Conversational AI in Insurance and Where the Rules Reach

Most people’s first encounter with AI in insurance is a chat window. You ask about a quote, or report a fender-bender, or check where your claim stands, and something answers instantly. That something is conversational AI, and it is now doing a large share of the industry’s front-line contact. Customer service is one of the operational areas an examiner counts in Exhibit A.

Conversational AI in insurance, job by job

Conversational AI is software that responds in natural language by text or voice. The interface hides three different systems. A rules bot retrieves approved answers and follows a fixed tree. A language-model bot generates a response from instructions and retrieved material. A transactional agent can also update a record, submit a claim, bind a step in a quote, or trigger another workflow. The same chat window can move through all three, while the governance hub supplies the controls shared with other uses.

Whichever system sits behind it, the work clusters around four jobs. One job is quoting: a prospect answers a few questions in a chat and gets a price. Another is intake, where a policyholder reports a claim and the bot captures the first notice of loss, the details a person used to type into a form. A third is service, fielding the routine questions, coverage, deductibles, billing, that make up most of a call center’s volume. The last is tracking, telling a customer where a claim or application stands without anyone picking up a phone.

What these jobs have in common is volume. Their judgment requirements are different. Claim status retrieval can be tightly bounded; interpreting a coverage exclusion cannot. Carriers therefore need to govern the authority behind the window, not the appearance of the window itself.

The clearest example: Lemonade’s bots

The cleanest end-to-end case is the insurtech Lemonade, which built its customer experience around two bots. Maya runs the quote-and-signup conversation. Jim runs claims.

As of the end of 2025, Lemonade reported that AI Jim takes the first notice of loss without human intervention 96 percent of the time, and that roughly 55 percent of its claims were automated end to end.¹ Lemonade is reporting on Lemonade here; the figures sit in its own annual report rather than an audit, and the same filing says many incidents are still reviewed by a person before the claim is approved. The numbers are still worth having. A carrier that builds its whole customer path around the bot can take a claim, assess it, and pay it inside one chat.

Most carriers are not Lemonade. The common pattern is narrower, a bot that handles billing questions and claim status while routing anything hard to a person. Narrower is not stationary, though. The simple, repeatable conversations keep moving to the machine, and the people keep moving to the conversations that need judgment.

Where a chatbot is low risk, and where it is not

Not every bot carries the same weight. They sort the way the rest of insurance AI sorts: by how close the thing sits to a decision the consumer feels.

A bot that confirms whether a payment posted carries less decision authority than one that interprets coverage. Both can still harm a consumer if the answer is wrong. A false statement about towing coverage may change whether someone seeks service or files a claim, even though the bot never altered the policy.

A bot that decides, or that a consumer reasonably treats as deciding, is a different matter. One that quotes a price, that determines eligibility, or that closes a claim, is shaping an outcome. When a human is no longer in the loop on that outcome, the documentation bar rises, and the questions a regulator asks about any high-stakes model start to apply to the bot. The autonomy ladder for claims specifically is the subject of agentic AI in claims and the human review test.

When a bot pulls the carrier into the rules

A chatbot feels like a website feature, so teams often leave it out of the AI governance conversation entirely. That is a mistake, for two concrete reasons.

First, the inventory. Exhibit A of the NAIC AI Systems Evaluation Tool has a carrier count its AI models one operational area at a time, customer service among them, with a separate column for the models that carry direct consumer impact.² That is a count, and it is answered from an inventory rather than being one. The inventory itself belongs to the AIS Program, which the NAIC Model Bulletin expects to cover every AI system touching a regulated insurance practice, bought as readily as built.³ A customer-facing bot that shows up in neither is the easiest kind of gap to find.

Second, notice. Section 1.9 of the bulletin asks the program to tell affected consumers that AI systems are in play, and to scale how much it tells them to the stage of the policy life cycle involved.³ That is an AI disclosure expectation about AI in general rather than a bot-specific rule, and it reaches insurers through the states that issued the bulletin as their own, about half of them so far.⁴ What a given department will accept differs, so check what your own states have issued before you write the script. Whatever the wording, a consumer being handled by software should be able to tell.

Neither point makes chatbots forbidden. It means the carrier should inventory the system, identify its authority, document applicable notice, and monitor the outcomes that matter. The AI use cases by business line guide shows where customer service connects to claims, distribution, and underwriting.

What to check if you run one

Four questions separate a governed chatbot from an exposure.

Is it on your AI inventory? If the answer is no, that is the first fix, before anything else.

Does the consumer know it is a bot? If a reasonable person could mistake it for a human on a matter that affects their coverage or claim, your disclosure needs work.

Where does it hand off to a person? A clear, logged escalation path is both a service feature and a compliance record. A bot that cannot escalate a hard case is a bot that will mishandle one.

What does it do when it does not know? A model that guesses at a coverage question is worse than one that says “let me connect you.” The safe failure mode has to be handoff, not a confident-sounding guess.

The bill for this arrives at the carrier, in two places: the exam and the complaint file. Leave the bot off the count and an examiner has somewhere to start. Let a consumer take it for a person and the complaint writes itself, in language a department that issued the bulletin already keeps on hand. The cheapest version of this problem is the one where the chat window is already on the list.

Footnotes

1. Lemonade, Inc., Form 10-K for fiscal year 2025, filed February 25, 2026: as of December 31, 2025, AI Jim took the first notice of loss “96% of the time” without human intervention, and “roughly 55% of our claims were automated.” https://www.sec.gov/Archives/edgar/data/1691421/000169142126000016/lmnd-20251231.htm . Company-reported operational figures from Lemonade’s own SEC filing, not an independent audit. ↩

2. NAIC, “AI Systems Evaluation Tool 4.0.” Exhibit A (“Quantify Regulated Entity’s Use of AI Systems”) lists Customer Service among the fifteen operations or program areas and asks for four counts per area, including “Number of AI System Model(s) with Direct Consumer Impact”: https://content.naic.org/sites/default/files/inline-files/AI%20Systems%20Evaluation%20Tool%204.0%20%28Clean%29.pdf ↩

3. NAIC, “Model Bulletin: Use of Artificial Intelligence Systems by Insurers,” adopted December 4, 2023. §1.8 (scope of the AIS Program, insurer-built and vendor-supplied alike) and §1.9 (notice to impacted consumers that AI Systems are in use, with access to information appropriate to the phase of the insurance life cycle), both p.5: https://content.naic.org/sites/default/files/inline-files/2023-12-4%20Model%20Bulletin_Adopted_0.pdf ↩ ↩²

4. NAIC, “State Adoption Map for the AI Model Bulletin,” Big Data and Artificial Intelligence (H) Working Group. Status as of April 1, 2026: the Reference List on page 2 carries 24 states plus the District of Columbia. This site’s per-state record is at /states/. https://content.naic.org/sites/default/files/cmte-h-big-data-artificial-intelligence-wg-map-ai-model-bulletin.pdf ↩