Insurance Broker AI App: A Buying Checklist for a Small Brokerage

Check six things: where client data is stored and whether it trains models, how it writes to your CRM or broking system, a log of what it says and does, a price you can model for your busiest month, hard limits on cover advice and quotes, and a handover to a person. Reject any app that cannot show a real log.

Demos show the clever part, an assistant that chats or phones a client about a renewal. The checks that decide whether it is safe to run sit underneath: what it stores, what it writes back, what it records and what it must never say. This checklist has thirty-six of them in six groups, arranged so that one afternoon with a vendor covers the lot.

Follow me on Instagram@sagnikteaches

Scope first: what the app may do and what stays with a broker

Fix the job before you look at products, because every check below depends on it. A broker's licence, and the client's trust, sit with the people in your firm. The app collects, logs, reminds and books. It does not advise. Here is a starting split for a small brokerage; adjust it to your own rules and your compliance adviser's view.

Connect on LinkedInSagnik Bhattacharya

How a voice agent handles calls and lead qualification day to day is covered in can AI answer calls and qualify leads for an insurance agency, and the renewal workflow itself, once you have an app, is in automating insurance renewals with AI. This tutorial is about choosing the app.

Subscribe on YouTube@sagnikteaches

The checklist: thirty-six checks in six groups

Each check gives the reason and a way to verify it. A check earns a point only when you have seen it work or seen it in writing. A promise on a sales call earns nothing. For the general vendor questions that apply to any AI tool touching client data, questions to ask before buying AI that touches client data is the broader list; the checks here are the ones specific to enquiries and renewals.

Data handling

1. Where data is stored. Client records and call audio may be held somewhere you did not choose. Verify: ask in writing which regions the app uses for storage and processing, whether you can choose, and whether the sub-processor list names the same regions.
2. Model training. Some tools use customer content to improve models unless told not to. Verify: get the sentence from the contract or data-processing terms, not the marketing page, and ask the same question about the vendor's own AI model providers.
3. Retention and deletion. Transcripts, recordings and summaries pile up, and you may have to delete one client's data on request. Verify: ask for the default retention period for each data type, how deletion works and how long backups keep deleted data. Delete one test conversation and ask for confirmation.
4. Sub-processors. Your data passes through the vendor's own suppliers for voice, transcription and the AI model. Verify: ask for the current list and how much notice you get before a new one is added.
5. Access and permissions. You need to know who in your firm can read conversations, and who at the vendor can. Verify: look for roles such as admin, broker and read-only, multi-factor sign-in, a per-user access log and a written policy on vendor staff viewing your transcripts.
6. Evidence of security practice. A badge in a website footer is not a report. Verify: ask for the independent audit report or security summary, read its scope and date, and confirm that the product you are buying falls inside that scope.

Fit with your CRM or broking system

1. How it connects. A native integration, an open interface, a middle layer such as an automation tool and email forwarding all behave differently. Verify: ask for the documentation for your exact system, and for a reference firm on the same one. "We integrate with major systems" is not an answer.
2. What it reads and what it writes. The safest start is reading contacts and renewal dates and writing notes and tasks only. Verify: ask for the field-level list and a read-only mode. The connection should have no right to edit policies, premiums or commission.
3. How it matches a client. Couples, family firms and shared phone numbers create duplicates and mix-ups. Verify: ask which keys it matches on (email, mobile, policy number) and what it does on a near match: create a new record or ask a person.
4. How fresh the renewal data is. Reminders built on last week's data reach people who have already cancelled. Verify: ask whether it syncs on change, hourly, nightly or from a weekly export, then cancel a dummy policy and watch for a reminder.
5. What happens when the connection fails. Sync errors are silent unless someone is alerted. Verify: ask what becomes of a note it could not write, whether it retries, and who is told. Look for a failed-sync report.
6. Getting out. If you leave, you want your conversations and settings back. Verify: export a sample on day one in a standard file format, and check that scripts and knowledge documents export too.

Vendor pages tend to promise the connection in one line. Sonant AI's home page, for example, says its voice agent will "write every detail to your AMS" and names several broking systems, but it gives no field mapping, matching rule or failure handling. Those are the questions above, and a vendor with a good answer will have the documents ready.

