Clients are now asking ChatGPT what they should be asking their insurance broker, then arriving with the output. The questions are strange, confident, and often confused. Many brokers have no clean way to answer them.
This is a regulated-firm problem, not just a broker one. The cause is the same everywhere, and so is the fix. It is not another chatbot.
The new question brokers can't answer
A recent Insurance Times briefing put a number on something brokers have felt for months. Over the past half year, intermediaries have seen a surge of odd, clearly AI-generated questions from clients and prospects. People type "what should I be asking my insurance broker?" into a large language model and bring the answer to the desk.
Far from speeding things up, it slows everything down. Handlers spend time decoding questions that don't quite map to the cover, the risk, or the wording in front of them. One regional broker said the technology was now clouding the process rather than clearing it.
The same pattern is landing on claims teams, law firms and advisory desks. Anyone whose clients can open ChatGPT is now fielding questions shaped by it.
Why a raw LLM makes it worse
The instinct is to fight AI with AI. Bolt a chatbot onto the website and let it field the questions. In a regulated setting, that usually makes the problem worse.
A general language model answers confidently whether or not it is right. It has no sight of your products, your wordings, your underwriting rules or the answers your best people already give. When it hits a gap, it fills it by guessing, and the guess reads just as smoothly as a fact.
We have a recent, public example of where that leads. A High Court judge found that a junior solicitor had "almost entirely outsourced the thinking process" to AI, which had produced invented legal authorities. The firm self-reported to the regulator. The work looked polished. It was wrong.
In a regulated business, a confident wrong answer is not a curiosity. It is a liability, an audit problem, and a trust problem all at once.
What a curated knowledge base actually changes
The fix is to give the AI something true to work from. That is what a curated knowledge base is: your own material, gathered in one place. Product wordings, training documents, underwriting guides, compliance notes, and the answers your experienced people already give every day.
The flow changes completely. A query comes in. The system finds the relevant passage in your library. The answer is built from your documents, with the source attached. This is retrieval, not invention. The model's job moves from "produce a good-sounding answer" to "find the right one and phrase it clearly."
Willis made a related point from the governance side in its latest Risk and Resilience review. Adoption is outpacing control, and the answer is transparency and accountability over what AI produces. A sourced answer is accountable. You can see where it came from and check it. A black-box guess is neither.
The proof: how CreditHire Assist does it
We build this for a living. CreditHire Assist, our credit hire product, runs on a curated knowledge base of more than 130 verified UK case law authorities. It does not freestyle.
When a handler faces an insurer challenge, the tool retrieves the relevant authority and builds the argument from it, with the citation shown. The handler sees the source, not just the output. They stay in control, and they can defend every line.
Credit hire is a hard test for this. It is legally technical, it is adversarial, and every claim is checked by the other side. If sourced AI holds up there, it holds up for a broker answering a client question or a firm answering a regulator.
How to start without boiling the ocean
You do not need a year-long programme to begin. Start with one question type that keeps landing on the desk.
- Pick the recurring question. The one your team answers over and over, or the AI-generated one you keep seeing.
- Gather the known-good material. The wordings, guidance and expert answers that already settle it correctly.
- Load it, then test the answers against your own people. Expand only once the answers hold up.
You are not replacing the broker or the handler. You are giving them a fast, sourced first draft they can trust and stand behind.
The takeaway
The AI questions are not going away. Clients will keep arriving with them, and they will keep getting stranger. The firms that cope are the ones that answer from a source of truth, with the working shown, rather than from a confident guess. That is a knowledge base, not a chatbot.
Frequently Asked Questions
- What's the difference between a chatbot and a knowledge base?
- A chatbot is a conversation layer. On its own it answers from a general model with no access to your business. A knowledge base is a curated store of your own trusted material that the AI retrieves answers from, so the response is grounded in your documents rather than guessed.
- Why do AI tools give confident but wrong answers?
- A general language model is built to produce fluent text, not verified fact. When it lacks the right information it fills the gap with a plausible guess, and that guess reads with the same confidence as something true. Without a source to draw from, it has no way to know the difference.
- Can AI safely answer client questions in a regulated firm?
- Yes, if it answers from a curated, source-backed knowledge base rather than a general model, and a person stays in the loop. The safety comes from grounding every answer in approved material and showing the source, so the output can be checked and defended.
- How does a curated knowledge base reduce hallucination?
- It changes the task. Instead of inventing an answer, the system retrieves the relevant passage from your trusted library and builds the response from it. The model is constrained to your material, and the source is attached, so unsupported claims have nowhere to hide.
- Where should a firm start?
- Pick one recurring question, gather the known-good material that answers it, load that, and test the output against your own experts before going wider. Starting narrow proves the approach and builds confidence before you scale it.
Related articles
Instant Claims Has a Trust Problem. Where Automation Actually Belongs in Motor Claims
UK insurers want instant motor claims, but AI-built fraud is rising. Where straight-through processing works, where it breaks, and what to automate first.
Where AI Has Actually Moved UK Insurance Distribution in 2026 (and Where It Hasn't Yet)
A calm read on what genuinely shifted in the first half of 2026: the Aviva and Compare the Market ChatGPT launches, the 70 percent consumer expectation signal, and the parts of the business that have not moved despite the headlines.
What Credit Hire Teams Should Take From the Pinsent Masons AI Ruling
A senior judge has said a junior solicitor 'almost entirely outsourced the thinking process' to AI. Here is what the Pinsent Masons ruling means for credit hire operations and how to design against the three failure modes it exposes.