An insurance adjuster can use AI to turn a claim file into a one-page summary and to draft status letters from fixed templates, as long as every template contains a locked line stating that no coverage or payment decision has been made and the adjuster reads and sends each letter. AI handles the facts and the wording; the decisions stay with you.
A letter that reads badly is a small problem. The bigger risk is a friendly, confident sentence that implies a decision: "we expect this to be paid shortly" reads as a promise. So the method below starts with the line you will never let the tool write, then builds the summary and letters around it. Carrier and client guidelines come first; if they restrict third-party tools or claimant data, those rules win.
The line AI is never allowed to write
Before any prompt, define the boundary. Write it as a one-page rule and attach it to every template.
- AI may: summarise facts already in the file, list documents received and outstanding, draft acknowledgements, requests for information and neutral status updates, and tidy your own inspection notes.
- AI may not: state or imply that a loss is or is not covered, that a payment will or will not be made, any amount, any reason for delay not recorded in the file, or any reading of the policy wording.
- A human must: read the whole letter, check each date and amount against the file, and send it under their own name.
The practical device is a locked paragraph in every template that the AI is told to copy verbatim, for example: "This letter is a status update only. No decision has been made on coverage or payment, and nothing in this letter should be read as a decision or as a promise. Your claim is being handled under the terms of your policy." Wording must be settled with your client or compliance contact; the sentence above is illustrative.
A worked example: a water-damage claim on a small café
Say an independent adjuster has a loss file for a café that reports a burst pipe behind the kitchen wall. Illustrative file contents: the first notice of loss form (two pages), the policy declarations page, 11 photos, a plumber's invoice for $640, a contractor's repair estimate for $18,900, three emails with the owner and a call note. The adjuster wants a one-page summary for the carrier's file and a status letter to the owner after the site visit.
Without AI, assembling the summary takes about 50 minutes: reading back through the emails, pulling dates and amounts, listing what is outstanding. With AI the routine is: paste or upload the non-sensitive structured parts (see the data section), ask for the summary in a fixed layout, then verify. Illustrative time: 10 minutes to generate, 15 to check against the documents, so about 25 minutes. The saving is modest on one file and large across a caseload of 40, which is why it is worth having a template.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowThe claim file summary template
A summary is only useful if it has the same headings every time, so the next person can find things. Use this prompt with the file material under it.
You are helping an insurance adjuster prepare an internal file summary.
Use ONLY the documents below. Do not infer, estimate or interpret.
Do not say whether anything is covered. If something is not in the
documents, write "not in file".
Layout:
1. Claim reference, date of loss, date reported, insured, location
type (e.g. commercial kitchen)
2. What the insured reports happened (their words, short)
3. Documents received (list with dates)
4. Documents still outstanding (from the checklist I give you)
5. Figures stated in documents: label each with its source
(invoice, estimate, insured's email). Do not add them together.
6. Timeline: date, event, source document
7. Open questions for the adjuster (facts missing or inconsistent)
Documents:
[paste]An illustrative extract of the output: "Figures stated: plumber's emergency call-out $640 (invoice dated 3rd); contractor repair estimate $18,900 (estimate dated 9th, unsigned); insured email of the 5th mentions 'about $25,000 in stock and equipment', no itemisation. Open questions: estimate does not say whether the kitchen wall was opened by the plumber or the contractor; the first notice says the leak was found on the 2nd, the insured's email says the 1st."
That last line is the real value. A summary that surfaces inconsistent dates is doing a job a human would otherwise do on a second read. What you fix: the assistant occasionally merges two figures or attributes an amount to the wrong document, so tick each figure back to its source. The instruction "do not add them together" is deliberate; adding figures from different sources is how a summary starts to look like a valuation.
Status letter templates with the decision line locked in
Four letter types cover most routine correspondence: acknowledgement, request for information, status update and a "still awaiting" reminder. Each one has the same three parts: a facts block the AI fills, a locked paragraph it must not change, and a next-steps block with dates that you confirm.
Draft a status letter from the template below. Fill ONLY the fields in
[square brackets] from the facts I give you. Copy the section marked
LOCKED exactly, word for word. Do not add any sentence about coverage,
payment, likelihood or timing beyond the dates I give.
