What Walks Out With Them

The retirement wave is a deadline on your AI program, not an HR problem.

The short version (by humans, for busy humans)

You skim now, or your AI reads for you. Fair. This part was written by humans, for busy humans, in under 400 words. The full version below carries the evidence.

The actual point:

Your retirement wave is a deadline on your AI program. Treating it as an HR matter running on a separate track is the expensive mistake.

What walks out is judgment, and judgment was never written down. Ask a thirty-year underwriter how she decides and she will describe the guidelines, because the guidelines are the part she can see. The tells, the exception paths, the accounts she treats differently for reasons she has never said out loud, none of that survives an exit interview or a documentation project. Those produce the process people are supposed to follow. Watching real work on real files gets you closer to the process they use.

The path that used to carry judgment across is broken at both ends. Retirements pull from the top. At the bottom, junior seats are not getting created, and remote work took away the chair next to the veteran. Apprenticeship used to do this transfer as a side effect of sitting near someone. Now it has to be deliberate.

Not all of it is worth keeping. Thirty years in one seat also produces instincts calibrated to a rate environment that ended in 2013. The hard work is discrimination rather than collection, which is why the veterans have to be the reviewers instead of the interview subjects. It is also why judgment sitting in one person's head is a governance problem. Nobody can examine it, test whether it is still right, or check it against the underwriter in the next office.

Whoever owns the system owns the memory. We build it. You own it. The workflow, the data, and control over the inference sit with the carrier, and we plan and staff the maintenance as part of the engagement.

The practical move is to sequence the AI program against the succession plan instead of the technology backlog. Start with the workflows where the honest answer to "who else can do this" is one person's name. Those are the only ones with an input that expires.

That is the post. If you can name the two or three people whose retirement would actually hurt your book, reach out.

The full version (human on the loop — for depth, yours or your AI's)

It is a Friday afternoon in the third-floor conference room. Sheet cake, the good kind from the place on the corner, and a card most of the office signed. Your senior commercial underwriter is retiring after thirty-one years. Somebody makes a joke about his golf handicap, and he gives a short speech thanking the two people who trained him, both of whom retired years ago.

Monday, a submission comes in. A mid-size contractor, clean loss runs, in-appetite class code, priced where your guidelines say it should be priced. Your remaining underwriters look at it and see a normal risk.

He would have declined it in about eight seconds, and nothing in the file explains why. It was the broker who sent it, the way the payroll splits across two states, and a pattern he watched play out twice in the 2000s.

The delay is what makes this hard to catch. Nothing bad happens Monday. The account gets quoted, gets bound, and goes into the book at an adequate-looking rate. It surfaces eighteen months later in a claim review, as development on an account nobody had flagged as unusual, and no one connects it back to the sheet cake.

You can hire for the job. You cannot hire for the thirty-one years.

The two tracks that never meet

Every carrier of any size has two programs running right now, and they almost never touch. HR owns the succession plan, which is who backfills whom, what the bench looks like, and when each retirement lands. The AI program belongs to technology, and it tracks workflows, vendors, and what the 2027 roadmap promises. Nobody owns the handoff.

The gap usually gets filled with a knowledge-management push. Somebody stands up a wiki, asks the veterans to document their process, and ends up with a document that describes the guidelines. It goes stale in a quarter.

The scale is real even where the numbers are soft. The Bureau of Labor Statistics has projected that US insurance will lose roughly 400,000 workers to attrition by the end of 2026 (BLS projection, as reported by Insurance Business, 2026). That is a projection and not a measurement, so read it as a shape rather than a count.

The insurance-specific version is harder to wave off. NAMIC projects that up to half of today's underwriters will retire by 2028 (NAMIC, via Gradient AI, 2026).

What a decade inside a carrier taught us

We spent a decade building AI in production inside a live carrier, sitting next to the underwriters and adjusters whose work it changed. The durable lesson was that the judgment is the asset and the software is only where you keep it.

1. What walks out is judgment, and judgment was never written down

Ask a thirty-year underwriter to explain how she decides and she will describe the guidelines. She is not holding anything back. The guidelines are the part she can see, and the rest of it does not register with her as knowledge. It registers as the obvious thing to look at first.

That is why exit interviews and documentation projects produce the wrong artifact. People describe the process they are supposed to follow, because that is the process they can articulate on demand. The exception paths, the broker tells, the accounts they treat differently for reasons they have never said out loud, all of that sits one layer below anything a person can dictate into a document.

You get at it by watching the work itself. On a recent carrier engagement we pointed our own tooling at the operational record the carrier gave us access to, rather than at the people, and it read 4,850 Slack channels (public, private, and group DMs) in a matter of days. It surfaced more than 800 multi-person DM groups running as a hidden coordination layer, where real decisions were happening outside the searchable channels.

The output was 12 ranked AI opportunities, each cited back to the specific channels and message patterns that revealed it. Nobody was interviewed.

The speed is not the point. Recommendations built from the operational record come from what people actually did on live files, where nobody is performing for an interviewer.

What this means for your Tuesday morning. Pick your best underwriter, ask her to walk you through a recent decline, and count how many of her reasons appear anywhere in your guidelines. That count is your exposure.

