The Underwriting Constitution: Encode Appetite Before Letting AI Act

Carriers are putting AI agents into underwriting faster than they are deciding what those agents are allowed to do. Submission intake, data enrichment, triage and pricing support are already being automated across commercial and specialty lines, and the pressure to go further is real.

That pressure is understandable. Underwriting teams are being asked to process more information, respond faster and make better use of increasingly complex data, while experienced underwriting expertise remains difficult to scale. AI agents offer a compelling opportunity to increase capacity and support underwriters by interpreting submissions, gathering evidence and helping move work through the underwriting process.

But as agents move beyond supporting tasks and begin influencing — or potentially making — underwriting decisions, the nature of the question changes.

This is no longer simply about whether AI can automate another step in the workflow. Underwriting decisions determine which risks a carrier accepts, at what price, under what conditions and with what level of exposure. The further AI moves into that decision process, the more important it becomes to define the boundaries within which it is allowed to operate.

Most boardroom conversations understandably focus on how quickly these tools can be deployed, how much capacity they can create and where they can improve underwriting performance. A more fundamental question comes first: what exactly are we delegating?

Because when an AI agent moves from gathering information or recommending an action to binding, declining or pricing a risk, we are no longer delegating only a task.

We are delegating underwriting authority.

Delegating Authority, Not Just Tasks

An agent that can bind, decline or price a risk is exercising underwriting authority. In a human organization, that authority is defined with care. Underwriters hold letters of authority, referral limits and line-specific guidelines, and a chief underwriter answers for how all of it is used. When the same authority passes to software, the definition often gets looser, because much of what makes an underwriter good was never written down.

That judgment has always moved through apprenticeship. Junior underwriters sat near senior colleagues, watched how they read a loss run, and learned which exceptions were acceptable and which were warning signs. Agents cannot learn that way. If a carrier wants an agent to act on its judgment, it has to state that judgment in a form a machine can follow and an auditor can check.

I call that statement an underwriting constitution: the written operating authority of the underwriting function, set down before any agent acts on it.

Beyond the Rules Engine

Experienced underwriting leaders will fairly ask whether this is underwriting guidelines and rules engines under a new name. Carriers have encoded eligibility rules and referral thresholds for decades. The difference lies in where the work happens. Rules engines handled the clear cases and routed anything ambiguous to a person. Agents are being deployed to work in that ambiguous middle: reading unstructured submissions, weighing incomplete information and making calls that used to require an underwriter's discretion.

A rulebook governs thresholds. A constitution has to govern discretion. That means stating the principles behind the rules, the evidence an agent needs before it exercises judgment, and the conditions under which it must stop and hand the decision to a person.

AI researchers have a name for this pairing: neurosymbolic AI. Neural models, including the large language models behind today's agents, excel at perception and pattern recognition. They can read a loss run, pull exposures out of a broker email and spot resemblances to past risks. They are weaker at applying explicit rules consistently and at showing how they reached a conclusion. Symbolic systems have the opposite profile: precise, explainable logic that cannot make sense of a messy submission.

A neurosymbolic design lets each do what it does well. The neural layer interprets the submission. The symbolic layer — which is the constitution — decides what the agent may do with that interpretation and records which clause applied. That split also answers the audit question in advance, because every decision passes through rules a person wrote and can inspect.

The Five Articles of the Underwriting Constitution

I would write it in five articles.

Article I: Appetite

What the carrier will write, by line, segment and geography, stated precisely enough for software to enforce. Many carriers keep appetite in guideline documents that each underwriter interprets a little differently. For an agent, appetite has to be a live, versioned boundary with a named owner. If a carrier cannot state its appetite that precisely, it has found its first piece of work.

Article II: Authority and Referral

What the agent may bind, decline or price on its own, and what forces a referral: premium thresholds, hazard complexity, accumulation concerns, and risks unlike anything the agent has handled before. Each referral needs a named human owner and a response time. Calibration matters in both directions — an agent that refers too much adds cost without adding capacity, and one that refers too little takes on risk nobody approved.

Article III: Evidence

The minimum information the agent must hold before deciding, defined by line of business, and the decisions it must decline to make without it. Submission completeness becomes an underwriting control. On every submission, the agent's first task is to check the evidence standard, before any judgment begins.

Article IV: Overrides

When a person may overrule the agent, who signs, and what gets recorded. This article also sets when appetite is reviewed. Override patterns are among the most valuable signals an underwriting function has. When underwriters repeatedly overrule the agent on the same class or hazard, the market has usually moved and Article I needs amending. Tracking override rates by clause turns that signal into a routine governance step.

Article V: Audit

Every decision should be traceable to the appetite clause that permitted it and to the evidence and rating inputs that informed it. Regulatory expectations are moving in the same direction. Twenty-five U.S. jurisdictions have adopted the NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, reinforcing expectations around AI governance, accountability and documentation.

A carrier that cannot explain why its agent quoted a price will struggle to defend that price.

Getting Judgment Out of People's Heads

The hardest part of the work is extracting tacit knowledge, and three sources help:

1. Decision history. Years of referrals, declinations and bound exceptions record how underwriters actually applied appetite, which often differs from what the guide says.

2. Structured sessions with senior underwriters, working through real submissions and asking them to explain each call.

3. Replay. Running past submissions through a draft constitution and comparing its outcomes with what underwriters decided. Each disagreement shows where the constitution is incomplete.

Start Here: A Bounded Pilot

The practical first step is a bounded pilot. Pick one line of business where agents are already in use or planned. Draft its constitution, then replay the last year of submissions against it and compare the results with what your underwriters actually decided.

The gaps show where your appetite lives only in people's heads — judgment still waiting to be written down.