There is almost certainly a segment of your portfolio that looks fine on the surface, consistent premium, familiar clients, no obvious drama, that is quietly underperforming in a way which only becomes visible when the numbers are arranged differently. Not catastrophically. Just steadily, year after year, eroding margin you could have redirected somewhere with better economics. Most insurance professionals never find it, not because they are not capable, but because the analysis required has historically demanded a consultant, an analytics function, or a Tuesday afternoon that did not have three other things competing for it.
McKinsey’s analysis of generative AI in financial services puts the potential annual value of gen AI across banking and insurance at over USD 1.1 trillion, with claims, underwriting and policy work flagged as among the highest-impact applications. The professionals who are realising any of that value are not running data science projects. They are running portfolio reviews on a Tuesday afternoon with a tool they already had access to.
The data you already have, in the form AI needs it in
The starting point is not new software. It is a flat table of what you already know about your book: policy type or product line, client segment, annual premium, claims paid (last three years), claims frequency, average severity, and renewal rate. Pull that out of your broker management system, AMS, or whatever dashboard your team runs. If it is messy, that is fine. AI is unusually good at working with messy structured data, as long as you tell it what each column means.
ChatGPT and Claude both handle this well in a regular chat. For broader portfolio analytics with industry-grade models built in, Verisk, Guidewire, and Shift Technology offer products with specifically-trained risk and pricing models. The general-purpose tools are the right starting point. The sector tools are where most teams move to once they know what they are looking for.
The prompt that surfaces the blind spot
The instruction that produces a real commercial briefing, rather than a paraphrase of your own data, looks like this:
I am a [BROKER / UNDERWRITER / AGENCY PRINCIPAL] reviewing my book of business. Below is a summary table covering policy type or client segment, annual premium, total claims paid over the last three years, claims frequency by segment, average claim severity, and renewal rate. [PASTE TABLE]. Do the following analysis and present each section as a short, plain-language commercial briefing rather than a number dump:
1. Identify which policy types or segments have the strongest and weakest relationship between premium and claims cost (loss ratio direction matters more than the absolute number).
2. Highlight any segment where renewal rate is declining year-on-year and suggest two plausible commercial reasons.
3. Flag any segment where claims frequency is high relative to premium charged, which could indicate a pricing or risk-selection issue.
4. Identify any segment that looks profitable on a 12-month view but deteriorates over the 3-year window.
5. Give me the two or three findings you consider most commercially significant, why each one matters, and one specific action I could take this quarter to address it.
What comes back is not a verdict. It is an interpretation, with the patterns ranked, written in language you can take into a conversation with an underwriter, a principal, or a client. More than once, brokers running this exercise on their own book have found that the product line they assumed was their stable engine was the part most quietly eroding margin.
The counterintuitive finding that keeps showing up
The segments most often flagged as the real problem are not the ones that produced the obvious large losses. They are the segments with low-severity but high-frequency claims, in lines of business where the loss ratio sits in a “looks acceptable” band that does not trigger any internal review. The drag is in the operational cost (handling 40 small claims a year) and in the slow upward creep of frequency that nobody noticed because nobody looked at the trend.
This is also the kind of pattern AI is unusually good at noticing, because it is reading every column at once rather than scanning the row that is on fire. A worked example, lightly disguised, that one broker shared after running this analysis on their commercial book: trade-account public liability looked fine on a single-year loss ratio, but frequency had risen 18% over three years while average premium had risen only 3%. The fix was not to walk away from the segment. It was to retier it inside the agency structure, push for a 9% rate uplift at renewal on the loss-leading sub-segment, and decline two specific risk profiles that were generating most of the frequency. Loss ratio on that segment came back into target within four quarters.
A second prompt for the segments AI flagged
Once you have the top two or three findings, the second pass is what makes the work usable:
For the [SEGMENT NAME] segment you flagged above, write a one-page action plan I could share with [MY PRINCIPAL / UNDERWRITING TEAM / RENEWAL TEAM]. Cover: the specific commercial issue, three plausible drivers, two options for addressing it (one cautious, one assertive), the data I should ask for to confirm the diagnosis, and the way I should frame this with affected clients at renewal so it does not sound like a blanket rate hike. Plain English. No hedging. End with the one decision the principal needs to make this month for this not to drift another year.
That output is something you can send. It is also the thing that, over the course of a quarter, changes how your principal sees you, because most insurance professionals do not bring this level of analysis to a portfolio review meeting unprompted.
The failure mode to watch for
AI confidently misreading a column header. A “claims paid” column that secretly includes recoveries netted off, a “premium” column that is gross of brokerage on some rows and net on others, a “renewal rate” calculated on count not on value. Garbage in, polished garbage out. The model will not warn you that the columns are inconsistent unless you tell it to look. Always start the prompt by asking it to inspect the data first:
Before you analyse, scan the table I pasted and tell me: what each column appears to represent, any inconsistencies in units or scale, any rows that look like they may have been calculated differently from the others, and any data that is missing in a way that would change the analysis. If anything is ambiguous, ask me before continuing.
Two minutes of data-quality interrogation up front is the difference between a real briefing and a confident-sounding fiction.
What this changes about how you work
The most commercially effective insurance professionals have always been the ones who understood their own numbers. What is changing is that this capability is no longer locked behind an analytics team or a consulting engagement. A broker working independently, a small agency without a dedicated reporting function, an underwriting professional building a case for a new product, all have access to the same quality of analytical interpretation that previously required real money to produce.
The skill is not the tool. It is knowing which question to ask of which data and which segment is worth chasing once the answer comes back. That is the part that develops from doing this exercise four or five times on your own book, with the segments that turn out to matter for your business specifically.