AI Cross-Selling for Insurance Agencies: Account Review Prep, Not Advice

Use AI to prepare the review, not to conduct it: summarise the client file, list what has changed since the last review, and draft open questions for the agent to ask. The licensed agent then listens, advises and records the recommendation. Never let AI propose a product or say a client is underinsured.

The line is easy to blur because the same tool can do both. A prompt that asks "what should this client buy?" produces advice by machine; a prompt that asks "what has changed, and what should I ask about?" produces preparation. Everything below is built to keep the work on the second side of that line.

Follow me on Instagram@sagnikteaches

Why the preparation is where AI helps and advice is where it must not

Preparing for a review is mostly reading: policy schedules, last year's notes, the claims history, email threads. An agent with 300 accounts cannot reread each properly, so reviews become a call that starts cold. AI is good at condensing documents into a one-page brief, and an agent who arrives with the brief asks better questions.

Connect on LinkedInSagnik Bhattacharya

Advice is different. Recommending cover, comparing a replacement policy with an existing one, or saying a limit is adequate each depends on facts the file may not hold and on rules about suitability, disclosure and replacement that vary by product and place. Those duties sit with the licensed person. If an AI tool wrote the recommendation, the agent is still responsible for it, and the agency's errors and omissions insurer will ask who checked it. Keep the machine out of the recommendation, and the question never arises. For the insurer's side of this, see the scope and E&O questions when hiring an AI consultant.

Subscribe on YouTube@sagnikteaches

The review-prep routine in six steps

1. Pull the minimum file. Export the policy list for the client with types, limits, deductibles, renewal dates and claims count, plus the last review note. Strip dates of birth, identifiers, bank details and medical information before anything goes near an AI tool.
2. Ask for a summary, not a verdict. The prompt asks what is in the file and what has changed. It forbids recommendations.
3. Ask for questions. A second prompt turns the summary into open questions grouped by topic, none naming a product.
4. The agent checks the brief against source documents. Ten minutes on the policy schedules, comparing every figure in the brief.
5. The agent runs the review. The questions open conversations. The agent's answers and any recommendation come from the agent.
6. Log it. A short record: brief used, errors found, outcome, and who advised.

The summary prompt, with a sample output

You are preparing a brief for an insurance agent's annual account review.
Use ONLY the information below. Do not recommend, rank or suggest any
product or limit. Do not say whether any cover is adequate.

Return:
1. In-force policies (type, limit, renewal month)
2. Changes since the last review note
3. Claims in the last 3 years (count and type only)
4. Gaps in the information (things the file does not say)

FILE:
[paste the minimised file here]

Illustrative output for a household file:

1. In-force: home (limit as on schedule, renews March); two cars
   (renew June); personal umbrella (renews March)
2. Changes: last note records a new teenage driver; no note since
   about the second car's usage
3. Claims: one windscreen, one water damage (2 in 3 years)
4. Gaps: no record of home office use, rental of any room, valuables
   over the schedule limit, or whether the teenage driver was added

What you would fix: item 4 is the most valuable part, because it lists what the file cannot tell you. The "changes" line is only as good as the notes. If it invents a change that is not in the file, that is a hallucination; the source-check step in the routine exists to catch it. Fix the prompt by adding "quote the file line you used for each change".

The question prompt, and what separates a question from a pitch

The second prompt works from the summary and produces things to ask. The test for each output line: does it open a topic, or does it name a product? "Has anyone started working from home or running a business from the house?" is a question. "Offer a home-based business policy" is a pitch and has no place in the list.

From the brief below, write 10 open questions for the agent to ask at
the review. Group them: household changes, property and belongings,
vehicles, work and income, anything the file leaves unclear.
Rules: questions only; no product names; no statements about what the
client needs; no leading wording such as "wouldn't you want".

BRIEF:
[paste the checked brief]

Illustrative output (first four lines):

Household changes
- Has anyone joined or left the household since we last spoke?
- Has your teenage son started driving, and which car does he use?
Property and belongings
- Have you bought anything valuable in the last year, such as jewellery,
  a bike or electronics?

