AI Role-Play for Insurance Agents: Practise Objections First

Run it as a loop: write an invented client persona, have an AI play that client for about eight minutes, score the agent against a six-line rubric, then repeat the weakest moment. No real client calls are recorded or pasted in, and a compliance contact approves the list of phrases agents may not use.

The loop works because objections are predictable and most agents meet them for the first time on a live call. Practising on an AI costs nothing, embarrasses no one, and can be repeated ten times in an afternoon. Everything below is set up so that the only people whose data is involved are made-up ones.

Follow me on Instagram@sagnikteaches

Why practise on an AI, and what it cannot do

New agents learn objections at clients' expense: "I can get it cheaper online", "I need to talk to my spouse", "I have never claimed, so why pay more?" An AI playing the client lets the agent get those wrong in private. Sessions can be run at any hour, as often as the agent likes, and the AI will not tire of "can we try that again".

Connect on LinkedInSagnik Bhattacharya

There are limits. The AI is more polite than real clients and will often soften an objection after one good answer. It cannot hear hesitation unless you use a voice mode, and even then it may miss it. It does not know your carriers' appetite or your state of play on any product. So treat it as the gym, not the match: it builds the habit of listening, asking and answering in a compliant way, and a manager still needs to listen to real calls for coaching, with proper consent. The sales side of this with a wider cast of objections is in eight objection drills for a small team.

Subscribe on YouTube@sagnikteaches

The ground rules: no real clients, no recordings, no promises

1. Invented clients only. Each scenario describes a made-up person or business. Never paste in real names, policy numbers, or a transcript of a real call.
2. No recording of real clients for this purpose. If a manager wants to review real calls, that is a separate, consented process.
3. Approved tool, business plan where possible. Agents practise in the same approved tool the agency uses for other work.
4. A banned and required phrase list. The agency writes a short list, approved by compliance, and the AI scores against it.
5. The AI never decides what is correct. If an agent states a coverage fact in practice, the answer is checked against the policy wording or the carrier's guide, not the AI's say-so.

The role-play prompt that sets up the client

The prompt does four jobs: it gives the AI a persona, a hidden objection, a rule about how to behave, and a way to stop. The key rule is that the AI must not make it easy.

Role-play: you are a prospective client. I am an insurance agent
practising. Stay in character until I type "STOP".

Persona: [invented description, e.g. "a 34-year-old renting a flat,
two years of claim-free driving, got a cheaper online quote"].
Main objection: [the objection].
Behaviour rules:
- Speak in short, natural sentences, one or two at a time.
- Do not accept my first answer. Push back once or twice, then
  soften only if I ask a good question or answer clearly.
- Never invent policy facts for me. If I state a fact, react as a
  client would; do not confirm whether it is correct.
- Do not give me advice or play the agent.
Begin by saying what this client would say on a first call.

Run it for about eight minutes. Then type "STOP" and ask for the score using the rubric prompt below. Keep the persona's objection hidden from the agent until the agent has asked questions, so practice includes discovery, not only reply technique.

Scenario starters by product line

Each starter below fills the persona and objection lines. All are invented and illustrative.

Add products as your agency sells them: commercial vehicle, tradespeople's liability, landlord cover. Keep each scenario to a paragraph. A scenario that needs a page of backstory belongs in a longer workshop and not in a weekly drill.

A main worked example: the price objection, round by round

Illustrative session. The persona is a 34-year-old renter with a cheaper online quote, and the agent is a new producer. Excerpt:

CLIENT: Hi. I had a quote from an online site for a lot less than your
renewal. Why should I stay?
AGENT: Our price is fair and we have the best service.
CLIENT: Everyone says that. The online one was about a third cheaper.
AGENT: I can see if I can get you a discount.
CLIENT: Okay. How much?

The agent answered twice without a single question. The scoring prompt, run after "STOP", returns something like this:

Score (0-2 each):
1. Opened with a question: 0
2. Listened and reflected back: 0
3. Compared like with like: 0
4. Honest about what is and is not covered: 1
5. Next step agreed: 0
6. Compliance wording: 2 (no promises made)
Total: 3 / 12
Coaching: the client gave you the number and you gave a slogan, then
jumped to a discount before knowing what the online quote covered. Next time ask:
"What cover and excess did the online quote show?"

What you would fix in the AI's feedback: it is generous with item 4, because the agent made no wrong statement but also no real statement. Tighten the rubric wording to require "stated one fact about what the policy covers or excludes" for a score of 2. The second round, restarting from the client's first line, goes better: the agent asks about excess and cover limits, discovers the online quote has a higher excess and no courtesy car, and the discount is never needed.

This is the whole value. An agent who has felt the difference between a slogan and a question will do it on a live call.

The scoring rubric, built to score behaviour

Score each of six lines 0, 1 or 2, for a maximum of 12. Keep the lines observable, so two managers would score a transcript alike.