Template:
Dear [insured name],
Re: claim [reference], loss on [date of loss]
Thank you for [documents received this week]. I visited on [visit date]
and photographed [areas]. I am still waiting for [outstanding items].
LOCKED: This letter is a status update only. No decision has been made
on coverage or payment, and nothing in this letter should be read as a
decision or as a promise.
What happens next: [next steps and dates from the adjuster].
If anything in this letter does not match your records please tell me.
[adjuster name, licence number, contact]
Facts: [paste]Illustrative output for the café: "Thank you for the plumber's invoice and the contractor's estimate. I visited on the 8th and photographed the kitchen, the rear wall and the storeroom. I am still waiting for a stock list with purchase dates and the equipment make and model numbers." Then the locked paragraph exactly as written. If the AI adds "we aim to settle this quickly", the template instruction has failed; delete it and tighten the prompt.
Keep a library: four templates, one per letter type, each with its locked paragraph. When your client updates its approved language, change it in one place.
Where letters quietly make decisions
The dangerous sentences are the helpful ones. Learn to spot them in AI drafts and in your own.
| Draft says | Why it is a problem | Replace with |
|---|---|---|
| "Your claim should be paid within two weeks." | Promises payment and timing | "I will update you by [date]." |
| "The water damage is covered under your policy." | States a coverage decision | Delete; coverage is addressed by the carrier |
| "Unfortunately gradual leaks are excluded." | Interprets the policy | "The cause of loss is under review." |
| "The $18,900 estimate seems reasonable." | Evaluates the amount | "I have received the estimate dated [date]." |
| "Delays are due to the carrier's backlog." | Gives an unrecorded reason | "I am waiting for [specific item]." |
Run a simple scan before sending: search each draft for "covered", "excluded", "will be paid", "approved", "denied", "should", "expect", "likely" and any currency amount. Every hit needs a source in the file or comes out. It takes 30 seconds, and is the cheapest control you have.
Photos and estimates: what AI can and cannot do with them
Photo handling is where adjusters most want help and where the tools are weakest. An assistant that can read images can describe what it sees: "stained plasterboard below a window, approximately 1.2 metres of visible discolouration, a metal shelf unit against the wall". It cannot tell you the cause, the age of the damage, or whether the staining pre-dates the loss. Use the descriptions as a first draft of your inspection note, then replace guesses with your own observations. Never paste a description into a report as if it were your finding.
Estimates are better suited, because they are text and numbers. Ask for a table of line items with quantity, unit and amount, and a check that the lines add up to the stated total. Typical catches: a line that duplicates another under a different heading, a quantity that does not match your measurements (a 14 square metre floor priced as 40), or a total with a transposed digit. These are queries to put to the contractor, not conclusions. For more on what the tools misread see what breaks when ChatGPT reads PDFs, spreadsheets and photos, and for a photo-led report workflow see roof damage photo reports for insurance claims.
Data, carrier guidelines and who sees the file
Claim files contain personal details, financial information, sometimes health information and the carrier's own confidential material. Settle these points before you use any tool.
- Read your service agreement and the carrier's guidelines. Some restrict third-party processing or require written consent. If it is silent, ask in writing and keep the reply.
- Use a business plan with training off by default, and read the vendor's data terms yourself. As of November 2026, OpenAI says ChatGPT Business content is not used to improve its models by default and Anthropic says the same of Claude Team; their individual plans may train on chats unless the setting is switched off. A free consumer chatbot is not suitable for claimant data.
- Minimise what you paste. The summary needs facts and amounts, not the claimant's phone number, date of birth or bank details. Replace identifiers with the claim reference.
- Do not use AI on files with legal complications: suspected fraud referrals, litigation, regulator complaints. Those follow your carrier's protocol.
- Keep the original file as the record. The AI summary is a working copy; save it with a note on which tool produced it and that you checked it.
Disclosure expectations are changing. A carrier may ask whether AI was used on a file, and the claims-handling rules that apply to your licence may say something about automated processes. Ask your client what it expects instead of assuming. The wider habits are in AI mistakes that damage customer trust.