What you would fix: the second line assumes the driver is male and a son. Strip any assumption that is not in the file, because a wrong guess makes a client feel unheard. Another common failure is a question that is really a closed offer, such as "Do you want us to add valuables cover?". Rewrite it as "How would you replace those items if they were stolen?", which invites an answer without steering.

A main worked example: a small furniture maker's commercial account

Take a small furniture maker with a workshop and three employees, with a commercial package, a commercial vehicle policy and cover for employee injuries in force. Illustrative file, minimised: the package renews in September with a stock limit that was set two years ago; one van; no change notes since last year.

The brief says: stock limit unchanged for two years; no record of turnover, workshop contents value or any new equipment; no note of any online selling; no record of subcontractors. The questions it drafts include: "Roughly what is the value of timber and finished stock on a typical day, and does it rise before a busy season?", "Have you bought any major machinery since last year?", "Is anyone working for you who is not on the payroll?", "Do you sell directly online or deliver to customers' homes?", "Is any work done at customers' premises?"

None of these names a product. The agent hears the answers: the maker bought a CNC router for a large amount and has started doing installations at customers' homes. The agent now has real facts and can advise, using the carrier's wording and eligibility rules, on whether the package needs to change. That advice, and its record, belong to the agent. The AI contributed a brief that saved twenty minutes of file reading and a list that did not forget the contractors.

Without the AI the same agent might still have found the router. The difference is consistency: across 300 accounts the questions get asked the same way every year.

A quick sum, with assumed figures to replace: if preparation, including the ten-minute check against the schedules, falls from 25 minutes to 12 per review, and each of the agent's 300 accounts is reviewed once a year, that is 13 × 300 = 3,900 minutes, or 65 hours a year. Time five reviews the old way and five the new way before you quote any such figure, and count the source check every time: a saving that comes from skipping it is not a saving.

A second example: a car repair garage with a business-use question

Illustrative: a four-bay car repair garage insures its building contents, a courtesy car and a recovery truck. The file holds a package renewal in January and one claim for a break-in two years ago. The brief lists the three policies and flags that no note records how many courtesy cars are on the road, whether customers' vehicles are held overnight, or whether the garage has started any mobile servicing.

The questions drafted are: "How many vehicles are on your premises overnight in a typical week, and how many are customers'?", "Do your mechanics ever work on a vehicle away from the workshop?", "Who drives the courtesy car, and does anyone under 25 or with a recent conviction use it?", "Has your security setup changed since the break-in?" The agent might learn that mobile servicing began in the summer. That is a fact the policy file did not hold, and it matters whatever cover the agent eventually discusses. The tool did nothing except make sure the question was asked.

The same review without the brief

Without the brief, the agent opens the file twenty minutes before the call, sees the package and the old claim, and opens with "Anything changed?". The client says no, because mobile servicing has not seemed like a change to him. Specific questions get specific answers; "anything changed?" gets "no". That is the practical reason to prepare, and it has nothing to do with selling more.

Drafting the agenda email before the review

A review goes better when the client has seen the topics. An AI draft of a short pre-review email, based on the brief's gaps, asks the client to think about a few things in advance. Illustrative wording the agent would edit: "Hi [first name], ahead of our review on Thursday, it would help to have a rough idea of three things: any big purchases or building work in the last year, anyone new driving your vehicles, and any change in how you use your home or premises. No need to dig out paperwork. [your name]". It names topics, not products, and it makes no claim about what the client needs.

Keep the email under 80 words, send it from the agent's own address, and let the agent adjust the tone to suit the client. The AI draft is a starting point, and an agent who knows the client will cut half of it.

After the call: notes that stay factual

AI can also tidy the agent's own call notes, and the same boundary holds. Paste the agent's rough notes (no recording, no transcript of the client unless your agency has approved that and told the client) and ask for a clean, dated record of what the client said and what the agent said they would do next. Forbid additions: "Use only what is in the notes. Do not add advice, reasons or recommendations that are not written there."

