What AI Can — and Can't — Do for Insurance Brokers
Something surprised me when we started building structured intelligence for insurance. I expected the hard part to be reading the documents — parsing dense legal language, handling endorsements, dealing with inconsistent formatting. That was hard, but it wasn't the hardest part. The hardest part was knowing where to stop. Because the moment an AI starts making recommendations instead of surfacing evidence, it crosses a line that matters in regulated advice. Here's where I've landed on what works and what doesn't.
Where AI genuinely helps
Document comparison is the clearest win. A human reading two 80-page wordings side by side will miss things — not because they're careless, but because the volume is beyond what manual attention can sustain. AI reads every clause, normalises the terms, and surfaces the differences. That's not a nice-to-have; for complex commercial placements, it's the difference between catching a gap and missing one. Term normalisation is another. The same concept — 'business interruption indemnity period', 'loss of rent period', 'period of insurance for BI' — appears under different names across different wordings. Structured intelligence maps them to the same concept so comparisons are apples-to-apples. Gap detection follows naturally: once terms are normalised and clauses are mapped, you can surface what's missing or what changed between versions. Every finding traces to the clause.
Where AI falls short
Client empathy. Risk appetite. Commercial negotiation. The conversation where a broker says 'I know the cheaper policy looks fine on paper, but given what you told me about your expansion plans, I'd recommend the one with broader BI cover — here's why.' No model produces that. It requires knowing the client, understanding their business, and making a judgement call about what matters more. I've also seen AI struggle with market context — knowing which insurers are tightening appetite, which ones are flexible on endorsements, who's likely to pay a claim without a fight. That knowledge lives in the broker's experience, not in a document.
Being honest about the limits
I think the companies that are honest about what AI can't do will earn more trust than the ones that promise it can do everything. In insurance, trust is the product. If your AI tool claims to 'automate advice', a thoughtful broker should be sceptical — and a regulator definitely will be. The better framing is: AI does the reading, the broker does the advising. The system shows its working, the human makes the call. That's not a limitation of the technology. It's how regulated advice should work.