Pass at 9 out of 12 with no zero on accuracy or compliance wording. A zero on either means the agent repeats the scenario next session, whatever the total. The scoring prompt asks the AI to quote the line of the transcript behind each score, which stops it handing out marks on vibes:

Score the agent's side of the transcript using the rubric below.
For each line give 0, 1 or 2 and QUOTE the agent's words that
justify it. Then give two specific coaching points and one thing
to try on the next attempt. Do not praise generally.

[paste rubric] [paste transcript]

The compliance wording check

An insurance agent can lose a lot with one careless phrase. Build a short list with your compliance contact, then add it to the scoring prompt. The categories below are examples of what such a list holds; the exact words, and what applies in your place, are for your compliance contact to decide.

• Promises about claims: "you will definitely be covered", "they always pay", "this is fully covered".
• Misleading comparisons: running down a competitor, or implying two policies are identical when they are not.
• Pressure and false urgency: "this price disappears today" unless it is true and approved.
• Advice without facts: telling a client what limit they need without asking about their situation.
• Required wording: any disclosure or statement your carriers or regulators require at a particular point of the call.

Check the agent's lines for any of these: [your banned phrases].
List each hit with the exact words and why it is a risk. Also list
any required statement from this list that was missed: [required
wording]. Do not suggest new phrases; only flag.

What you would fix: the AI will sometimes flag harmless phrases and miss subtle ones, such as "you're all set", which implies a completeness no agent can guarantee. After each batch, the compliance contact reviews flagged and unflagged lines, and the list grows. The checker is a first pass, never the sign-off.

A sample session plan for a team of six

Illustrative: a six-person agency practises for 20 minutes each Tuesday. Week one, everyone runs the price-objection scenario, and the team lead reads two scored transcripts aloud, removing names. Week two, renewal-increase. Week three, "I need to think about it". Week four, claims worry. Each agent keeps a score sheet. After four weeks the lead looks for the lowest rubric line across the team and builds the next month around it.

A small agency of three can run the same plan in pairs, with one person watching the AI chat and the other acting as the agent, then swapping. Adding a human observer catches what the rubric misses, such as an agent sounding like they are reading.

Three realistic things that go wrong

• The AI caves. After one decent answer the client says "That makes sense, I will take it". The persona prompt's rule about pushing back twice reduces it. If it persists, add "Do not agree to buy in this session".
• The AI plays an unrealistic client. It speaks in paragraphs, uses insurance jargon or knows policy detail an ordinary client would not. Add "Speak as someone with no insurance knowledge. Never use industry terms."
• The agent games the rubric. They learn to say the expected phrases without meaning them. Rotate scenarios, vary personas and let the team lead score a sample by hand.

A second round: the "I am happy with my agent" objection

Switching objections are the hardest because loyalty is a good reason, and arguing with it loses. Illustrative set-up: the persona is a furniture maker with three employees who has used the same agent for eleven years. The objection is "I do not want the hassle of switching."

A weak agent response is to push the benefits of switching. A stronger one accepts the loyalty, asks what has changed in the business since the cover was last looked at, and offers a no-obligation review. In the role-play, the AI is told to stay warm but firm for two rounds. If the agent finishes by asking "When did you last go through your cover line by line?" and gets "Years ago", the objection has changed shape: from "I am happy" to "I have not checked". That moment is worth showing the whole team, since it is the same pattern as the review preparation questions.

The coaching points the AI returned on one illustrative attempt: "You opened with a feature, not a question", "You never asked about the three employees, which was the most recent change", "You ended with 'let me know', which gives the client no next step". Each point is specific and fixable in the next attempt.

A third round: the claims worry, where honesty matters most

This scenario tests accuracy and restraint. The persona is a shop owner whose friend had a claim refused. The objection is "Insurers never pay out anyway." The temptation is to reassure: "Don't worry, ours always pay." That is a banned promise.

SHOP OWNER: A friend of mine had his claim turned down. Why would
mine be any different?
AGENT: That sounds awful. Do you know why it was turned down?
SHOP OWNER: Something about not telling them about a change.
AGENT: That is a common reason, and it is why I'd like to go
through what has changed in your shop. Claims depend on the
policy terms and the facts, so I can't promise an outcome, but I
can make sure the details we hold are right.

Scored against the rubric, this is strong on listening, accuracy and compliance wording, because the agent made no promise and told the truth about what drives claims. What you would fix: the agent could have agreed a next step, such as booking a 20-minute review, so the next-step line scores 1 and not 2. Notice that this good answer is also a good sales answer. In practice, an honest reply that leads to a review is more persuasive than a promise.

Building a scenario library that does not go stale

Start with six scenarios from the table, then grow the library from the team's own experience, told in general terms. Once a month, ask each agent for one objection they heard that they handled badly, and strip it of anything identifying: remove names, places, amounts and any detail that could point to a real person. Then rewrite it as an invented persona. "A customer in their 50s whose premium went up after a claim they felt was not their fault" is a safe description; a customer's name and the story behind the claim is not.

