How Insurance Brokers Use AI to Handle Claims Enquiries

Insurance brokers can use AI to capture and sort claims enquiries, not to decide them. It can turn an emailed loss into a structured first notification, flag urgency and signs of a vulnerable client, pull the policy details a handler needs, and draft plain status updates. Whether something is covered stays a broker's statement, checked against the policy wording.

The biggest risk sits in a single sentence. An automated reply that says "your policy covers this" or "your claim has been accepted" creates an expectation the insurer may not meet, and turns a declined claim into a complaint against you. The biggest win sits in speed: a complete notification sent to the insurer on day one, with photos and the right policy number, moves a claim faster than anything you can do later.

Follow me on Instagram@sagnikteaches

Four kinds of claims enquiry, and who answers each

The fourth row is the one to get right on day one. If a message contains words like "complaint", "unacceptable" or "ombudsman", or simply reads as angry, the AI's only job is to flag it so a person picks it up under your complaints procedure.

Connect on LinkedInSagnik Bhattacharya

Keyword lists miss the politest complaints, which are often the most serious. Take an illustrative message from a dental practice client: "Third time I've asked for an update on the surgery flood claim. I'll be reviewing where we place our insurance at renewal." It contains none of the trigger words, yet it expresses dissatisfaction and it's a retention risk. So the prompt shouldn't ask "does this contain complaint words?" but "does the sender express dissatisfaction with our service or the insurer's, however politely? Answer YES or NO and quote the sentence." Asked that way, the model has a reason to flag the dental practice's message, which a keyword search would wave through. Asking it to quote the sentence also lets the reviewer see in a second why it said yes.

Subscribe on YouTube@codingliquids

Turning a first notification into a complete one

Clients report losses in whatever state they're in. The handler's first half hour usually goes on working out what's missing. Give the AI your notification fields and let it do that sorting. The fields a small broker typically needs:

• Policyholder and policy number (matched against your system, not guessed)
• Date, time and address of the loss
• What happened, in the client's words
• Damage or injury, and whether anyone was hurt
• Third parties involved, and their details
• Police or incident reference, where relevant
• Emergency steps taken (tarpaulin, water turned off, premises secured)
• Photos, receipts, estimates received
• Whether the premises are usable or the business can trade
• Best contact method and times

An illustrative email from a charity shop client:

Hi, water came through the ceiling at the shop overnight, the flat
above had a leak. Loads of donated stock soaked, the till area too.
We've shut for today. The upstairs tenant says it was his washing
machine. What do we do?? Photos attached. - Shop manager

The prompt asks the AI to fill your notification template from the email and the policy record you paste alongside, list missing items and draft a reply asking for them. A sample result (illustrative):

Policyholder: [charity name] - shop premises policy CS-2217
Date of loss: overnight, date not stated (assume [today])
Cause: escape of water from flat above (washing machine, per tenant)
Damage: donated stock, till area, ceiling
Business impact: closed today
Third party: upstairs tenant, name and contact not given
Photos: 6 attached
Missing: exact date and time discovered; third party's name and
contact; whether the water is now stopped; estimated stock value;
whether the till and card machine still work.
Urgency: HIGH - premises closed, possible ongoing leak.
Suggested reply: "Your claim is covered under escape of water..."

Most of that is exactly what a handler wants. Two lines need correcting. "Assume [today]" should never be written into a notification; a missing date stays missing until the client confirms it. And the suggested reply opens with a coverage statement the AI isn't entitled to make. Replace it with your standard acknowledgement: what happens next, what you need, and what they can do now to stop further damage.

Motor claims bring a different error: the AI can swap the parties. An illustrative client email reads "The van behind didn't stop at the lights and went into the back of us; the other driver gave me his details, photo attached." The photo of the other driver's licence carries the only name in the message, so a model can easily put that name in the policyholder field and the client's van registration under "third party vehicle". The check that catches it takes seconds: the policyholder name and registration in the notification must match your policy record exactly, and the prompt should say "take policyholder details ONLY from the pasted policy record; anyone named in the email or attachments is a third party unless they appear there".

The acknowledgement, and the phrases it must never contain

Write the acknowledgement once, have your compliance lead approve it, and let the AI fill only the blanks. A filled-in version for the charity shop:

"Thank you for letting us know about the water damage at the shop. We're sorry you're dealing with this. Here's what happens next: we'll report the claim to your insurer today and send you their reference. To do that we need four things: the date and time you found the water, the upstairs tenant's name and phone number, confirmation the leak has stopped, and a rough value of the stock affected. In the meantime, please keep damaged stock if you can and take photos before throwing anything away. If the premises aren't safe, call us on [number]."

