AI for Public Adjusters: Claim Packages and Carrier Letters

Use AI for the paperwork around a claim, not the numbers inside it. It can index documents, draft carrier letters, summarise estimates and build timelines, but every figure must trace to a document you hold. A figure-check step before anything leaves the office is what keeps a package accurate and defensible.

A public adjuster is paid, in most places, as a percentage of the recovery, which creates a pull towards bigger numbers. A writing tool that "improves" a scope narrative can quietly turn a 30 square-foot water stain into "extensive damage across the ceiling", or round $18,740 up to a tidy $19,500. Neither is a typo; both are the kind of change that carriers, regulators and courts treat as a problem. The defence is one habit: a figure ledger that every package has to pass.

Follow me on Instagram@sagnikteaches

What AI is good for in a claim file, and what it must not touch

The bottom two rows matter. Do not ask a chatbot "is this covered?" and relay the answer to a client. Coverage turns on the policy language, the facts and the law of your jurisdiction, and a confident wrong answer in a letter to a carrier can damage the claim and your standing.

Connect on LinkedInSagnik Bhattacharya

Licensing and contract rules to check before you automate

Public adjusting is a licensed activity where it is permitted, and the rules sit with each jurisdiction's insurance regulator. They commonly cover points such as a written contract with specific wording, limits on fees or on charging for certain claims, a cancellation window for the policyholder, rules on soliciting, record-keeping periods and what you may say or promise. I cannot summarise yours here and it would be wrong to guess, so make a one-page list from the regulator's own pages and keep it with your templates. Then ask of every AI-assisted step: does it touch any of these?

Subscribe on YouTube@sagnikteaches

• Do your letter templates carry the licence number and wording your regulator requires?
• Does anything an AI drafts promise an outcome ("we will recover the full replacement cost")? Strip it.
• Do client intake forms tell the client how their documents are handled, including any AI tools?
• How long must claim files be retained, and is an AI tool's chat history part of that record or outside it?

Choose a tool setup that is safe for a claim file

A claim file includes names, addresses, policy numbers, financial details and photographs of a home or business. Treat it as confidential client information.

• Use a business plan with written no-training and confidentiality terms. ChatGPT Business lists at $25 per user per month ($20 on annual billing) and Claude Team has a similar structure; both need at least two seats, so a solo adjuster either pays for two or uses an individual plan with training switched off. Check the vendor pages for current prices.
• Paste the minimum. A carrier letter draft needs the claim dates, amounts and the point at issue, not the client's date of birth.
• Remove identifiers from text you paste where you can ("Insured", "Carrier", "Property") and put the real names back in the final letter.
• Tell clients in your intake paperwork that you may use secure software tools, including AI, to prepare documents, and that you review everything yourself.

The figure ledger: one table that every number must pass

This is the step that answers the risk of inflation. For each claim package, keep a table with one row per figure that appears in any letter, summary or schedule. The table is simple, which is the point.

The rules are short:

1. No source, no figure. If a number cannot be traced to a document in the file, it comes out of the letter.
2. Exact, not rounded. Use the source's number. Rounding is for conversation, not for claim documents.
3. Totals are recalculated outside the AI. Add them in a spreadsheet and put the spreadsheet in the file. Language models are unreliable at arithmetic, as the explainer on AI and maths shows.
4. Estimates are labelled. If a figure is your estimate rather than a bid or an invoice, say so plainly in the letter.
5. A second pair of eyes for big packages. Above a threshold you set (say $50,000), someone other than the drafter checks the ledger.

Ask the AI to help build the ledger, not to skip it: paste the draft letter and the list of source documents' key amounts, and ask it to list every number in the letter that does not appear in the list. That prompt gets you a short list of orphans to check.

Below are (A) a draft letter and (B) a list of figures taken from the source documents, each with its document name.
List every number, date and percentage in the letter. For each, say which line of (B) it matches exactly, or write "NO MATCH".
Do not correct anything and do not suggest new figures.

(A) [letter text]

(B) [figures list]

An illustrative output: 14 numbers found, 12 matched, one "NO MATCH" for "$19,500" (the source says $18,740), and one for "30 days" which is the carrier's policy time limit and not in your list. You fix the first (it is an upward rounding error, exactly the kind to catch) and add the second to the ledger with the policy page.

Drafting the carrier letter

A good carrier letter is specific, dated and polite, asks for something clear and attaches evidence. AI is useful for turning your notes into that shape. The prompt below keeps the model inside the facts you supply.

Draft a letter from a public adjuster to a property insurer's claims handler.
Use only the facts below. Do not add figures, dates, policy clauses or legal arguments that are not listed.
Facts: claim number [number]; loss date [date]; carrier estimate received [date] totalling $[carrier total];
our contractor's bid dated [date] totalling $[bid total]; the difference is in these items: [list with amounts and the document for each].
Request: a written explanation for each item the carrier excluded and a re-inspection within [time period stated in the policy or required by rule].
Tone: professional, concise, no threats.
List attachments at the end, naming each file.

What you would typically fix in the draft that comes back: an assertive phrase such as "the carrier has clearly breached its obligations" (remove it unless a qualified adviser has advised you to say so); a deadline the model invented (use the one in the policy or the regulator's rule, quoted from the source); and any adjective that carries an unsupported claim, like "catastrophic" or "total". Replace those with the measured fact: "water damage affected 420 square feet across three rooms, per the field measurements attached."

Before and after: a scope sentence

The first draft did two things at once: it widened the scope from two slopes to the whole roof, and it rounded $21,460 up to a tidy $25,000. The second version is longer and duller, and far harder for anyone to dispute. It is also the one the carrier's reviewer can verify without a phone call.

