledging Fed independence while declining to call the President’s public rate demands a threat. Treasury yields ticked lower on the news. The Fed meets April 28–29 with a 99.9% probability of holding at 3.50%–3.75%.

My take: Whoever runs the Fed in 90 days, the structural fact remains: short-rate decisions matter less to the borrower than what bond markets do between meetings. A change of Fed leadership probably means more political pressure, more market reactivity to administration commentary, and more weeks like this one — where a single news cycle moves rates 10 bps and your entire applications volume responds. If you can’t reprice, re-qualify, and re-engage borrowers automatically when the input shifts, you’ll spend the next two years catching up to the market instead of being in it. Rate volatility is becoming the operating environment. Build for it, or absorb it.

CFPB keeps withdrawing its withdrawals

The Bureau has spent the past year pulling back rules it had previously tried to roll back. The most visible example, still on the books and still relevant this spring: a direct final rule that would have rescinded state-official notification procedures under the Consumer Financial Protection Act, withdrawn last summer and never replaced. The procedural detail matters less than the pattern. The federal compliance framework keeps reversing itself. State AGs keep filling the gap.

My take: I keep writing this section and the underlying point keeps holding. Federal compliance is becoming a calendar of withdrawn rules and rescinded rescissions, while state AGs are doing the actual enforcement. If your compliance stack was built around a stable federal framework, it’s getting more out of date every week. Configurable, state-aware compliance infrastructure is no longer a roadmap item. It’s the thing that keeps you from a multi-state RESPA action you didn’t see coming.

Canada

The Bank of Canada will hold. Fixed rates already moved.

The Bank of Canada announces Wednesday at 9:45 ET. All 41 economists in the most recent Reuters poll expect a hold at 2.25%. Prediction markets are at 96.5% for no change. Macklem and Rogers will hold a press conference at 10:30 with a fresh Monetary Policy Report.

The policy rate isn’t where the action is. Government of Canada bond yields climbed into the 3.1% range this week — their highest level since mid-2024 — pushed by Strait of Hormuz developments and oil price reaction. Best 5-year fixed through brokers is sitting at 4.04%; major banks range from 4.3% to 4.9%.

My take: Two consecutive weekly editions where I’ve made the same point and the data keeps confirming it: the BoC overnight rate has nothing to do with what Canadian borrowers actually pay. Fixed mortgage pricing is being set by 5-year GoC yields, which are being set by global oil prices, which are being set by geopolitics. The lenders still talking to renewing borrowers about “BoC moves” are sending the wrong signal. The conversation that retains those borrowers is about bond spreads, term selection, and the volatility premium baked into every fixed product right now. That’s an architecture problem — you need infrastructure that can model and surface that conversation in seconds, not a loan officer manually pulling rate sheets.

Carney’s housing announcements are starting. They won’t move the math fast enough.

Prime Minister Carney announced an expanded federal-Ottawa housing partnership on Wednesday — $400 million committed, with the agreement targeting roughly 3,000 homes overall. The first eight Build Canada Homes projects approved this week deliver 1,100+ affordable rental units inside that pipeline. Nationally, Build Canada Homes has now committed to over 10,000 units since September, with 1,400 under construction or breaking ground in the next two months.

CMHC’s standing estimate of the supply gap to restore affordability by 2030 is 3.5 million units above the current trajectory. The current Build Canada Homes commitments are roughly 0.3% of that.

My take: Credit where it’s due — the Carney government is moving from announcement to construction faster than recent predecessors. They are. The honest read of the math is that the rate of progress is still an order of magnitude below what affordability requires. That’s a constraint problem in the Goldratt sense. You can’t add demand-side incentives without adding supply at scale, and you can’t add supply at scale by approving rental projects one city at a time. Until the supply throughput changes, every demand-side measure pushes prices up. CMHC said this clearly two weeks ago. Watch whether the next budget treats supply as a flow problem or a project list.