2. The path that used to carry judgment is broken at both ends

For most of this industry's history there was one transfer mechanism, and it was apprenticeship. You sat the new hire next to the veteran for five years, gave them the easy accounts to start, and let the judgment move across by osmosis. It worked well enough for decades. What is happening now is that it fails at the top and the bottom at the same time.

The top end is the retirement math above. At the entry level something newer is happening. Finance and insurance job openings fell to their lowest monthly level in a decade in December 2025, roughly 138,000 against an annual average near 281,000 (Jacobson Group and Aon, Q1 2026 Insurance Labor Market Study, as reported by Insurance Journal, March 2026). The analysts who published that were careful about the cause, suggesting carriers may be pausing hiring to see how AI lands rather than cutting because of it. A Harvard working paper on US résumé and job-posting data from 2015 to 2025 found that junior employment at firms adopting generative AI declined relative to non-adopters, while senior employment trends stayed largely unchanged (Hosseini Maasoum and Lichtinger, "Generative AI as Seniority-Biased Technological Change: Evidence from U.S. Résumé and Job Posting Data," Harvard working paper, SSRN 5425555).

That decline came from slower hiring rather than a rise in separations, which changes how you should read it. Nobody is cutting juniors. The seats where juniors used to learn are not getting created in the first place.

A skeptical reader is already thinking the obvious thing, so I will say it. Some of this is our category's doing. AI took the entry-level work that used to fund the apprentice's chair, and remote work took the chair itself, which is a competing explanation worth taking seriously (Insurance Business, July 2026). Both are partly true, and they land in the same place, which is that the transfer used to happen as a side effect of proximity and now it has to be designed.

That is an argument for building the system rather than against it. A junior underwriter who can see why a senior one walked away from an account, in the system, on a live file, is learning something the empty chair beside her cannot teach.

What this means for your Tuesday morning. Count the junior seats you have quietly not refilled in the last two years. Then name the person a new hire would learn the exception paths from today.

3. Not all of it is worth keeping

The romantic version of this argument treats every veteran as a vault of pure signal. They are not. Thirty years in the same seat also produces instincts calibrated to a rate environment that ended in 2013, habits that survived because nobody ever audited them, and the occasional rule of thumb that stopped being true and never got retired.

So the hard work here is discrimination, not collection. When you build alongside the expert, the expert reviews the output, and the review is where the sorting happens. She watches the system flag an account the way she would have flagged it herself, and she moves on. The next one gets flagged for a reason she recognizes as a habit she should have dropped in 2015, so she says so and the system stops doing it.

Collection on its own cannot do that, which is why "capture everything before they go" is the wrong instruction.

There is a governance consequence most carriers have not connected to their succession plan. Judgment that sits in one person's head cannot be examined by anyone, including you, which means you cannot review it, cannot test whether it is still right, and cannot tell whether two underwriters are applying the same standard to the same risk. Making it explicit is what makes it reviewable at all, and reviewability is what you lose permanently when the person leaves. That is a real cost of waiting, separate from the loss of the judgment itself.

What this means for your Tuesday morning. Judge any knowledge-capture effort by what it throws away. A program that keeps everything has not evaluated anything.

4. Whoever owns the system owns the memory

The judgment you capture has to live somewhere, and where it lives decides whether it is still yours in five years. This is the least exciting of the four points and the one with the longest tail.

Our terms are the same on every engagement. We build it. You own it. The workflow, the data, and control over the inference sit with the carrier.

Much of the surrounding stack is rented and should be, since the model layer is swappable by design and today's frontier model will not be next year's. The layer worth owning is the operating layer, the one that holds how your people decide.

Ownership also does not mean we hand you software and walk away. We plan the maintenance and staff it as part of the engagement, because a system you own that nobody maintains is not much of an asset either. Most carriers have already learned that lesson the hard way somewhere else in their stack.

What this means for your Tuesday morning. On the next AI proposal you read, find the sentence that says who owns the workflow and the data when the engagement ends. If there is no such sentence, that is your answer.

So what

Most AI decisions at a carrier carry no deadline. You can defer the core-system question, the data-layer question, and the build-versus-buy question, and the cost of deferring is real but diffuse. This one has dates on it. Somebody in your organization knows roughly which quarter your most valuable underwriter hands in her notice, and that date does not move because your roadmap slipped.

So sequence the AI program against the succession plan instead of the technology backlog. Find the workflows where one person is the single point of judgment, the ones where the honest answer to "who else can do this" is a name and not a team. Those go first, whatever the roadmap says, because they are the only ones with an input that expires.

Then fund the capture as a phase of the AI program, with its own weeks and its own owner, and assign your most-tenured people to it as teachers and reviewers. They are the only ones who can tell you which of their own instincts are still true. That is the only quality control this work has.

Take one question into your next planning meeting. Which of the judgments on your AI roadmap sit with exactly one person, and how much time is left on each?

We do this work with carriers by building alongside the people who hold the judgment, on live workflows, while they are still in the building. The deliverable is a working system, and the veterans are its reviewers. If you can name the two or three people whose retirement would actually hurt your book, that is a conversation we would like to be in.