Four quick examples from different files
A before and after on an acknowledgement. Before: "We have received your claim and will be in touch." After: "Thank you for reporting the loss on the 2nd. Your claim reference is [reference]. To start, I will need the plumber's invoice, 10 or more photographs of the damage and a list of affected stock. I will contact you by the 7th to arrange a visit. No decision has been made on coverage or payment." The second version gives the claimant three things to do and a date, and takes the AI one minute to draft.
A driving school's vehicle damage claim. The file has a repair garage's quote, a photo set and the instructor's account. The summary lists that the quote labels the damage "rear impact" while the instructor's note says the car was reversing into a bollard. The adjuster raises the discrepancy with the insured by email rather than guessing. The discrepancy was visible only because the summary placed the two statements side by side.
A mistake that showed up on review. A draft letter said "following your telephone conversation with the carrier on the 4th". The file had a call on the 14th. The date had been misread from a handwritten note. The adjuster's date check caught it before sending. Lesson: every date in a letter is checked against its source, no exceptions.
A quick sum on time. A caseload of 35 open files with a weekly status letter on each, at 12 minutes a letter by hand, is 7 hours. With templates and a check at 5 minutes each it is under 3 hours. That is illustrative; the saving is real only if every letter is still read, so measure honestly for a month.
From call notes to file notes without adding opinions
File notes are the other half of claim paperwork. After a call or site visit most adjusters have scribbled notes or a voice memo, and the file needs a clean, dated entry. This is a safe use of AI because the source is yours and the output is a restatement, but the same boundary applies: the entry records what was said and observed, not conclusions.
Turn my rough notes into a dated file note. Keep my wording where
possible. Use these headings: Contact, Date and duration, Who said what
(attribute statements to the person who made them), Observations I made,
Documents promised and by when, Follow-up actions. Do not add
conclusions, causes or opinions. Mark anything unclear as [unclear].
Notes: spoke owner 4th approx 10 min, says pipe went Sat night found Sun
am, cafe closed Sun anyway, says no prior leaks, plumber was [name]
came Sun 11am, stock in storeroom wet bottom shelves, will send
supplier invoices Weds, asked how long, told him i cant say yetA realistic result (illustrative) lists the owner's statements under "Who said what" as "owner states the pipe failed on Saturday night and was found on Sunday morning", the observations block as empty because this was a phone call, and the follow-up as "owner to send supplier invoices Wednesday". Notice the last line of the rough notes, "told him i cant say yet". A good assistant records it as: "Owner asked how long the claim would take; adjuster said no timing could be given." That is exactly the sort of line that protects you later.
What to fix: assistants sometimes replace "approx 10 min" with an exact duration, upgrade "says no prior leaks" into "no prior leaks" (turning a statement into a fact), or tidy away a hesitation that mattered. Search the draft for any statement that lost its "owner states" attribution.
A request-for-information letter, filled in
Requests for information are the highest-volume letters, and the most useful to template, because vague requests cause weeks of back and forth. A good one names each document, why it is needed in neutral terms and a date.
An illustrative result for the café file: "To continue with your claim I need the following by the 15th: (1) the supplier invoices for the stock you report as damaged, covering the last 3 months; (2) the make, model and purchase date of each item of kitchen equipment you report as affected; (3) the plumber's report on the cause of the failure, if one was provided; (4) photographs of the storeroom shelving taken before the clean-up, if you have them. If any of these are not available please tell me and I will note it on the file. No decision has been made on coverage or payment, and nothing in this letter should be read as a decision or as a promise."
Check that each item is a document, not a question about fault; that the reason, where stated, is neutral ("to continue the assessment" not "to verify your claim is genuine"); and that the deadline is realistic for the claimant, not just convenient for you. Where a document does not exist, a good adjuster records that instead of chasing it again, which also reduces complaint risk.
When the claimant is frustrated about delay
The hardest letter is the one to a claimant who has waited three weeks and is angry. An AI draft tends to over-apologise and then slip into reassurance ("we are doing everything we can to settle this quickly"), which is precisely the implied promise the template exists to prevent. Give the assistant the facts and the constraints, and ask for plain, short sentences.
