Before AI Can Follow Your Rules, You Have to Identify Them

The standard advice on AI implementation goes something like: fix your processes, map your workflows, clean your data, update your guidelines. Then deploy.

It’s a sensible approach that describes virtually no insurer’s actual experience.

What most insurers discover mid-implementation is the deployment itself is the first real test their underwriting logic has ever faced. 

Someone configuring the system asks a simple question - what's the rule? - and the answer is never as straightforward as the documentation suggests.

The gap between documented and governing

Most insurers have guidelines - appetite statements in Word documents, triage criteria in rule sets, assignment logic configured into workflows.

But how much of this documentation reflects how underwriting actually works? Insurers preparing for technology deployments tend to discover the answer isn't clean.

Some say they don't have guidelines at all. They almost always do. At minimum, they know what they won't write. 

‍What's missing is the gradient.

Most insurers can define what's out of appetite. What they haven't formalized is what makes one in-appetite submission better than another - and which ones deserve an underwriter's time first. The hard nos are documented. Everything above that lives in people's heads.

In rare cases, it's genuinely true - there are no formal guidelines. A handful of senior underwriters built the book. They know the market, the brokers, and what good risk looks like. 

None of it is written down, and it didn't need to be, until the team started growing beyond what their judgment alone could cover.

These are both versions of the same gap: the logic that runs underwriting has never been validated against reality.

Why the gap stays invisible

The gap between documentation and practice doesn’t produce visible problems until a system demands a single, explicit answer.

When an underwriter reads a guideline document, they interpret. They adjust for context, fill gaps with experience, and apply judgment the documentation never captured. 

Submissions get triaged, risks get assessed, and policies get bound. Nothing in the day-to-day signals that the underlying logic is incomplete.

Software doesn't interpret. It requires a single explicit answer: what is the threshold, what triggers a referral, which conditions route to which team. 

So when it’s time for insurers to codify how underwriting decisions are made, they’re asking underwriters to make explicit what has always been implicit: the judgment calls, the contextual adjustments, the "it depends.” All of it needs to resolve into rules a system can follow.

That's where the gap shows up. Not in whether underwriters can work with ambiguity, but in whether the ambiguity can survive being written down.

Often it can't. An insurer that operates in a handful of regions has triage criteria around location-based risk factors - without enough locations to create meaningful differentiation.

Other rules cast such a wide net they flag more than half of all submissions. A few rules fire less than one percent of the time, and not even for hard declines.

The scoring was supposed to separate a book into clear tiers. Instead, everything landed in the same narrow band.

Human judgment had been doing the differentiation work all along, reading between the lines of criteria that reflected how underwriters thought about risk, but were never precise enough to stand on their own.

Nothing forced the question until now.

When scale forces the question

Undocumented logic works until it needs to scale. 

At some point, the team outgrows what proximity and institutional memory can hold. New underwriters need to learn how decisions get made, but the knowledge that helped build the book was never written down.

Quality variance increases. The book drifts from the strategy that built it. And the insurer faces a question it's never had to answer: how do we formalize what we've never articulated?

For insurers at this inflection, the technology deployment is often the first structured opportunity to answer that question. 

The exercise of configuring a system to follow underwriting logic produces something the insurer didn't have before: a validated, explicit account of how decisions actually get made.

Capgemini's 2026 P&C report draws a sharp line between the insurers getting this right and everyone else. The top 10 percent - what the report calls "intelligence trailblazers" - don't treat process clarity as a box to check before deploying AI.

They treat strategy, technology, and organizational readiness as simultaneous, reinforcing investments. The other 90 percent sequence them: fix first, deploy second, hope it holds.

The difference shows up in outcomes. "Trailblazers" aren't just deploying faster. 

They're building an organizational capability that compounds, with each implementation cycle producing clearer logic, tighter alignment between documentation and practice, and a team that understands its own underwriting discipline well enough to scale it.

What implementation actually produces

The insurers extracting the most from AI implementation aren't the ones who had clean processes going in. They're the ones who recognized the implementation itself as the diagnostic.

Every deployment surfaces the same questions. Where does documented logic diverge from practice? Which criteria actually differentiate, and which exist because nobody revisited them? Where does expertise live in people's heads rather than in systems?

Conventional advice says insurers need these answers before they deploy. In practice, deployment is what produces them.

The organizations that treat this as a feature rather than a setback emerge from implementation with something more valuable than a technology platform: a validated understanding of how their underwriting actually works, documented precisely enough to scale, audit, and improve.

That understanding was always the harder problem.

‍