How to Use AI as an Underwriter Without Losing the Judgement
How to use AI as an underwriter: document review, drafting decision letters, a worked example, and the risk decision that has to stay a named human call.
A lot of underwriting is reading: financial statements, applications, inspection reports, medical records, prior claims history — pulling out the handful of facts that actually change a decision from pages of material that mostly do not. That reading and summarising is exactly what a general-purpose AI tool is good at. The decision itself — approve, decline, price, on what terms — is a judgement call that carries regulatory and financial responsibility, and that stays with a named underwriter, the same as it always has.
This sits alongside how to use AI as an insurance agent and how to use AI as a banker — other roles where the paperwork can move faster and the underlying risk decision cannot be handed off.
Where AI genuinely helps
- First-pass document review. Feed in an application packet or a set of financial statements and ask for the specific figures that matter for this decision — debt ratios, coverage gaps, claims frequency — pulled out and listed, rather than reading the whole file cold.
- Flagging inconsistencies. Comparing figures across an application and its supporting documents — does the stated income match the attached statements — is exactly the kind of side-by-side check a model is good at when both documents are right there.
- Plain-language summaries. Turning a dense inspection or medical report into a short summary of what actually affects the decision, for your own use or for a broker who needs to know what is missing.
- Drafting routine correspondence. The request for missing documentation, the explanation of what additional information a file needs — high-volume writing that eats time without needing your judgement on every sentence.
- Explaining a decision in plain language. Once you have made the call, drafting the explanation of why in language the applicant or broker will actually understand, for you to check and send.
General-purpose models have got measurably better at exactly this kind of structured, document-grounded task in a short span of time — the Stanford AI Index tracks that capability curve year over year. It is also worth knowing the market is already pricing this skill: PwC's 2026 Global AI Jobs Barometer found a 62% wage premium for workers who can use AI skills well, up from 57% the year before, across close to a billion job ads.
Where it stays your call
The approve-or-decline decision, the price, the terms — that is the judgement your role and your licence exist for, and it is where fair-lending and anti-discrimination rules bite hardest. A general AI tool has no visibility into which factors your regulator allows you to weigh and which ones it does not, and it has no accountability if a pattern in its output turns out to correlate with a protected characteristic. Treat anything it produces here as a first draft from a junior colleague who has read the file quickly — a reasonable starting point, never the decision itself.
The data and documentation rule
Applicant financial details, health information, and anything else in a file does not go into a general consumer AI tool. Once submitted, you have limited visibility into how it is retained, who can access it, or whether it trains a future model. Deciding in advance what a tool may and may not see is the core idea behind NIST's AI Risk Management Framework — treat it as a checklist for what stays out of a general tool, not an afterthought for once something has already gone in.
There is a second reason to be careful specific to underwriting: a decisioning process has to be explainable to a regulator and, in many jurisdictions, to the applicant. A summary or a suggestion an AI tool generated is fine as a draft you reviewed and adopted as your own reasoning. It is a real problem if the actual reasoning behind a decision lives only inside a tool's output that nobody can fully account for afterward. Keep your own documented reasoning as the record of the decision, with the AI-drafted material as an input to it, not a replacement for it.
A worked example
Weak: "Review this application."
Better: "Here is an application packet with the identifying details removed: stated income, three years of financial statements, and a claims history summary. List every figure relevant to a standard affordability and risk assessment — debt-to-income ratio, claims frequency, any gap between stated and documented income — and flag anything inconsistent between the application and the supporting documents. Do not recommend approve or decline. Just surface what I need to look at."
The second version removes identifying information first, asks for specific figures rather than an open-ended review, and explicitly keeps the decision itself out of the model's hands — it is doing the reading, not the judging. How to write a prompt that works on the first try covers that same discipline more generally: state exactly what you want extracted or flagged, rather than asking for an open-ended opinion.
Check the flagged figures against the source documents before the summary shapes your decision at all. This is a search-and-summarise task, and it is worth verifying the same way you would check a junior colleague's first pass at a file — what AI is actually bad at covers the general pattern behind why a fluent summary still needs that check.
Where this goes wrong
A confident, complete-looking summary is not the same as an accurate one — hallucination in large language models is well documented, and a missed or invented figure in a risk summary reads no differently from a correct one until someone checks the source document. Asking the same tool to double-check its own summary is not a reliable fix either: research on sycophancy has found that models trained on human approval tend to agree with a suggestion rather than genuinely re-examine it, including their own prior output. Checking an AI answer when you are not the expert is the general routine — for a risk file specifically, that means opening the actual document a flagged figure came from, not just trusting the summary that cites it.
The draft can surface what is in the file. Only you can decide what it means for this decision.
There is also a related risk worth naming for underwriting specifically: any AI-assisted process that touches a decision affecting whether someone gets coverage, credit or a job carries real exposure to producing a discriminatory pattern, even without anyone intending it — a model trained on historical data can reproduce historical bias in ways that are hard to spot from the output alone. That risk is exactly why the actual decision, and the reasoning behind it, has to stay documented and owned by a named person rather than delegated to a tool's summary. Checking an AI answer when you are not the expert applies here with extra weight, given what is riding on the decision.
What to do this week
- Find out what your organisation has already said about AI tools and applicant data — many now have an explicit policy, and it is worth reading before assuming anything is fine to paste in.
- Pick one document-heavy step in your process — the first-pass read of an application packet — and try it with anonymised data, checking the output against the source documents before trusting the pattern.
- Write down, in one line, which parts of the decision never move to a general AI tool untouched — the approve-or-decline call and the pricing are the obvious ones to name explicitly.
- Ask what your compliance or risk team already knows about AI use in underwriting specifically — this is a faster-moving area of regulatory attention than most, and a five-minute conversation now beats finding out the rules changed after the fact.
The short version
Use AI for reading and flagging what is in a file, drafting routine correspondence, and explaining a decision you have already made in plain language. Do not use it to make the decision, and do not put applicant financial or health details into a tool your organisation has not reviewed. The reading gets faster. The judgement, and the accountability for it, does not move.
Coursium teaches this kind of practical AI use directly — writing a request precisely enough to be useful, and the habit of checking the result before it shapes a real decision. Stay ahead of AI by learning the tools on your phone.