Then give the AI a list of phrases it may never produce, and have the reviewer search every draft for them: "you are covered", "your claim is accepted" or "approved", "the insurer will pay", "within [n] days", "don't worry", and any mention of excess or settlement amounts. A two-second search catches the drafts where the model slipped back into being helpful.

Triage: what the AI flags and a person decides

Write the triage rules down, and treat the AI's flags as prompts for a person rather than decisions. Two kinds of flag matter most.

Urgency: anyone injured; property open to the weather or uninhabitable; water, fire or gas still an active risk; a business unable to trade; a theft where the premises can't be secured; a third party threatening legal action.

Vulnerability cues: a recent bereavement or serious illness; mentions of money worries, such as "we can't afford to replace this"; confusion or repeated questions; language difficulties; a message sent in obvious distress at 3am. These clients need a phone call from a person, and your process should record that the need was noticed and met.

An illustrative morning's triage after a storm:

The church office row is why a person reviews every triage. The AI had no way of knowing that a "minor leak" in that building sat above the most expensive item on the schedule.

Status updates clients can actually read

A lot of "where's my claim?" traffic exists because insurer correspondence is written for claims professionals. AI is good at translating it. Here's a typical insurer update and the version the client received:

From the insurer: "We acknowledge receipt of the loss adjuster's preliminary report. Reserve has been established. We await the policyholder's completed claim form and stock schedule prior to consideration of an interim payment. Liability remains under review pending the third party's insurer's response."

AI-drafted update, after the handler's check: "The insurer has the loss adjuster's first report. Before they can consider paying part of the claim early, they need two things from you: the completed claim form (attached) and a list of the damaged stock with rough values. They're also waiting to hear from the upstairs tenant's insurer about who is responsible for the leak. We'll chase that on Friday and update you then."

Notice what the handler removed from the AI's first draft: "you should receive an interim payment within two weeks". Nobody had said that. Never let an update promise dates or amounts the insurer hasn't given, and send updates on a fixed rhythm (weekly on open claims works for most brokers) even when there's no news, because silence is what generates the chasers.

The weekly rhythm is affordable once the AI does the reading. A quick illustrative sum: with 40 open claims, a handler who spends 12 minutes per claim reading the latest correspondence and writing an update needs 8 hours a week. If the AI summary and draft cut that to 4 minutes (read the insurer's note, check the draft, send), the same round takes under 3 hours, and the 5 hours saved go on the stalled claims that need a phone call.

Stalled claims are the other half of this. When the insurer hasn't replied for a set period, say 10 working days, the AI can draft the chaser to the insurer as well as the update to the client:

Draft a chaser to the insurer's claims team. Claim ref [ref].
Our last message: [date], sent the stock schedule and claim form.
No reply since. We need: their decision on an interim payment and
the latest on the third party's liability response. Polite, firm,
under 90 words. Do not mention any amount.

Illustrative output: "Claim [ref]: we sent the completed claim form and stock schedule on [date] and haven't yet had a response. Could you confirm whether an interim payment is now being considered, and whether the third party's insurer has replied on liability? Our client's shop has been closed since the loss, so an update this week would help us keep them informed. Many thanks."

That draft is sound, but check the business-impact line against the file before sending: if the shop reopened last Tuesday, "has been closed since the loss" overstates the position to the insurer, and a claims handler who later spots that will trust your next message less.

"Am I covered?": let AI find the clause, not answer the question

A coverage question is advice. AI can speed up the research by locating the relevant sections of a long wording, but the answer comes from you, or from the insurer if it's genuinely unclear. The prompt:

Below is the policy wording and schedule for this client, and their
question. Do NOT say whether they are covered.
1. Quote, word for word with section numbers, every clause that could
   be relevant: the insuring clause, definitions, conditions,
   exclusions and any endorsements.
2. List facts we would need from the client to apply these clauses.
3. Note anything in the schedule (limits, excesses) that applies.

QUESTION: Are our volunteers covered when they use their own cars
to collect donations?
[paste wording and schedule]

The AI returns the motor exclusion in the liability section, the definition of "employee" (which includes volunteers in this wording) and a note that the schedule shows no business-use motor extension. What to check: whether it has quoted the clauses exactly (compare against the PDF), and whether it has missed an endorsement at the back. Then you write the answer, which here is probably that volunteers need business-use cover on their own motor policies, and you say so in writing. Comparing wordings across insurers is its own task, covered in how brokers use AI to compare policy wordings.

