A new homeowner is shopping three carriers the night before closing. A driver needs coverage before they can legally drive the car home. A shop owner compares liability quotes at 9pm because that's when the shop finally goes quiet. Insurance buying is deadline-driven, and the person on the other end usually fills in several forms and goes with whoever answers first.
That's the case for an AI chatbot for insurance agencies, and it's a real one. But we'll say our position early, because it shapes everything below: the bot should answer questions and collect details. It should never quote, recommend or bind. The moment a conversation needs a licence, it hands over to a person.
Written by the ChatterMate team. We build an open-source AI support agent, so we have a stake here.
What the bot is actually for
Look at what lands in an independent agency's inbox and voicemail, and most of it isn't a sale. It's a certificate-of-insurance request from a contractor who needs it by noon. It's "who do I call to report a claim." And somebody always wants to add a teenage son to the policy. Your producers handle these between quotes, and each one interrupts something more valuable.
Three jobs fit a bot well.
1. After-hours intake
Say a visitor lands on your site at 8pm. The bot asks what they're shopping for, then gets the things a producer would ask first: line of business, ZIP, current carrier, renewal date, a number to call. The producer opens a complete request in the morning instead of a form that says "need a quote."
2. Service questions with fixed answers
Certificate requests. Claims phone numbers. Office hours. Whether you write in Ohio. None of that changes from one caller to the next, and all of it already sits in a document somewhere. Point a bot at those documents and it answers in seconds, with the page it used. Our guide to training a chatbot on your own docs covers how.
3. Routing
Commercial risks, urgent bind dates, a claim in progress. The bot's job is to recognise these and get a human on it fast, with the conversation attached.
The line the bot must not cross
This is the part most chatbot pages for insurance skip, and it's the part that matters.
In most US states, selling, soliciting or negotiating insurance requires a producer licence. Regulators have drawn that line for human staff in detail. Pennsylvania's Insurance Department put this in writing in a 2013 notice. Ask an unlicensed employee for advice, it says, and a licensed producer has to answer. Recommending policy changes, discussing specific coverages and starting a sale all count as licensed work. It also says that when something requires a licence, the unlicensed person should tell the consumer they are being transferred to a licensee (Pennsylvania Bulletin, Notice 2013-09). It's older than chatbots and never mentions them. We're not claiming it governs yours. We're saying the logic is easy to carry over: if a person at your front desk couldn't say it without a licence, don't let your bot say it.
In practice that means the bot can say "liability coverage generally protects you if someone else is injured or their property is damaged." It shouldn't say "you should carry a $1 million umbrella." It can collect a vehicle's year and make. It shouldn't produce a number and call it a quote. We're not lawyers, and rules vary by state, so check yours before you write the bot's instructions.
There's a second layer. The National Association of Insurance Commissioners adopted a model bulletin on insurers' use of AI systems, and its adoption map, current to 31 August 2026, lists 26 jurisdictions that have adopted it: 25 states and the District of Columbia (NAIC adoption map). The bulletin is written for insurers, not for agencies, so it may not bind you directly. But your carriers are subject to it in those states, and carrier contracts and appointments tend to travel downhill. Ask your carrier reps whether they have expectations about chatbots on your site before they ask you.
What we're deliberately not citing
If you search for this topic you'll find a lot of confident statistics: percentages of leads lost, percentage jumps in conversion, billions saved. We looked at where the speed-to-lead numbers come from. A widely repeated one, that responses slower than five minutes cut your odds of qualifying a lead by 80%, appears in industry write-ups attributed only to unnamed "studies," with no sample size or date. Another says 78% of customers buy from the first company to respond, again with no source. Many of these figures come from vendors who sell response-tracking or chat tools.
So we won't hand you a number. The direction is plausible and you can test it yourself: pull your own quote requests from the last 60 days, note the time each arrived and the time someone replied, and see which ones became policies. That's a better argument for a bot than any statistic here, and it uses your data.
What to put in its knowledge base
Roughly in order of payoff:
- Service procedures. How to request a certificate, change an address, add a vehicle, report a claim. Include the carrier claim numbers.
- Lines you write and states you're licensed in. So the bot can politely redirect a prospect you can't help before anyone spends time on them.
- Office hours and who to call for what.
- The intake questions for each line. For auto: ZIP, vehicle, drivers, current coverage. For home: address, year built, coverage amount. For commercial: business type, revenue band, employee count.
- Plain-language definitions. Deductible, premium, endorsement. Approved wording only, written by your people.
Leave out carrier rate sheets, underwriting guidelines and anything you wouldn't hand a stranger on paper. A bot that can see something can be talked into repeating it, which we covered in the prompt injection post.
Handoff is the feature
For an agency, a good handoff matters more than a clever answer. When the visitor says "I need coverage by Friday" or "someone hit my car," the bot should stop answering and route to a licensed person, say what happens next ("a producer will call you today before 5") and pass the whole transcript. Nobody should have to repeat a claim story to a stranger. We covered what separates a handoff people accept from one they resent in our human handoff guide. And tell the visitor early that they're talking to an AI; our post on chatbot disclosure explains why that's worth doing even where it isn't required.
Think about data too. Prospects type addresses, birth dates and vehicle details into a chat box. If that makes you uneasy about sending transcripts to a third party, self-hosting is an option. ChatterMate is open source and self-hostable, and our self-hosted chatbot guide explains what that involves.
A first month that won't embarrass you
Week one: read the last two months of inbound emails, calls logged and web forms. Group them by question. Count; don't guess.
Week two: write the answers down properly. If the certificate process only lives in one CSR's head, that's the first thing you'll find out.
Week three: load the documents, put the widget on your contact and service pages only, and route handoffs to a monitored inbox.
Week four: read every conversation. Fix the document, not the bot. Then decide whether to widen it. The staged approach is laid out in our chatbot rollout guide.
What it won't fix
A bot won't make a slow agency fast if the follow-up after handoff is slow. If a producer takes two days to return a call the bot promised for today, you've made the problem visible, which is useful but not pleasant. And it won't replace the relationship. Clients stay with an agent because someone picks up when the claim happens. The bot's job is to clear the certificate requests and office-hours questions so your people have time for that call.
Try it on your own documents
ChatterMate is open source and free to start: the Free plan includes 50 AI replies a month, and you can self-host it. Load your service procedures and licensed-states list, then ask it the ten questions your CSRs are tired of answering. If an answer is wrong you'll see which document it used. Start at chattermate.chat. The same pattern works in other back-office-heavy businesses; see our post on staffing agencies.