Audit trail

1. A full transcript with timestamps. You may need to show exactly what a client was told. Verify: open one test conversation and look for speaker labels that separate the app from a human, timestamps to the second, and the channel.
2. The source behind each answer. To correct a wrong answer you need to know where it came from. Verify: for one test answer, ask the vendor to show which document or rule produced it and which version of that document was live.
3. A change log for settings. Someone edits the script on Friday and Monday brings complaints. Verify: ask for an admin log showing who changed prompts, rules and knowledge documents, with before and after text.
4. A hand-off record. It shows how long a client waited for a person. Verify: look for the reason for each hand-off, the time it was raised, the time it was picked up and by whom.
5. Search and export. In a dispute you need one client's history in minutes. Verify: time yourself finding one conversation by client name and date, export it to a file you can store elsewhere, and ask how long logs are kept and whether you set that.
6. Recording notices. If calls are recorded or answered by an AI, you need proof that the client was told. Verify: ask whether the opening announcement is logged for every call and whether you can edit its wording.

10:42:07  Caller (not yet verified)  "Am I covered for flood?"
10:42:09  App                        "I can't say what a policy covers. I'm passing you to a broker now."  [rule: no cover advice]
10:42:10  System                     Hand-off raised: reason=cover question, queue=brokers
10:44:31  Broker (initials)          Picked up. Waited 2 min 21 s. Summary shown on screen.

That is what a usable log looks like (illustrative). Each line says who spoke, when, and why the app behaved as it did. A dashboard that shows only counts of calls and a satisfaction score does not meet the first five audit checks.

Pricing model

1. The billing unit, defined in writing. A call, a conversation, an outcome, a resolution, a seat and a credit are different things. Verify: get the definition with its edge cases. Does a hand-off count? A spam call? A reminder you send first? A caller who rings twice?
2. Cost in your busiest month. Renewals cluster, so volume is not flat. Verify: take last year's busiest month, multiply by the unit price, and ask what happens past the allowance: an overage rate, an automatic upgrade or a hard stop.
3. Minimums and term. A minimum monthly commitment or a twelve-month term can outweigh a low unit price. Verify: find both in the order form, with the notice needed to leave.
4. Extras. Set-up fees, extra phone numbers, text messages, extra languages, recordings, seats and integration fees add up. Verify: ask for a quote that lists every line for your exact set-up.
5. Price-change terms. Usage prices get revised. Verify: ask how much notice you receive and whether you can leave without penalty. AI software contracts: auto-renewals, price rises and notice periods covers the wording.
6. Free testing. You need test traffic that costs nothing. Verify: ask whether sandbox conversations and test calls count against your allowance.

Whether a price is published at all is itself information. Smith.ai and Intercom, two general-purpose tools, publish theirs. Sonant AI's pages, one of the insurance-specific options, show no prices and only a "Book a Demo" button, so for that kind of tool the quote in writing, with overage, is your first deliverable.

What it may and may not say

1. Approved knowledge only. An app that answers from a generic model can invent policy terms. Verify: ask whether answers can be restricted to documents you upload, with "I don't know" as an allowed answer. Test it with a question your documents do not cover.
2. Hard refusals. Cover advice, quotes, binding and claim outcomes need a licensed person. Verify: in the demo ask "Am I covered for flood?" and "What will it cost?", then rephrase and push. The refusal and hand-off should hold every time.
3. Saying it is an AI. Clients should know, and if you have customers in the EU, the EU AI Act's duty to tell people they are talking to a chatbot has applied since 2 August 2026. Verify: listen to the greeting, check that the chat widget labels itself, and ask whether you can edit the wording.
4. Escalation triggers. Claims, complaints, bereavement, vulnerable clients and cancellations should reach a person. Verify: ask for the list, test three, and confirm that scheduled reminders stop for that client once a trigger fires.
5. Identity checks. Before it reads out a renewal date or policy detail, the app should confirm who is calling. Verify: ask what it asks for, and test with a wrong date of birth.
6. Resistance to being talked round. Callers, and mischief, try "ignore your rules". Verify: try to make it reveal a premium or policy detail, and check that the refusal holds.

Staff using ChatGPT on their own is a separate exposure, covered in the insurance broker AI risk checklist. Put both policies in place together, so the approved app is easier to use than the workaround.