One edge case catches brokers who keep wordings in a shared folder: the AI quotes last year's wording. Picture the charity renewed with the same insurer, which reissued the policy with a new volunteer driver endorsement, but both PDFs sit in the client folder and a handler pastes the older one. The quoted clauses look perfect and the answer built on them is wrong. It usually shows up as a section number or endorsement code that doesn't appear in the current schedule. Two habits prevent it: paste the schedule with the wording, so the AI can see the policy period, and add a line to the prompt asking it to state the wording's reference and edition date, taken from the document itself, at the top of its answer. If that date doesn't match the current schedule, stop.

Storm week, worked through

The brokerage in this illustration has six people and about 1,400 commercial and household policies. A normal week brings around 20 claims enquiries. A storm brings 90 in three days.

Before AI, capturing each first notification, chasing the missing details and submitting took about 25 minutes. With the AI structuring the email and drafting the missing-details request, the handler's time drops to about 10 minutes: read, correct, call if flagged, submit. Across 90 enquiries that's roughly 22 hours saved in the three worst days of the year, which is the difference between acknowledging everyone on day one and still working through the inbox on day three.

The triage did its job too: 14 enquiries flagged urgent, 6 with vulnerability cues, and handlers rang all 20 on the first morning. Two of the "routine" ones were upgraded by people who knew the clients. The acknowledgements used your standard wording, and no message said anything about cover.

If you already use Microsoft 365, Copilot in Outlook can do some of the drafting inside the mailbox; otherwise a business-plan chat assistant with pasted emails works, and moving to an automated inbox connection is a later step. The data here includes injuries and personal circumstances, so use only plans that don't train on your content by default.

Checks before claims enquiries go through AI

• Test on your own history. Take 30 past claims enquiries, run them through the prompts, and compare the output with what your handlers actually did. Carry on in shadow mode, with AI drafting and people sending, for at least a month.
• Check your agreements. If you handle claims under delegated authority from an insurer, read what the agreement says about systems and outsourcing before adding any tool.
• Log every complaint flag, including ones a person decides aren't complaints, with the reason.
• Keep inbox sorting separate. If claims arrive in a shared mailbox alongside renewals and new business, the sorting step comes first; AI email triage for shared inboxes covers the routing, and AI submission intake handles the new-business side.

Score the 30-enquiry test on a sheet like this illustrative one, filled in from a brokerage's past claims:

The rule for moving on is simple: no missed urgent case in two consecutive weeks of shadow mode, and every banned phrase still caught by the search. Until both hold, the AI drafts and a person does everything else.

Claims enquiries and AI: what brokers ask next

Can the AI send notifications straight to the insurer?

Not at first. Run it in shadow mode, where the AI prepares each notification and a handler sends it, until you've seen several weeks without a material error. Even then, keep a person approving before submission, because an incomplete or wrong notification can delay the claim or cause a dispute. Check each insurer's notification requirements and portal rules; many want specific fields or their own forms.

Should clients know an AI read their claim?

Say in your privacy notice and terms of business that you use software, including AI tools, to process enquiries, and name the kinds of provider. If a client talks to a chatbot, tell them so: that has been a legal duty for people in the EU since August 2026. Always give a simple route to a named handler, and never let automation stand between a distressed client and a person.

What about claims reported by phone?

Phone is where the most distressed clients call, so keep a person on it. The handler takes notes during the call, or uses a note-taker with the caller's consent, and AI turns the notes into the structured notification afterwards. That keeps the conversation human and still saves most of the rekeying time. If you want AI answering calls themselves, test it on quotes and admin first, not claims.

Further reads

• Where AI Saves Time in a Small Insurance Brokerage — The other broker tasks where AI saves the most time.
• Can AI Answer Calls and Qualify Leads for an Insurance Agency? — Whether AI should answer the phone at all, and for what.
• How to Automate Insurance Renewals With AI and Keep More Clients — The renewal side of the same client relationship.
• Can AI Handle Customer Complaints Without Making Them Worse? — Where complaint handling and AI drafting meet.
• How to Monitor AI That Talks to Customers: Hand-Offs and Errors — Watch hand-offs and errors once anything faces clients.
• What Does AI Cost a Small Insurance Brokerage in Year One? — Budget the tools and time for a brokerage's first AI year.
• How to Handle Sensitive Customer Conversations Without AI — Which customer conversations to keep away from AI, how to route them to a named person quickly, and what that person says when they pick up.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ChatGPT Business, Claude Team and Microsoft 365 Copilot plan pages (business-data defaults); EU AI Act Article 50 transparency duties.