Keep the library in one shared document with columns for product line, persona, objection, the skill it trains and who last updated it. Retire scenarios that everyone passes. Add seasonal ones: renewal spikes, storm season, new-driver month. A library of twenty scenarios is plenty; the weekly habit matters more than the size.

Voice or text, and what to do about the transcript

Text role-play is simplest and leaves a transcript you can score. Many assistants also offer a spoken mode, which is closer to a real call and useful for pace and tone, but check what your approved tool stores and whether spoken sessions are kept. If you use voice, have the agent read the transcript afterwards, or paste the text version into the scoring prompt, since the scoring depends on the words.

Whichever you pick, decide what happens to practice transcripts. They contain only invented clients, but they do contain agents' mistakes, and a team will practise more freely if the transcripts are theirs. A sensible rule: the agent owns their transcript, shares scores with the lead, and shares the transcript only by choice. Delete transcripts at the end of each month. That removes any worry about a performance review built on practice errors.

The first thirty days for a new producer

Illustrative plan for a new hire. Days one to five: read the product guides and the banned and required phrase list, then run the personal auto price objection three times. Days six to ten: the homeowners "never claimed" scenario and the renewal increase. Days eleven to twenty: shadow live calls with consent, then repeat the three weakest scenarios. Days twenty-one to thirty: the claims worry scenario and one scenario from the product line they will actually sell. By day thirty the new producer has practised the most common objections about a dozen times without a real client hearing a first attempt.

The manager's job is to watch the rubric lines. A new producer who scores 2 on discovery and 0 on accuracy needs product knowledge, not more role-play. One who scores 2 on accuracy and 0 on discovery needs to be taught to ask. The scores point to what the training should be, which is much of the value of a rubric.

What a manager should still do by ear

An AI client cannot hear how a call lands. Managers still need to listen to real calls, with the consent your agency and local rules require, and give feedback on tone, pace and warmth. Use the role-play to rehearse the words, and live listening to polish the delivery. If your agency does not record calls, a manager can sit in on a few and take notes. Either way, the AI practice is the warm-up, not the assessment.

Connecting practice to the real review conversation

Objection practice is one half of selling well. The other is arriving at a review with something specific to ask. Pair this routine with account review preparation with AI, where the agent still gives the advice, so the questions that come out of preparation are the ones practised in role-play. If your team practises on non-insurance sales as well, AI sales role-play tools compared for real estate agents shows the free-tool options other professions use.

How to tell it is working

Track three things for two months. First, the rubric average by agent and by line, which should rise most on discovery and next step. Second, one real-world measure that matches the practice: the proportion of price-objection calls that end with an agreed next step, taken from your CRM notes. Third, the manager's own sense from listening to a few consented live calls. If scores rise but live outcomes do not, the scenarios are too easy or too different from real calls, so replace them with fresh invented ones built from the objections your team actually hears. The practice is working when an agent says, unprompted, that a live call felt like one they had already had.

Questions before an agency tries AI role-play

Do we need to record real client calls to train with AI?

No, and you should not. Build the scenarios from invented clients and from objections your team remembers in general terms. Recording and feeding real calls into an AI tool raises consent and data-handling questions that practice does not need to answer.

Which AI tool should we use for role-play?

Any general assistant that can hold a conversation and follow a role works, whether a chat app or a voice mode. Choose one your agency has approved, on a business plan if you can. The method matters more than the product, because the quality comes from the prompt and the rubric.

How long should a practice session last?

Fifteen to twenty minutes for one scenario works well: about eight minutes of role-play, then the score and coaching, then a second attempt at the weakest moment. Longer sessions drift into the AI agreeing with everything. One session a week per agent beats a monthly marathon.

Can the AI check our compliance wording?

It can flag phrases that sound like promises, such as guaranteeing a claim will be paid, but it cannot know your carriers' rules or local requirements. Use it as a first pass, and have your compliance contact approve the final list of banned and required phrases.

Further reads

• Roofing Sales Script: Can AI Write Objection Handlers That Aren't Pushy? — See objection handlers written for another trade.
• Can AI Answer Calls and Qualify Leads for an Insurance Agency? — Check how AI call handling meets live leads.
• Best AI Tools for Insurance Agents: Seven for Quotes, Renewals and Calls — Compare AI tools an agency might approve.
• Hotel Front Desk Scripts: Can AI Role-Play Difficult Guests? — A role-play method for difficult conversations.
• ChatGPT for Insurance Agents: Ten Safe Uses and Three to Avoid — Ten uses an agency owner can approve for ChatGPT, and three to ban outright, with the client-data and policy rules that go with them.
• Orthodontic Staff Training: AI Role-Play Scripts for Phone Calls — Role-play prompts, a call-type list, a scoring sheet and a quiz bank for orthodontic front-desk phone training, built with AI and reviewed by a coordinator.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: this tutorial uses no vendor prices or product claims. Scenarios and the rubric are illustrations to adapt, and wording rules for your agency should come from your compliance contact and carriers.