OSFI’s CAR Guideline (2026) is in effect — and almost nobody is talking about it

The Capital Adequacy Requirements Guideline (2026) took effect November 1, 2025 and January 1, 2026 depending on fiscal year-end. The big change: a mortgage gets classified as Income-Producing Residential Real Estate when more than 50% of qualifying income comes from rent. OSFI was explicit this month that the IPRRE classification governs capital, not borrower qualification — Guideline B-20 still controls underwriting.

My take: The capital classification doesn’t change how the borrower qualifies, but it changes how expensive the loan is for the lender to hold. That’s going to bleed into pricing for investor mortgages, especially at the smaller institutions where capital is tighter. Lenders running pricing engines that don’t model risk-weighted capital cost at the loan level are about to discover they’re underpricing investor product. Quietly, this is the kind of change that punishes a flat pricing stack and rewards one that reflects the real economics.

The bottom line

Three weeks of falling rates. A 7.9% applications jump. A 44% refi share. Two AI launches. A confirmation hearing that didn’t change anything yet. A federal-municipal housing announcement. Two central banks are holding in the coming 24 hours.

None of this is the actual story. The actual story is that every visible data point this week was downstream of one input — bond yields reacting to a ceasefire headline. When the input twitched, the data twitched. When it stops twitching this week, the data will stop too.

That’s borrowed motion. That’s a system without architecture.

The lenders building infrastructure that doesn’t depend on the next external twitch are spending these weeks making boring, structural progress. The ones who treat each rate-driven uptick as proof of recovery are running on weather.

The Fed and BoC won’t move this week. Your pipeline’s response to that fact will tell you whether you’ve built infrastructure — or you’re still running on weather.

See you next Monday.

Chris Grimes is the founder of FundMore, an AI-native loan origination platform. FundMore builds agentic mortgage and lending infrastructure for institutional clients across Canada and the US.

The other two days and twenty-three hours? Waiting. Waiting for a document that had been uploaded to one system but hadn’t synced to another. Waiting for an income calculation that required data from three different sources, none of which shared a format. Waiting for a condition to clear that had already been resolved in a different tool but hadn’t propagated to the one the underwriter was looking at.

The underwriter wasn’t slow. The infrastructure around her was.

This is the part that most people in lending get wrong. “Underwriting is the bottleneck” has become gospel, repeated in board decks, vendor pitches, and ops meetings like it’s an obvious truth that just needs a better solution. So lenders hire more underwriters. Add more training. Layer on more oversight. Buy AI tools that promise to make credit decisions faster.

None of it moves the needle. Because the bottleneck was never the decision. It was everything that happened before the decision could be made.

That conversation stuck with me, because it wasn’t the first time I’d heard it. I’ve had some version of this same discussion with lenders across Canada and the US. The specific file changes. The number of days changes. The tools involved change. What doesn’t change is the diagnosis: a person gets blamed for a problem that a system created.

Walk through the actual time breakdown on a typical file once it hits the underwriting queue. From submission to decision, call it 72 hours. How much of that is genuine underwriting judgment? The part where a skilled human evaluates creditworthiness, weighs compensating factors, applies guidelines to an actual borrower’s situation?

Thirty minutes. Maybe an hour on a complex file.

The rest is infrastructure. Document collection. Format reconciliation. Data validation. Condition chasing. Re-keying information that exists in one system but can’t be read by another. Waiting for third-party verifications that were ordered late because the system didn’t trigger them automatically. Staring at a queue that isn’t moving because the previous step hasn’t released the file yet.

The underwriter isn’t slow. They’re idle. Not because they aren’t working (they’ve got thirty other files they’re toggling between) but because on any given file, the system hasn’t delivered what they need to make the next decision. So they move to another file, and the cycle repeats. Context-switching across a portfolio of half-ready files, each one stuck at a different point in a broken data pipeline.

That’s not an underwriting problem. That’s a plumbing problem.

Eliyahu Goldratt spent his career studying why systems underperform. His insight, laid out in The Goal, was deceptively simple: every system has a constraint, one point that limits overall throughput. Optimize anything other than the constraint and you haven’t improved the system. You’ve just made a non-bottleneck more efficient, which means you’ve changed nothing that matters.