Before (AI left to its own devices): "We sincerely apologise for the inconvenience and want to assure you that your claim is a priority. We expect to have this resolved very soon." After (rewritten to the rule): "I understand the wait has been difficult. Here is exactly where the file stands: I received your stock list on the 12th and I am waiting for the contractor's report, which I requested on the 9th. I will contact you by the 19th with an update whether or not it has arrived. No decision has been made on coverage or payment." The second version contains an apology in spirit, concrete facts and a date you control. Related tone guidance is in can AI handle customer complaints without making them worse.
Setting up the routine for a small firm
If three or four adjusters share templates, a little governance prevents drift.
- One prompt library. The summary prompt, the four letter prompts and the file-note prompt, held in a shared document with a version date and an owner.
- A second pair of eyes for new templates. A template does not enter the library until the firm's principal and, where relevant, the carrier's contact have approved the locked paragraph.
- A log, not a policy. A one-line record per file: AI used for summary, letters, or neither; who checked; any problem found. After 50 files you will know which steps pay off.
- A stop list. Files where AI is not used: suspected fraud, litigation, bodily injury, anything the carrier has excluded. Put the list at the top of the library.
- Quarterly review. Reread ten letters sent in the quarter against the table of risky phrases. Adjust the prompts, not the people.
If you would like to talk through how this fits your own caseload and carrier agreements, that is what my AI implementation consultation covers. The rest of this tutorial works without it.
Checking it worked, file by file
Run a monthly audit of ten sent letters. For each, confirm: the locked paragraph is present and unaltered; no sentence promises timing, outcome or amount; every date and figure ties to a document; the next step names who does what by when. Keep the audit sheet. If an issue appears twice, change the template rather than counselling yourself to be more careful.
Track also the claimant side. Watch whether complaints about unclear communication fall once letters list what is outstanding and by when. Brokers who sit between client and carrier will find how insurance brokers use AI to handle claims enquiries useful for what policyholders actually ask. For summarising long statements and transcripts safely, see summarising bundles and transcripts with AI.
Questions adjusters ask about AI on claim files
Can AI decide whether a claim is covered?
It should not. Coverage and payment decisions depend on the policy wording, the facts and the carrier's authority, and they carry legal consequences. Use AI to organise facts and draft routine correspondence, and keep every decision with a licensed human who owns the file.
Do I need the carrier's permission to use AI on a file?
Often yes, or at least you need to check. Carrier and client guidelines may restrict third-party tools, require disclosure or ban uploading claimant data. Read your service agreement and ask in writing before using any tool on a carrier's file.
Can AI read photos and estimates?
It can describe photos and extract line items from an estimate, but it misreads quantities and cannot confirm cause of loss. Use its output as a prompt for your own inspection notes, and check every figure against the source estimate.
What stops a status letter promising too much?
A fixed template with a locked paragraph that states no coverage decision has been made, and a rule that the AI may fill only the facts fields. The adjuster reads and sends every letter; nothing is sent automatically.
Further reads
- AI Answering Service for Insurance Agents: Claims Calls and Quote Leads — How agents handle claims and quote questions from callers.
- How to Handle Warranty Claims Faster With AI Triage — Triage ideas from warranty claims that carry over.
- AI in Delivery and Courier Services: Handling Damage Claims and Disputes — Evidence and wording on a different kind of damage claim.
- Does Your Business Insurance Cover AI Mistakes? — Whether insurance responds when AI makes a mistake.
- Which Legal Tasks Should a Small Firm Never Hand to AI? — Tasks a professional should keep for themselves.
- AI Insurance Agency Management System: What to Ask Before You Switch — Questions to ask about AI in agency systems.
- Loss Adjuster Software: Claim Files and Reports, and Where AI Fits — What a small adjusting firm needs from claims software, with verified prices and a clear line on where AI drafts and the adjuster signs.
- 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: OpenAI help-centre article on how data is used to improve models and Anthropic's privacy centre article on model training for commercial products (checked 4 November 2026); general claims-handling practice; each carrier's and client's own guidelines govern tool use. No vendor pricing is quoted.