Your Chatbot Just Did Something Only a Licensed Human Is Allowed to Do
Insurance licensing does not regulate job titles. It regulates three acts — solicit, negotiate, sell — defined so plainly that a conversational assistant performs them without anyone deciding it should.
A number of states have issued guidance on insurer use of artificial intelligence, and most of it is about governance: written programmes, model risk management, oversight of third-party data and models. That guidance is real and worth following. It is also aimed at a different problem. Producer licensing is older, simpler and sharper-edged: it asks whether the act performed required a licence, and if it did, whose. Most conversational insurance products have a considered answer to the first question and none at all to the second.
Six Interactions, Six Verdicts
The line is not where product teams intuitively place it. It is not at the point of payment, or at the word "recommend". It sits earlier, at the moment the assistant starts conferring about the substantive terms of a specific contract with a specific prospective purchaser.
General education about insurance concepts is not solicitation, negotiation or sale. This is the zone unlicensed marketing has always occupied.
Gathering information for referral to a licensed producer is generally permissible. It stops being passive when the intake starts steering.
Filtering by consumer characteristics and surfacing particular products from particular carriers begins to look like urging application for a specific kind of insurance from a specific company.
This is conferring directly with a prospective purchaser about the substantive benefits, terms or conditions of a specific contract — the statutory definition, almost verbatim.
Recommendation plus a call to apply is the core of what the licence exists to regulate. Suitability duties may also attach depending on the line.
Exchanging a contract of insurance for money is the clearest of the three acts, and the point at which no characterisation argument is available.
The Conversion Gradient Runs Straight Through the Line
There is a reason this keeps happening, and it is not carelessness. Every increment of helpfulness that moves an assistant across the licensing line also improves the funnel. An assistant that answers "what is a deductible" converts worse than one that says which of these three plans has the deductible that fits what you just told me. The product instinct and the regulatory boundary point in opposite directions, and in a weekly experimentation cadence the product instinct wins by default unless someone has drawn the line explicitly.
So the control that works is not a policy document. It is a named boundary in the system prompt and in the routing logic, tested the way any other requirement is tested, with transcripts sampled against it. Treat "did the assistant compare named products" as a regression to catch, not a judgement call to make after launch.
Four Operating Postures
Each of these is defensible. What is not defensible is being in one posture in the contract and a different one in the interface.
- What it requires
- No carrier-specific comparison, no recommendation, no application flow. Handoff on any product question.
- How it breaks
- Scope creep. The assistant is upgraded to be more helpful and quietly crosses the line without a compliance review.
- What it requires
- Intake and routing only. Compensation structure matters — per-policy compensation to an unlicensed party raises separate issues in several states.
- How it breaks
- Answering the follow-up question. Also: routing to a producer not licensed or appointed for that state and line.
- What it requires
- Transcripts retained, advertising reviewed, appointment verified before carrier-specific discussion, suitability documented where applicable.
- How it breaks
- Supervision that exists on paper only. A designated licensee who has never seen a transcript is the weak point a market conduct exam finds.
- What it requires
- Contract should be explicit about which party's licence covers the interaction and who owns the advertising review.
- How it breaks
- Both parties assuming the other handled it. This is the most common arrangement and the least documented.
Appointments Are the Quiet Second Requirement
A producer licence is necessary but not sufficient. In most states a producer must also be appointed by the carrier whose product is being sold, and licences are per state and per line of authority. A routing system that hands a Michigan health enquiry to whichever producer is available is making a licensing decision every time it fires, silently and at scale.
This is one of the few places where automation genuinely helps: an appointment and licence check against the consumer's state and the line of business, run before the handoff rather than after, is a straightforward lookup and removes an entire class of exposure. It is also the check most often missing from products that have otherwise thought carefully about the solicitation boundary.
What a Market Conduct Exam Will Ask For
- Transcripts. Whether the assistant's conversations are retained, for how long, and whether they can be produced for a named consumer.
- The script. For a generative assistant this means the system prompt, the retrieval corpus and the guardrails — the analogue of an approved sales script.
- Advertising review. Who signed off on the content the assistant presents, against which state's advertising rules for that line.
- Supervision. Who the designated responsible licensee is and what they actually reviewed, with dates.
- Complaint handling. How a consumer disputes something the assistant told them, and whether that path is distinguishable from ordinary support.
Related Reading
- The AI model bulletin and insurer governance — the governance layer that sits alongside, not instead of, licensing.
- Unauthorised practice of law and medicine by AI chatbots — the same licensure structure in two other regulated professions.
- CCPA and AI underwriting — the privacy obligations attaching to the same intake data.
Your Marketing Pages Describe the Assistant Too
"AI insurance advisor", "get a personalised recommendation in 60 seconds", "licensed-agent-quality advice" — the words on your site characterise the act as surely as the transcript does.
See every claim your site is making in one pass. Run a free scan and check each against what the assistant is actually permitted to do.
This article is general information and not legal advice. Producer licensing definitions, exemptions, referral-compensation rules, appointment requirements, advertising regulations and AI-specific bulletins vary by state and by line of business, and are changing quickly. Consult qualified insurance regulatory counsel before relying on any conclusion here.