I kept coming back to Goldratt when we were building FundMore. Every time a lender told us their underwriting was the problem, I’d ask them to walk me through where the file actually spent its time. The answers were always the same. The underwriter wasn’t the constraint. The data movement was.

In lending, everyone assumes the constraint is the underwriter. So they optimize the underwriter. Faster guidelines engines. Better decision support. AI-assisted risk analysis. And the underwriter does get faster, at the thirty minutes of actual judgment they were already doing well. The other seventy-one hours and thirty minutes don’t move.

Donald Reinertsen, in The Principles of Product Development Flow, went further. He showed that in any complex process, wait time dominates work time, often by a factor of ten or more. The problem isn’t that individual steps are slow. It’s that invisible queues form between steps, and nobody measures them. The work sits. The clock runs. And because the queue is invisible, it doesn’t show up in any dashboard, any KPI, or any vendor report, so nobody manages it.

Reinertsen showed that invisible queues are the root cause of poor performance in any development flow, what he described as “the root cause of the majority of economic waste in product development.” In lending, they’re the time between document upload and document availability. The gap between condition request and condition response. The lag between data entry in one system and data appearance in the next. None of these are “underwriting.” All of them are what makes underwriting take three days.

Here’s what makes this worse: the standard response to a misidentified bottleneck compounds the problem instead of solving it.

I asked that lender what they’d already tried. More underwriters, they said. Better training. A new QC layer. They were about to look at an AI underwriting tool. Each one of those responses sounds reasonable. Each one misses the point entirely.

Hire more underwriters? Now you have more people waiting for the same broken data pipeline. The queue upstream hasn’t changed. You’ve just added capacity at the one point in the system that wasn’t actually constrained.

Add more oversight and QC checkpoints? Each one is another gate in an already slow flow. Another place where a file can sit, waiting for someone to look at it, before it moves to the next stage of sitting and waiting.

Deploy AI underwriting on top of the existing stack? You now have a system that can make a credit decision in three seconds, then wait seventy-one hours and fifty-nine minutes for the next file to be ready. You’ve built a jet engine and bolted it to a horse cart.

Every one of these interventions feels productive. Every one of them misses the point. The constraint isn’t judgment speed. It’s data movement. And until you fix data movement, nothing else you do to underwriting will show up in your cycle times.

What changes when you treat the stack, not the underwriter, as the constraint?

When I walked that lender through this framing, something shifted. They’d been thinking about underwriting as a stage with a person at the center. The reframe is simpler than it sounds: underwriting is a continuous activity that happens whenever the data is ready. The person was never the problem. The pipeline was.

Document collection becomes parallel, not sequential. Validation happens at the point of entry, not as a separate step downstream. Data moves through a single architecture instead of being exported, transformed, and imported across six different systems.

The underwriter stops toggling between thirty half-ready files and starts working a queue of files that are actually ready for a decision. The context-switching drops. The idle time drops. The cycle time drops. Not because the underwriter got faster. Because they stopped waiting.

This isn’t theoretical. Lenders who’ve moved to integrated, event-driven architectures, or even started agentic system, where data flows continuously rather than being batched and handed off, are seeing cycle times compress by days, not hours. Not because they replaced underwriters with AI. Because they removed the invisible queues that were making underwriters look slow.

The lender I spoke with last week isn’t unusual. Most people reading this have probably been in some version of that conversation, either telling someone their underwriters are the problem or being told it themselves.

The mortgage industry has spent two decades trying to fix underwriting. Faster tools. Smarter models. More automation at the decision layer. And underwriting cycle times have barely moved, because the decision was never what took three days.

The plumbing took three days. The data movement. The format reconciliation. The handoffs between systems that were never designed to talk to each other.

Your underwriter isn’t your bottleneck. Your stack is.

Stop hiring more people to wait faster. Fix the pipes.

Chris Grimes is the founder of FundMore, an AI-native loan origination platform. FundMore builds agentic mortgage and lending infrastructure for institutional clients across Canada and the US.

In the middle of it all, Fannie Mae’s new rate buydown disclosure requirements go live today.

The macro is quiet — briefly. The structural questions haven’t gone anywhere.