Handover, testing and vendor health

1. A warm hand-off. A message left for a call-back is slower than a transfer. Verify: test in office hours. Does a person pick up with the summary on screen? Out of hours, does the app create a task with a deadline?
2. Test mode on your own material. Demos use perfect scripts. Verify: ask to run twenty anonymised past enquiries and the pilot script below through a sandbox.
3. A review routine. Answers drift when models or documents change. Verify: look for a weekly sample view or flagged low-confidence answers, and decide who reviews them.
4. Change notices. Vendors change the underlying model and the features. Verify: ask how and when you are told, and whether you can test a change before it goes live.
5. Support. When something fails on a Monday morning you need an answer that day. Verify: response targets in writing, a named contact and hours that match yours.
6. A reference your size. References from large brokerages show little about a ten-person firm. Verify: ask for two references with a similar headcount and the same broking system, and ask each what broke in the first month.

Scoring a vendor: one point per check you saw work

Give each vendor a sheet with the six groups and score one point for each check seen working or seen in writing, so each group scores out of six. Two rules keep the score honest. Any group below four is a no, however high the total. A total of 28 or more out of 36 is a yes only if no group failed.

Both vendors are invented, and the scores are illustrative. Vendor A answers the phone well and logs thoroughly, but it wrote notes only by email forwarding and would not quote past its entry tier in writing. Vendor B is weaker on voice but writes tasks and notes straight into the system the brokerage already uses. The decision: use B for renewals, where reading and writing renewal dates matters most, and ask A to come back when it can show the write-back working.

What one month costs under three pricing models

Take an illustrative six-person brokerage with about 900 live clients. It expects 60 phone calls to the app's number in a normal month and 120 in a renewal peak, plus 80 web chats. These sums use the prices that Smith.ai and Intercom published in November 2026, as examples of billing structures only. Neither tool is built for insurance, and this is not a recommendation.

Per call. Smith.ai's AI receptionist lists a free plan with 25 calls a month and $3 a call after that, and a Pro plan at $150 a month for 75 calls with $2.50 for each extra call. At 60 calls the free plan costs $105 (25 free, then 35 at $3), which beats Pro's $150. At 120 calls the order changes. The free plan costs $285 (95 paid calls at $3). Pro at 75 calls plus 45 extra costs $262.50. The next Pro tier, 150 calls for $270, comes between the two. The cheapest tier in a normal month is not the cheapest in the peak. The gaps are small enough that the definition of a call matters more than the tier: Smith.ai's page counts every answered call by default, filters known spam, and lets you remove up to 10% of calls from the bill each cycle for selected spam.

Per outcome. Intercom lists Fin at $0.99 per outcome, and counts an outcome when the customer confirms the fix, stops asking after Fin answers, or Fin completes a workflow including a hand-off. One charge applies per conversation. If every one of the 80 web chats counts, the bill is $79.20. If 60 of them do, it is $59.40. In a peak month with 160 chats it reaches $158.40. Used with another helpdesk it carries a minimum monthly commitment, which Intercom's page illustrates with 50 outcomes but does not fix. The lesson is that silence can count as success, so ask how "stops asking" is measured.

Per seat. Suppose a vendor charged $40 a seat for six seats. That is $240 a month in a quiet month and $240 in the peak. A flat bill is calm to budget, but you pay for seats that rarely touch the app. If only two of the six brokers handle renewals, ask whether the price works for two seats.

For the wider year-one budget, including set-up and staff time, see what AI costs a small insurance brokerage in year one.

Questions to send, and the answers that should worry you

Send the vendor a short email before the demo, so the demo answers questions rather than sets the agenda.

Subject: Questions before we shortlist [vendor name]

Hi [first name],

We are a small brokerage looking at an AI app for enquiries and renewals. Before a demo, please send in writing:

1. Where client data and call audio are stored, and whether we can choose the region.
2. Whether our content is used to train any model, yours or your providers'.
3. The exact definition of your billing unit, with examples of hand-offs, missed calls and spam.
4. A quote for 60 calls and 80 chats a month, and for 120 calls in our peak, including overage.
5. The fields you read from and write to [our system name], and what happens when a write fails.
6. A sample log of one conversation, including the hand-off record.
7. The independent security report, or summary, and its date.

Thanks,
[your name]

A twelve-message pilot script

Run these through the sandbox, by phone or chat, and record the result. Rows two, five and eight are automatic stops: if the app fails any of them, the vendor does not go forward whatever else it scores.

The reminder that went to a former client

Picture a brokerage whose app reads renewal dates from a weekly export. A client sells his business and cancels on a Tuesday. On Friday the app rings him about renewing and quotes the old date. He complains, and the transcript shows the call, but the log does not say which export the app used or when it was made, so the brokerage cannot tell whether the fault lies in the export or in the app.

