Every AI conversation in insurance starts in the wrong place.

It starts with the underwriter. Always the underwriter. Someone shows a model that can read a submission, price a risk, and spit out terms in nine seconds, and the room does the same two things it always does: the vendors say this changes everything, and the underwriters fold their arms.

Both reactions are wrong, and they're wrong for the same reason.

Because the job that's actually exposed isn't sitting in the underwriting room.

It's sitting in mine.

Start with why the underwriter survives

Not because underwriting is too complex. It isn't — a model can absolutely learn the pattern of a submission and produce a number. If you think the machine can't price a risk, you haven't been paying attention.

The underwriter survives for a reason that has nothing to do with capability:

Somebody has to be accountable for the number.

That's the whole job. Not the pricing — the ownership of the pricing. When a large account goes wrong, someone sits in a room and explains why they wrote it. When a broker pushes for terms that don't make sense, someone says no and takes the relationship hit. When the model has never seen a risk like this one — a new peril, a weird structure, a market that just moved — someone decides anyway, on incomplete information, and signs their name to it.

You cannot delegate accountability to a model. Not legally, not practically, not culturally. A model can produce a recommendation. It cannot absorb a consequence. And insurance, stripped of everything else, is an industry built entirely on who absorbs the consequence.

So the underwriter is fine. Augmented, faster, freed from grunt work — but fine.

Which brings us to the uncomfortable half of this.

Now look at what the analyst actually does

Here's the honest inventory of a typical week in my profession.

Sit in a meeting. Write up what was said. Turn a vague business request into a structured document. Format acceptance criteria. Chase people for clarification. Compare the "as-is" to the "to-be." Produce a deck summarising all of it. Send the deck. Discover half the room didn't read it. Rewrite the deck.

Now be honest with yourself about how much of that a competent model can already do.

Not "will do, in five years." Already does. Today. Passably. Sometimes better than the tired human on the fourth meeting of the day.

This isn't a threat I'm predicting. It's a description of the present tense.

And here's the part that should genuinely unsettle every BA reading this: the machine isn't coming for your judgment. It's coming for your output. And for most analysts, the output has quietly become the entire job.

The distinction that decides everything

So let me draw the line as sharply as I can, because everything in this argument lives on one side of it or the other.

AI displaces work where the output is the product. AI augments work where the output is only evidence of the judgment behind it.

Read that twice. It explains everything you're seeing.

The underwriter's number isn't the product — the decision to stand behind it is. So AI augments them. It reads the submission, surfaces the exposure, drafts the rationale, and the human does the one thing that was always the actual job: decides, and owns it.

The analyst's document, though? For an enormous number of analysts, the document is the product. The spec is the deliverable. The summary is the value. The deck is the point. And the moment the artefact is the job, the artefact is the vulnerability — because artefacts are exactly what these models manufacture at zero marginal cost.

This is the same argument I made when I wrote that I'm a business analyst who builds, not just specs — that the translation layer is always the first thing to get automated. I wrote that as a career thesis. GenAI just turned it into a deadline.

"But the AI gets it wrong"

Yes. Constantly. And this is where most analysts stop reading, fold their arms exactly like the underwriters did, and feel safe.

Don't.

Because the model being wrong isn't the defence you think it is. It's the job description you've been handed.

The model produces a confident, fluent, plausible answer. Someone has to know it's wrong. Someone has to know that the "requirement" it just wrote is generic to the point of uselessness, or that the process flow it produced skips the exception path where all the money actually leaks, or that the summary it generated is technically accurate and practically misleading.

That someone is an analyst who knows the domain cold. Which is the good news — and it's only good news if you show up.

Here's the trap: a model's output is only as trustworthy as the data and structure underneath it. I've written before that the hardest part of a reinsurance platform isn't the math — it's the data lineage, that these systems are really engines of defensible memory, and that the dangerous number isn't the wrong one, it's the right-looking one nobody can explain.

Now hand that world a technology whose defining characteristic is producing right-looking answers with no provenance whatsoever.

That's not a small risk. That's the same failure mode, on rocket fuel. A hallucinated recovery calculation doesn't crash. It doesn't error. It renders beautifully in a report and sits quietly on a balance sheet until an auditor asks the only question that has ever mattered in this business: where did that number come from?

So no — the model's fallibility doesn't make you safe. It makes you necessary. But only if you're in the loop, close enough to catch it. The analyst who refuses to touch the tool isn't guarding the gate. They've just walked away from it.

The part nobody wants to say out loud

Let's be precise about the mechanism of replacement, because "AI takes your job" is lazy and false.

AI is not going to replace you.

An analyst using AI is going to replace you.

They'll be one person doing the documentation output of three, and spending the reclaimed time where the value actually is — in the domain, in the edge cases, in the room where the decision gets made. They'll turn a two-week discovery into a two-day prototype. They'll walk into a stakeholder meeting with something working instead of something described.

And when the budget conversation happens — and it always happens — nobody is going to fund three analysts producing documents when one analyst produces the documents and the working prototype and the domain answer.

That's not a hot take. That's arithmetic.

Where this actually lands in insurance

Now put the two threads together, because insurance is a uniquely brutal test case for AI, and almost nobody frames it correctly.

I've argued that reinsurance is the most under-digitised corner of insurance — that it still runs on spreadsheets, email, and slightly different versions of the truth held by every party in a deal. Here's the punchline everyone selling AI into this industry is desperate for you to miss:

AI does not fix that. AI is downstream of that.

You cannot automate your way out of a data problem with a technology that is made of data. Point a language model at a treaty ecosystem where the participants can't agree on what was bound, when it was amended, and whose share it was — and you don't get intelligence. You get fluent confusion. Confusion with better grammar. Confusion that sounds like it went to business school.

Which is why the under-digitisation story and the AI story are not two stories. They're one. The firms that will get real value from AI in insurance are the ones who did the unglamorous work first — the data standards, the lineage, the provenance, the ability to answer where did this come from. The rest will buy a very expensive machine for generating plausible nonsense at scale, and they will call it transformation.

So what do you actually do

Stop asking whether AI will replace you. That question has no useful answer and it makes you passive.

Ask the only question that matters: what part of my job is the artefact, and what part is the judgment?

Then go and be honest about the ratio.

If most of your value is producing the document, the summary, the deck, the ticket — you're not being replaced by AI. You're being replaced by the fact that you built your career on the one thing that just became free.

And if the judgment is real — if you know where the money leaks, which requirement is load-bearing, why the recovery is wrong, what the treaty actually meant — then you have exactly what the machine doesn't and can't. But invisible judgment loses every budget fight. You have to show it. Which means you have to build with it. Which means picking up the tool, not folding your arms.

The underwriter's job survives because it was never really about producing the number.

Make sure yours was never really about producing the document.

Because the analysts who lose their seats in the next five years won't be the ones the machine outsmarted.

They'll be the ones who were still finishing a beautiful requirements document while someone else, sitting one desk over, shipped the thing it described.