Here’s what this week is actually telling you.

United States

Rates fall for a second week. Buyers aren’t buying.

The 30-year fixed averaged 6.30% as of April 16 (Freddie Mac), down from 6.37% the week prior and 6.46% two weeks ago. The 15-year fixed came in at 5.65%. Mortgage News Daily is showing rates essentially flat at recent lows as of this morning — a four-week low by Freddie Mac’s measure.

On paper, this is a good news story. In practice, it isn’t moving the market.

NAR’s March existing-home sales report came in at 3.98 million annualized — down 3.6% month-over-month. Median price: $408,800. Months of supply: 4.1. Inventory is up 8.1% year-over-year nationally, with 11 states — Arizona, Colorado, Florida, Texas, and others — now sitting above pre-pandemic 2019 levels.

NAR quietly revised its 2026 forecast downward this week. Existing-home sales are now expected to grow 4% for the year, down from a prior projection that assumed rates moved lower, faster.

My take: Rates are falling and buyers are still sitting. That tells you the problem isn’t purely cost anymore. It’s confidence. Every week the tariff situation remains unresolved, buyers face a calculation they can’t close: is this a good time to lock in a 6.25% rate, or will conditions shift materially in 90 days? Most are choosing to wait. That hesitation doesn’t show up on any lender’s pipeline report — it shows up as volume that never arrived. Lenders who can surface pre-approved buyers instantly when confidence briefly returns will capture the bounce. Lenders who depend on borrowers initiating the process themselves will miss it.

New GSE disclosure requirements go live today

Starting today, Fannie Mae’s new loan-level rate buydown disclosure requirements are live. Freddie Mac’s equivalent change — originally scheduled for April 20 — was postponed in March with timing TBD. If you’re selling to Fannie and haven’t updated your disclosure workflows, this is the week it matters.

Also quietly this month: the GSEs extended the maximum term for manufactured housing cashout refinance loans from 20 to 30 years, eased prefunding rules, and pushed back the implementation deadline for the new Uniform Closing Dataset — recognizing that lenders already have competing mandates with the November 2026 UAD 3.6 requirement on the horizon.

My take: The GSEs are releasing a steady stream of policy changes with tight effective dates and no coordinated rollout calendar. Every one of these is a workflow change, a disclosure update, or a compliance configuration that someone on your team has to catch and implement. If you’re tracking this manually — a spreadsheet, a weekly email scan, someone’s inbox — you’re one missed memo away from a sellability problem. This is exactly the kind of operational surface area that compliance infrastructure should handle automatically. Most lenders aren’t there yet.

State enforcement is accelerating as the CFPB pulls back

The pattern from the last several months is now a trend: CFPB examinations are down sharply, and state attorneys general are picking up the slack. RESPA cases and state probes have risen meaningfully across the Mid-Atlantic and other regions. A coalition of state AGs secured hundreds of millions in restitution against a lender last week for RESPA disclosure failures related to bundled add-on products.

The fragmentation is real and accelerating.

My take: Every lender operating across multiple states is now navigating different enforcement timelines, different disclosure standards, and different AG priorities — with less federal clarity to harmonize against than at any point in the last decade. The compliance stack that was built around a unified federal framework is now operating in an environment it wasn’t designed for. Configurable, state-aware compliance infrastructure was a nice-to-have two years ago. Today it’s the difference between a manageable audit and an eight-figure lawsuit.

The Fed: hold expected, again

The April 28-29 FOMC meeting is expected to end with rates held at 3.50%–3.75%. Markets are pricing in near-certainty of no cut. The Fed’s own projections put PCE and Core PCE at 2.7% for the year — though actual readings have been running higher. Some banks have abandoned 2026 cut forecasts entirely.

My take: The Fed is watching inflation numbers that tariffs are keeping elevated, in a growth environment that’s softening. That’s stagflation-adjacent — not a comfortable place to cut from. The near-term rate environment is going to be determined more by trade policy and geopolitical risk than by anything the Fed controls. Which means for the third week in a row, the real story isn’t the rate. It’s how fast your stack can reprice, re-qualify, and re-engage a borrower when conditions shift.