A before and after. Before, the agent's note: "new router, 3 staff more, van mid-yr, said send quote for extras". After, the cleaned record: "Client reported a new machine purchase, three additional employees and a van acquired mid-year. Client asked the agent to send a quotation for changes. Agent to confirm which policies are affected." The tidy version adds nothing that was not in the note, so it is safe to file. If it had added "client was advised to increase machinery cover", you would delete that line, because no one said it.

Check the output against the original note every time during the first month. If the tool adds content, stop using it for notes. A note that records advice the agent never gave is worse than a messy note, because it creates a record of something that did not happen.

What to tell clients about AI use

Most clients will not ask, but some will, and your agency should have an answer. A short, honest line works: "We use software to help us prepare for reviews by summarising your file. Your agent reads and checks it, and every recommendation comes from your agent." Say it only if it is true, and do not claim the tool is never used for something it is used for. If your agency's insurer, regulator or contract with a carrier requires disclosure of AI use, follow that and keep a copy of the wording.

If a client asks you not to use AI on their file, honour it by marking the file, and have the agent prepare the old way. The cost is twenty minutes, and the goodwill is worth more.

Rolling it out to a team of eight

Illustrative plan for an eight-person agency. Week one: the principal and one senior agent run the routine on five accounts each, noting every error in the summaries. Week two: fix the prompts from those errors, write the rule sheet, and agree with the compliance contact what may go into the tool. Weeks three and four: two more agents join, each reviewing the others' first three briefs for steering language. Month two: open it to everyone, with the monthly ten-review sample in the diary. Resist rolling out in week one to everybody; the prompts are the part that needs two iterations, and errors discovered across eight agents at once are harder to trace.

Decide up front who owns the prompts. One named person edits them, and changes need a short note on why. Treat the question prompt as an agency document, in the same class as a standard letter, because a wording change affects every review that follows.

The replacement and suitability line, in plain terms

Two situations call for particular care, and both are beyond what a prep tool should touch.

• Replacement. Moving a client from an existing policy to a new one can disadvantage them: waiting periods restart, benefits change, surrender charges or pricing differences may apply, and many places require specific disclosures and comparisons. A prep brief may note that a client's life policy has a particular renewal or term, but it must not suggest swapping it. The agent follows the replacement rules that apply to the product and place.
• Suitability or best-interest standards. For some products, a recommendation has to match the client's circumstances, and the agent has to document why. An AI-drafted question list can support that by prompting you to ask about circumstances, but the reasoning in the file must be the agent's own.

The rules differ by product, by licence type and by place, and I cannot tell you which apply to you. Ask your licensing body or compliance contact, and write the answer into your agency's AI policy. Where the file shows a possible replacement, flag it for the agent in the brief as "possible replacement: agent to apply replacement rules" and let the human take it from there.

Data handling: what goes into the tool

A client file is personal data. Before choosing a tool, read its privacy and data-use terms (what to check in an AI tool's privacy policy lists the points), confirm whether inputs are kept or used for training, and decide whether your agency allows client details in it at all. Where a tool is not approved, you can still use the routine with a redacted file: replace names with "Client A", drop addresses, and keep only the types, limits and dates.

A before and after for one redacted line. Before: "Mrs [name], 14 [street], DOB [date], teenage son [name] added as driver on the second car". After: "Client A, home and two cars, one teenage driver added, vehicle make not needed". The brief works as well, and the file leaves nothing behind that identifies anyone.

Four ways the routine goes wrong

1. The questions turn into a script. The agent reads them in order, and the call stops sounding like a conversation. Tell agents to pick three, and to follow the client's answers.
2. Prompts drift. Someone adds "and suggest what they should add" to the prompt to save time. Lock the prompt text in a shared document and review changes.
3. Stale files. The AI summarises a file last updated two years ago and presents it as current. Add "state the date of the latest note" to the prompt, and have the agent confirm basics at the start of the call.
4. Uneven use. One producer uses it for every review and another never does. Pick one team for a month before widening.