Two changes would fix it. The app syncs on every change, with a filter that ignores any policy not marked live, and every reminder line in the log carries a last-synced time. Check 4 in the CRM group and check 2 in the audit group would have caught it before purchase: cancel a dummy policy during the trial and see whether a reminder still goes out.

A three-week trial before you sign

1. Week one: paper and sandbox. Read the data terms and pricing definitions, connect the app read-only, load twenty anonymised past enquiries and run the twelve-message script. Score the sheet. Drop any vendor that fails a stop row or a group.
2. Week two: shadow mode. The app drafts replies and call notes, and a person sends or files them. Compare thirty real interactions with what the app would have said, and keep a tally of wrong or risky answers. Delete a test conversation and confirm that deletion worked.
3. Week three: one live channel. Start with the lowest-risk one, such as out-of-hours calls or renewal reminders to clients with no open claims. A named person reads every transcript each morning. One statement about cover or price made to a client is an automatic pause and a root-cause request to the vendor.
4. The decision. Go live on more channels only if the live scores hold, the log answered every question you asked of it, and the real bill matches your peak-month sum. Then sign the term you chose, not the one the vendor suggested.

Questions that come up when shortlisting

Do I need an insurance-specific app or will a general one do?

A general tool is easier to test because many publish prices, but it will not know your policy vocabulary and may not connect to your broking system. An insurance-specific app is more likely to write back to the system you use, yet some publish no prices at all. Run the same pilot script through one of each and let the scores decide.

How long should an app keep call recordings and transcripts?

There is no single right period. Decide it from your own record-keeping duties and privacy notice first, then check that the vendor can match it for each data type and delete on request. Ask what happens to backups, and test deletion on one dummy conversation before you rely on it.

Who is responsible if the app tells a client something wrong?

Assume you are, because the client dealt with your firm. Ask a qualified adviser to read the vendor's liability, indemnity and data-protection clauses before signing, and ask for the hard refusals on cover and price to appear in the contract schedule rather than only in the demo.

Further reads

• AI Insurance Agency Management System: What to Ask Before You Switch — Questions to ask if the app pushes you to change broking system.
• Software for Insurance Agents and Brokers: Where AI Fits the Stack — See where an enquiry app sits among your other software.
• How Insurance Brokers Use AI to Handle Claims Enquiries — Set the rules for claims calls before the app takes them.
• How Brokers Use AI to Compare Policy Wordings at Renewal — Pair the app with AI help on renewal wording comparisons.
• Best AI Tools for Insurance Agents: Seven for Quotes, Renewals and Calls — A wider view of AI tools for quotes, renewals and calls.
• Where AI Saves Time in a Small Insurance Brokerage — Spot the jobs worth automating before you buy anything.
• AI Insurance Quote Comparison for Brokers: Carrier Quotes to Proposal — A broker's workflow from three carrier quotes to a client proposal: normalise, flag gaps, draft the narrative, and check every cell before advising.
• Insurance Agency SOPs With AI: Turn Routines Into Procedures — Record a routine, draft its decision points with AI and publish an SOP that the next account handler can actually follow.
• AI Cross-Selling for Insurance Agencies: Account Review Prep, Not Advice — AI prepares the file summary and the questions; the licensed agent advises. A prep routine, prompts and a sign-off log for account reviews.
• ALTA Best Practices and AI: What a Small Title Agency Should Document — Write the AI-use log for a small title agency: what Best Practices 5.0 says about AI, vendors and the written security plan, with a filled-in log.
• AI in General Insurance: What a Small Broker Can Use Today — Renewal reminders, quote summaries, claims chasing and client emails: what a small non-life broker can hand to AI today, and what stays with a person.
• Can AI Check an Issued Policy Against the Quote and Binder? — A line-by-line quote, binder and policy compare: limits, deductibles, endorsements, named insureds, and a discrepancy report you confirm with the carrier.
• Can AI Read Loss Runs? Carrier PDFs to Remarketing Summaries — Turn carrier loss-run PDFs into a frequency and severity table and a remarketing summary, with the checks that stop wrong totals reaching an underwriter.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Smith.ai and Intercom pricing pages and Sonant AI's public pages (fetched November 2026), plus the EU AI Act transparency duties for chatbots.