Canada

Carney wins. Now comes the housing test.

Last week’s by-elections on April 13 handed Mark Carney’s Liberals a majority government — wins in University-Rosedale, Scarborough Southwest, and Terrebonne. Carney immediately vowed to focus on housing affordability. It was one of the central policy fault lines of the campaign, with competing platforms across parties on demand-side incentives, supply targets, zoning reform, and mortgage access expansion.

CMHC dropped timely analysis this week: demand-side measures like tax incentives or expanded mortgage access push prices up unless they are precisely calibrated and paired with equivalent supply. The math is blunt. Every government initiative that puts buyers into the market without adding housing supply generates upward price pressure.

My take: Carney inherits a housing system under structural stress. The renewal wall is active, supply is constrained, and affordability is at multi-decade lows. The political pressure to “do something” for buyers is real and immediate. The risk is that the interventions chosen are the ones easiest to announce — demand-side measures — rather than the ones that actually change the supply-side architecture. Watch what the new government does in the first 90 days, not what it promised on the campaign trail.

The Bank of Canada decision: Wednesday, April 29

The BoC holds its rate decision the same week as the Fed. Consensus expects another hold at 2.25%. The overnight rate has been steady for three consecutive announcements.

The rate is holding. Fixed mortgage rates aren’t. The best 5-year fixed through brokers is now around 4.04%; major banks are ranging from 4.3% to 4.9% depending on term and insured status. Variable rates are running around 3.3% with a prime rate of 4.45%. Government of Canada bond yields remain above 3%, pushed by the same geopolitical forces that are moving US markets.

My take: Two central bank decisions in two days, both widely expected to end in holds, both overshadowed by the political uncertainty that’s actually moving the bond market. The BoC doesn’t control fixed rates and hasn’t for months. The lenders who built borrower communication workflows around policy rate movements are sending the wrong message. The conversation that matters is about bonds, spreads, and the volatility premium baked into every fixed-rate product right now.

OSFI’s renewal warning deserves more attention than it’s getting

OSFI published its Annual Risk Outlook this month. The number buried in it: 3.1 million Canadian mortgages — 52% of all outstanding — will be renewing by the end of 2027. 1.3 million of those originated in the low-rate period of 2021-2022 and are facing material monthly payment increases.

OSFI has introduced institution-specific portfolio limits on uninsured mortgages exceeding a 4.5x loan-to-income ratio in response to household leverage concerns.

March housing starts came in with a six-month trend down 2.9% to 248,378 units, even as actual starts were up 10% year-over-year.

My take: The renewal wave is not a future risk. It’s a current one. 52% of all outstanding mortgages repricing over 20 months is the largest simultaneous payment shock in Canadian lending history. Lenders and servicers who are running manual renewal workflows — spreadsheet pipelines, reactive outreach, loan-officer-driven conversations — are going to be overwhelmed. The servicers who will retain borrowers through this are the ones with automated, proactive, scenario-based renewal infrastructure. The ones who wait for the borrower to call will see them go to a broker who already sent them three options.

The bottom line

Rates are falling. Buyers aren’t moving. A new majority government in Canada just inherited one of the most stressed housing markets in the country’s history. And next week, the Fed and the Bank of Canada both decide on rates within 24 hours of each other.

None of this is noise. All of it is signal.

The signal is the same as it’s been: the lenders who built for stability are going to find the coming weeks uncomfortable. The ones who built for adaptability are going to find them like any other.

See you next Monday.

Chris Grimes is the founder of FundMore, an AI-native loan origination platform. FundMore builds agentic mortgage and lending infrastructure for institutional clients across Canada and the US.

Two weeks ago, I’d blocked Tuesday. Not casually. Intentionally. For the first time in weeks I had real work I wanted to accomplish — things that had been stacking up for months, some of them longer. I finally had the tools and the space to actually attack them.

By 10 am, two quick syncs had claimed most of it.

By noon, a check-in materialized in the last open slot.

By 4pm, I’d attended six meetings and written maybe 200 words.

I headed out to dinner that evening, thinking: I didn’t get any of it done. Again.

