There is a moment in a renewal meeting that changes the relationship. It is the one where the broker, instead of opening with rate, says: “I have been looking at your claims history over the last five years, and I think there is something worth paying attention to.” The client leans in. The conversation that follows is not about price. It is about risk. That is the conversation worth having, and the part that has changed is how much time it takes to be ready for it.
McKinsey’s analysis of generative AI in financial services flags claims work as among the highest-impact AI applications across banking and insurance, with potential annual value over USD 1.1 trillion across the sector. The brokers and claims professionals turning that into something useful are not running data science projects. They are reading loss runs differently, with a tool that interprets rather than just summarises.
The data was always there
Loss runs, claims summaries, incident logs, large-loss reports. This information has been sitting in your management system for years. What has changed is that you no longer need a dedicated analytics team or specialist software to extract a pattern from it. A general-purpose tool like ChatGPT or Claude can read a block of claims data, hold the structure in mind, and come back with a plain-English interpretation: which claim types recur, which periods saw unusual activity, and what the pattern as a whole suggests about the client’s risk profile.
For carriers and larger brokerages, sector-specific platforms go further. Shift Technology is built for fraud detection and claims triage. Verisk and Guidewire ship claims-analytics products that draw on industry-wide loss data, not just your own book. The general-purpose tools are usually the right starting point. The sector-specific ones are where teams move once they know which questions to ask.
The prompt that produces a real briefing
Take a loss run for a single client, redact anything personally identifiable, and paste it into ChatGPT or Claude with this:
I am an insurance broker preparing for a renewal meeting with a commercial client in the [SECTOR, e.g. hospitality / manufacturing / professional services] sector. Below is five years of claims data including claim dates, claim types, descriptions, settlement values, and reserves where shown. [PASTE].
Read the data carefully and produce the following:
1. The most frequently occurring claim types, what they suggest about the client’s risk exposure, and whether the pattern is consistent with what you would expect for this sector.
2. Trends over time. Is claims frequency or severity increasing, decreasing, or flat? Is there a particular year that looks anomalous, and what could plausibly explain it?
3. Any periods of unusually high activity (cluster of claims in 90 days, repeat claim type in 12 months) and a few candidate explanations.
4. The two or three risk-improvement themes that, if addressed, would most plausibly reduce future claims frequency or severity.
5. Three specific, non-judgemental questions I could open the renewal meeting with to invite a productive conversation about risk improvement, before we discuss renewal terms.
6. Anything in the data that looks unusual or that I should ask my colleague to verify before relying on it.
What comes back reads like an analytical briefing. The kind that would have taken an experienced risk consultant most of an afternoon, or your own block of focused time, is now produced in roughly a minute. You can take it directly into the meeting, into a written risk report, or into a recommendation to the underwriter.
A counterintuitive finding worth sitting with
The single highest-value pattern AI tends to surface in a loss-run review is not the obvious cluster of large claims. It is a slow, low-severity rise in frequency that has gone unremarked because it has never breached a threshold that triggered internal review. A worked example: a broker ran this prompt on a five-year loss run for a mid-sized hospitality client that everyone, including the underwriter, considered stable. The output flagged that public liability frequency had risen 22% over three years while severity had stayed flat, with a clear cluster around weekend evening incidents. The renewal conversation that followed, opened with question three from the prompt above, surfaced that the client had changed their late-night door policy two years earlier and had not communicated it. The underwriter agreed to hold the rate flat in exchange for a documented door-management change. The renewal closed cleanly. The broker became, in the client’s words, the first one who had noticed.
That is the version of “what AI helps you see” that is actually worth having. Not the ten-minute miracle. The pattern that was always in the data and was never the main story.
From individual policy to whole book
Once you are comfortable doing this at the individual-client level, the same prompt scales upward. Pull anonymised loss-run data across a segment of your book (one industry, one client size, one product line) and ask:
Below is anonymised three-year claims data for [NUMBER] clients in the [SECTOR] segment of my book. Each row includes claim type, year, severity band, and whether it was a first-party or third-party claim. [PASTE]. Identify: which claim types are most common across this segment, whether the segment as a whole is trending up or down on frequency and severity, which sub-segments (by client size, region, or sub-industry if visible) appear to be driving the trend, and two or three risk-improvement themes a broker working with this segment should be raising proactively at renewal.
This is the analysis that turns a broker from someone who renews policies into someone who has a defensible point of view about a segment. That difference is what gets remembered when an underwriter is choosing who to back on a new product, or when a principal is deciding who to put on the larger accounts.
The failure mode to watch for
AI being confidently wrong about a column it misread. A “settlement value” column that includes recoveries, a “claim date” column that mixes incident date and reported date, a “claim type” column whose codes mean different things in different years. The model will synthesise around the misread without flagging it. Always start with a data-integrity check:
Before analysing, scan the data and tell me: what each column appears to represent, any inconsistencies in formats (dates, currency, codes), any rows where values look implausible, and any data that is missing in a way that would change the analysis. If anything is ambiguous, ask me before continuing.
The other failure mode is over-trusting a single-client run. Five years of claims for one mid-sized client is a small dataset by any analytical standard. AI is good at finding patterns in it. It is also good at finding patterns that are noise. Treat the output as a hypothesis worth taking into the meeting, not as a settled finding to present as fact. The right framing in the renewal conversation is “I noticed something in the data and wanted to ask about it”, not “the data shows”.
What it changes about the work
The brokers and claims people doing this consistently report two things. The conversations with clients shift from price to risk, which usually means renewal margins hold or improve and retention goes up. And the relationship inside the agency, with underwriters and principals, shifts as well, because they start to see you as the person who comes to meetings with a defensible commercial point of view. Both of those compound. Neither of them require any technical background. They require the discipline of asking the data the right questions, and the patience to read what the answer actually says.