Where this sits among the other AI uses in an agency is covered in where AI saves time in a small insurance brokerage, and the renewal side is in automating insurance renewals with AI.

Check that it worked: a monthly sample

Each month, pull 10 reviews and check five things: was a brief used, did the agent record corrections to it, did the questions asked match the client's situation, is there a recorded reason for any recommendation made, and is there any sign of a product being named by the AI output. Score each review pass or fail on those five. A pass rate below 8 in 10 means the prompt or the training needs work, not that the tool is useless.

Track two outcomes alongside: the number of questions that surfaced a change (a new driver, a new asset, a new activity), and time spent preparing. If preparation dropped from 25 minutes to 12 and the share of reviews that found a change rose, the routine works. If time dropped and the changes found fell, you are skipping the conversation.

To practise the conversation itself with colleagues, not clients, try AI role-play for insurance agents. For the email side of cross-selling, which has its own rules on consent and tone, see building an upsell and cross-sell email sequence with AI.

A one-page rule sheet for the office

Print this and pin it where agents work.

• AI summarises and drafts questions. The agent advises.
• No product names, recommended limits or coverage-gap verdicts in any AI output; gaps in the information are fine.
• Minimised file only: no identifiers, no bank or health details.
• Every figure in the brief is checked against the policy documents before the call.
• Anything that looks like a replacement goes to the agent with a flag, and the replacement rules apply.
• Log the brief, the corrections and the advising agent's name.

Questions agents ask about AI and account reviews

Can AI tell a client what cover they need?

It should not. Recommending cover is advice, and in most places that is an activity for licensed people held to suitability and disclosure rules. Use AI to organise the file and suggest questions to ask, and keep every recommendation with the agent who knows the client and signs the file.

Which client details should I give the AI?

Give the least that does the job: policy types, limits, renewal dates, claims count and recent life or business changes. Leave out dates of birth, national identifiers, bank details and health information. Check your agency's AI policy and the tool's data terms before pasting anything.

How do I stop staff using the questions as a sales script?

Frame the output as a listening list, not a pitch list. Each question should open a topic, such as whether anyone new drives the car, and never name a product. Review a sample of reviews each month to confirm the call followed the client's answers.

What if the AI summary contains something wrong?

Treat every summary as a draft. The agent compares it with the policy documents before the call, and any figure quoted to a client comes from the source document. Log errors you find, because a pattern tells you what to strip from the input.

Further reads

• How Brokers Use AI to Compare Policy Wordings at Renewal — Compare wordings at renewal with the same human check.
• Best AI Tools for Insurance Agents: Seven for Quotes, Renewals and Calls — Compare AI tools for quotes, renewals and calls.
• Can AI Answer Calls and Qualify Leads for an Insurance Agency? — See where AI call handling sits next to reviews.
• AI Insurance Agency Management System: What to Ask Before You Switch — Check what your agency system can do before adding tools.
• HubSpot for Insurance Agents: Pipeline, Renewals and What Breeze Adds — Hold review dates and notes in a CRM.
• Insurance Broker AI App: A Buying Checklist for a Small Brokerage — Use a buying checklist before approving a tool.
• AI Tools for Life Insurance Agents: Follow-Up Within the Rules — A rule-first method for AI follow-up in life insurance: guard-railed prompts, replacement flags, recording consent and a pre-send gate.
• Does Agency E&O Cover AI Mistakes? What to Ask Your Carrier — A per-tool list of questions for your E&O carrier, the exclusion wording to look for, and the written AI policy that gives you evidence. Not legal advice.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: no vendor pricing is quoted here. The advice rests on general licensing and disclosure principles, which vary by place and product; confirm specifics with your licensing body and errors and omissions insurer.