But I stopped being annoyed at the people who booked the meetings. I started asking what each one was actually doing. And every single one had the same job: collecting information so someone could make a decision. A status check. A loop that needed to close before the next step could happen. Information moving from one person to another through thirty minutes of scheduled time.

That’s it. That’s all they were.

On Wednesday, I blocked the time again and intentionally ignored everyone. I went and built agents to do it instead. A status reporter. A decision router. A check-in replacement. A financial outlook. All of them run on Claude Cowork, integrated with the business’s existing tools via a set of connectors.

I don’t know yet if “Tuesday” is fixed. I’ll tell you when I find out.

But the diagnosis got clearer. Every meeting on my calendar stopped looking like coordination. It started looking like a system I hadn’t built yet.

Turns out that’s true of every calendar. It’s not a scheduling problem. It’s debt. Every founder I’ve talked to has a version of Tuesday, the blocked afternoon that didn’t survive contact with the week, the work that keeps not getting done. The details change. The structure is always the same.

The calendar wasn’t designed for your company. It was designed to schedule shared physical resources in an era when coordination required presence. The hourly grid made sense when getting twelve people in the same room was the only way to move something forward.

Paul Graham named the problem in 2009. The hourly calendar is a manager’s tool. It assumes time is fungible, that an hour is an hour regardless of what came before or after it. That assumption works for coordination. It destroys creation. A single meeting, he wrote, can blow a whole afternoon by breaking it into two pieces, each too small to do anything hard in. That essay is nearly two decades old. Nothing about the calendar has changed since.

The work has changed. The calendar hasn’t.

Walk through the recurring meetings on your calendar and ask what each one is actually solving.

The weekly status meeting exists because no one built a system that surfaces project state without being asked. That’s what the status reporter does. It pulls from the tools the team is already working in, formats it for whoever needs it, and delivers it on a schedule. Nobody has to ask. Nobody has to attend.

The approval meeting exists because authority is undocumented, so decisions route upward by default until they hit someone willing to say yes. That’s what the decision router handles. It picks up the request, attaches the context, and moves it to the right person with everything they need to act. No sync required.

The quick check-in exists because the async channel isn’t reliably moving context between people, so the gap gets patched with scheduled time. That’s the check-in replacement. The loop closes automatically. No calendar invite, no thirty-minute block, no recovery window on the other side.

The financial outlook meeting exists because the numbers aren’t visible unless someone pulls them together and presents them. So I built an agent that does the pulling, formats the summary, and delivers it when it’s needed.

Every recurring meeting is a maintenance contract on a broken system. These four agents are my attempt to cancel four of those contracts.

The cost of running a company on interrupt-driven coordination is real, and almost no one measures it.

Gloria Mark at UC Irvine found it takes roughly 23 minutes to fully recover focus after an interruption. Not 23 minutes to finish the meeting. 23 minutes after it ends before the interrupted work resumes at depth. A calendar with four meetings spread through the day doesn’t cost four hours. It costs four hours plus four recovery windows. For a team of ten, that math compounds fast.

And because the cost never shows up on a P&L, it never gets treated as a cost. It hides in slower cycle times, lower quality decisions, and the persistent feeling that the work isn’t moving as fast as it should. Nobody looks at a week of meetings and calculates the recovery tax. They just wonder why output is falling despite everyone being busy.

That’s what I was doing every Tuesday. Not measuring it. Just absorbing it.

Building the four agents took a few hours. Claude Cowork connected to the tools we already use — Slack, calendar, project tracking, financials. The agents aren’t sophisticated. They’re doing simple things: pulling data that already exists, formatting it, routing it to the right place at the right time. The meetings were doing that work manually. Now the agents do it.

What I don’t know yet is whether it actually gives the time back. Whether the meetings stay gone or just get replaced by new ones. Whether the information lands the way it needs to or generates a different kind of follow-up. I’ll run it for a few weeks and report back.

But here’s what I already know. The meetings I cut weren’t valuable. They were necessary. There’s a difference. Valuable is something you’d protect. Necessary is something you do because the alternative doesn’t exist yet. The agents are the alternative.

If the alternative works, the meetings don’t come back.

I’m still in meetings. Some of them are genuinely irreplaceable. The conversations where something needs to be felt, not just communicated. The decisions where alignment can only happen in real time. Those exist. They’re maybe 20 percent of what most calendars contain.

The other 80 percent is sync debt. Coordination problems that were easier to schedule than to solve.

The question isn’t how do I have fewer meetings. It’s: what is this meeting compensating for? And then: can a system do that instead?

For four of mine, the answer was yes. Build took a few hours. If it works, I get every Tuesday back. Permanently.

That’s the trade I’m trying to make. I’ll let you know how it goes.

Chris Grimes is the founder of FundMore, an AI-native loan origination platform. FundMore builds agentic mortgage and lending infrastructure for institutional clients across Canada and the US.

April 13, 2026

A little late, instead of reading this with your morning coffee, it will have to be with your afternoon pick-me-up. Stay with me, it’s only my second week, and hopefully, this doesn’t put you to sleep! Onto the pause edition….

On Wednesday, Trump paused most of his reciprocal tariffs for 90 days. The stated reason was that the bond market was getting “a little queasy.” The 10-year yield had surged toward 4.5%. The administration blinked.

Mortgage rates dropped. The 30-year fixed fell from 6.46% to 6.37% in a week (Freddie Mac). Refi applications ticked up. Mortgage Twitter exhaled.

But here’s what I kept thinking about: a political decision caused a 50+ basis-point swing in borrowing costs over a few days. Not a Fed decision. Not an economic data release. A tweet, then a reversal, then another statement.

If your operation can’t absorb that kind of volatility, if your pipeline, your pricing, your lock management, your borrower communication requires a stable rate environment to function, that’s not a rate problem. That’s an architecture problem.

The pause didn’t fix anything. It just gave everyone a week to catch their breath before the next move.

United States

Rates: a reprieve, not a resolution

The 30-year fixed averaged 6.37% as of April 9 (Freddie Mac), down from 6.46% the prior week, driven by the tariff pause and brief Treasury rally. By April 13, Bankrate was showing 6.41% (6.48% APR). The improvement evaporated almost as fast as it arrived.

The 10-year Treasury, which was briefly at 3.86% during peak panic, is now back above 4.2%. Markets are pricing a 97%+ probability that the Fed holds at its April 28-29 meeting. There are no cuts on the near-term horizon, regardless of what the administration would prefer.

Fannie Mae still projects rates dipping below 6% later in 2026. That forecast assumes the macro stabilizes. That assumption is doing a lot of work.

My take: The rate number this week is almost beside the point. What matters is volatility: a 50+ bps swing in days driven by a political decision rather than economic data. Lenders managing locks and pipeline commitments in this kind of environment quickly learned how much manual handling their stack requires when conditions shift that quickly. The ones with event-driven pricing and automated lock extension workflows felt it less. The ones running rate sheets manually felt it everywhere.

The pause paused most tariffs. China’s went to 145%.

While Trump suspended tariffs on most trading partners for 90 days, tariffs on Chinese goods escalated to 145%. That matters for housing because China is a primary source of building materials, appliances, and fixtures. The pause didn’t end the trade war. It concentrated on the country most embedded in the supply chains that build and equip homes.

More importantly, it tells buyers that nothing is resolved. The 90-day clock is running on an unknown outcome. Rates might be 6.3% today and 6.8% in two months if negotiations collapse, or 5.9% if they don’t. No one knows. And when buyers can’t model the environment 90 days out, they wait.

The housing data reflects it. There are now roughly 630,000 more active sellers than buyers, the largest mismatch since tracking began in 2013. Inventory is up about 20% year-over-year. Supply isn’t the problem. Confidence is.

My take: A rate problem you can model. Uncertainty that resets every news cycle is harder to build a workflow around. Lenders whose pipeline depends on borrower conviction are navigating something the stack wasn’t designed for. The ones who’ll hold margin through this are the ones who can move fast when confidence briefly returns — automated pre-approvals, instant scenario updates, borrower communication that doesn’t require a loan officer/broker to manually restart every time conditions shift. Uncertainty as a default condition, not a weather event.

The CFPB vacuum is becoming a compliance stack problem

The regulatory picture continues to fragment. CFPB examinations are reportedly dropping from roughly 600 annually to about 70. The agency is reviewing TILA/RESPA integrated disclosure requirements under a new executive order, potentially moving to materiality-based standards. Disparate-impact enforcement is deprioritized.

Into that vacuum: state attorneys general. A coalition of 13 state AGs this month filed suit against a lender alleging a bait-and-switch scheme involving add-on products bundled into loan balances without clear disclosure. The action is seeking hundreds of millions in restitution.

The federal framework is thinning. The state enforcement patchwork is getting denser.

My take: Every lender operating across multiple states is now navigating different enforcement regimes, different disclosure standards, and different risk thresholds, with less federal clarity to standardize against. This is exactly the kind of problem that a fragmented compliance stack handles worst: rules that vary by state, change frequently, and carry material risk when missed. If your compliance layer is spreadsheets, tribal knowledge, or a single vendor built for the old federal framework, you feel it here. Flexible, configurable compliance infrastructure isn’t just an efficiency play anymore. It’s a risk management requirement.

Canada

Fixed rates climbing despite the hold

The Bank of Canada held at 2.25% in March and is expected to hold again at the April 29 decision. But fixed mortgage rates are moving anyway.

Government of Canada bond yields have climbed above 3%, driven by the same geopolitical tensions rattling US markets. Lenders have responded by raising fixed rates 20 to 30 basis points over the past few weeks. The best 5-year fixed rate through brokers is now around 4.0%; at major banks, rates range from roughly 4.3% to 4.9% depending on term and insured status. The spread between the policy rate and what borrowers actually pay keeps widening.

For the 1+ million homeowners facing renewal this year, locking in at current fixed rates means materially higher payments than the 2020–2022 cohort signed up for, before any further policy rate movement.

My take: This week’s lesson for Canada: the Bank of Canada doesn’t control fixed rates. Bond markets do. And bond markets right now are controlled by geopolitics. Lenders who built their borrower communication and renewal workflows around rate stability are discovering that “hold” doesn’t mean “quiet.” Proactive outreach, scenario modelling, and fast product switching matter most precisely when rates are moving in ways no one predicted two weeks ago.

Ottawa in the market: $30B of CMBs

A detail worth understanding: Ottawa confirmed in January that it will purchase up to $30 billion in Canada Mortgage Bonds through 2026. The government is actively buying CMB primary issuances to keep a floor under fixed-rate mortgage liquidity.

This is the federal government functioning as a direct participant in mortgage infrastructure, not just a regulator, but a buyer. The CMB program keeps the spread between government bond yields and mortgage funding costs tighter than it would be without intervention. It’s one of the structural reasons Canadian fixed rates haven’t spiked as hard as they might have given bond market conditions.

My take: Ottawa is running a quiet systems play here. By acting as a guaranteed buyer of mortgage-backed liquidity, the government is absorbing volatility that would otherwise land directly on lenders and borrowers. It’s worth knowing this mechanism exists, because it’s also fragile. Suppose the fiscal picture changes or political priorities shift, the floor moves. Lenders whose business model depends on this spread remaining stable should consider what their funding stack would look like if that backstop narrows.

The bottom line

This was a week where the pause got all the headlines and the architecture got none of them.

Rates swung 50 basis points in a week on political news. The tariff pause calmed markets for most countries and escalated them for the largest one. The federal compliance framework is thinning, and state enforcement is filling the gap. In Canada, a government bond-buying program is the only thing keeping fixed rates from fully repricing in response to geopolitical risk.

None of these are things you can control. All of them test the same thing: whether your operation is built to absorb chaos or to assume it won’t arrive.

The pause didn’t resolve anything. It just reset the clock.

See you next Monday.

Chris Grimes is the founder of FundMore, an AI-native loan origination platform. FundMore builds agentic mortgage and lending infrastructure for institutional clients across Canada and the US.