Summarising the carrier's estimate and denial letters

Reading a 40-page carrier estimate against a contractor's bid is slow work. AI can build a comparison table of line items, which saves an hour. The risk is mismatched lines and silent omissions, so use it as a guide to where to look.

1. Paste each estimate in turn and ask for a plain table: line, description, quantity, unit price, total, exactly as printed.
2. Ask for a second table pairing lines from the two estimates that appear to cover the same work, with unmatched lines listed separately.
3. Check the totals of each table against the totals printed on the documents. If the AI's table does not sum to the printed total, a line is missing or wrong; find it.
4. Read the pairings yourself. Models sometimes pair "remove and replace" with "repair", which is a real difference in scope.

The same method suits long denial letters: ask for a list of every reason given, with the paragraph number quoted, and then read the paragraphs. The tutorial on summarising bundles safely covers the checks in more depth.

A claim package from start to finish: a flooded retail unit

An illustration with invented figures. A small shop suffers a pipe burst overnight. The adjuster is engaged under a written contract.

1. Day 1 (30 minutes). After the first visit, the adjuster dictates notes. AI turns them into a tidy site-visit record with the sections "observations", "measurements", "actions taken by the client". The adjuster checks the three room measurements against the tape-measure sheet.
2. Days 2 to 5. Photos, the plumber's invoice ($640), the water-extraction invoice ($2,310) and two contractor bids ($14,200 and $15,800) arrive. AI proposes file names and an index. The adjuster fixes two mislabelled photos.
3. Day 6. The adjuster builds the figure ledger: 23 figures, each with a source. A spreadsheet adds the totals. The lower bid is used as the main figure, with the second bid attached.
4. Day 7. The letter is drafted with the carrier-letter prompt, cut to one and a half pages, and checked: one "NO MATCH" caught ($15,000 where the bid says $14,200). Sent.
5. Weeks 3 and 4. The carrier's estimate arrives at $9,850. AI builds the comparison table. The adjuster finds the carrier omitted the ceiling replacement and used a lower unit price for flooring. A second letter requests an explanation of each, quoting the policy sentence that applies, which the adjuster has read in full and copied exactly.

The AI saves perhaps three to five hours across the claim, concentrated in indexing, the first drafts and the estimate comparison. The adjuster's judgement, the visits and the negotiation are unchanged.

Building the timeline: a prompt and a realistic output

Carriers and reviewers read a clear chronology faster than a pile of emails. AI is good at ordering dated material into a table, as long as you give it the dates and it does not infer any. Paste the dated items (email headers, invoice dates, inspection notes) and use a prompt like this:

Build a chronology table from the dated items below. Columns: date, event, source document.
Use only dates that appear in the items. If an item has no date, list it separately under "undated".
Do not infer or estimate dates. Keep events to one short sentence each.

[items]

An illustrative result for a water-loss claim:

What you would fix: the model may sort "3 March" events in the wrong order if two share a date, or turn "early March" into a specific day. Confirm each date against its source and put the "undated" items right by asking the client or checking photo metadata. If a date is in dispute, say so in the table instead of choosing one.

Intake and engagement checklist that mentions AI

Your engagement paperwork is where client trust starts. Alongside whatever your regulator requires, add a short note on tools, written in plain language. A filled-in illustration:

HOW WE HANDLE YOUR DOCUMENTS
We keep your claim file in [secure storage system].
We may use secure software tools, including AI writing and summarising tools on a business account
with no training on our content, to prepare drafts, indexes and summaries.
We keep your full policy number, bank details and date of birth out of those tools wherever we can.
Every figure and statement in a letter that goes to the insurer is checked by me against your documents
and signed by me. AI does not decide what your claim is worth.
You can ask us not to use these tools on your file.

Keep a signed copy. If a client opts out, flag the file so nobody pastes its content into any tool.

Carrier requests: proof of loss, documents and deadlines

Carriers send formal requests with deadlines: a sworn statement, a contents inventory, receipts, a recorded statement. These are the moments to be precise. AI helps in three safe ways: turning the request into a checklist of what is asked, drafting a covering note, and tracking what has been sent. It must not answer the substance, and it must not promise to deliver something by a date you have not confirmed.

An illustrative request: "Provide an itemised inventory of damaged contents, with age, purchase price and receipts, within 30 days." Paste that, ask for a checklist, and you get something like: inventory spreadsheet with item, room, age, purchase price, replacement cost, receipt reference, photograph reference. You then check the deadline against the carrier's letter and the policy. If the inventory shows 62 items, the ledger rule applies to the unit prices: purchase prices come from receipts or the client's statement, replacement costs from a named source, and anything estimated is marked "estimate".

Contents inventories: where inflation creeps in

Contents claims are where well-meaning rounding does the most harm. A client remembers a television as "about $1,200" and an AI tidy-up changes it to $1,200.00, which looks like a receipt figure. Keep a column called "basis" with values such as "receipt", "bank statement", "client estimate" and "retailer price list, accessed date", and keep it in the schedule you send. It signals honesty and speeds the carrier's review.

Where this goes wrong in practice

• The helpful paraphrase. You ask for a summary of a policy clause and the model reworks it. The summary says "includes sudden water damage", but the clause also says "unless the loss results from long-term seepage". Quote the clause word for word in letters; never rely on a paraphrase.
• Mixed-up claims. You use one chat for several files and the model blends the details. Start a new conversation for each claim, and keep claim identifiers out of chat titles.
• Template drift. A letter template is edited by an AI over months until it carries a clause your regulator never approved. Keep a master copy and compare.
• Over-trust in tone. Polished writing can make an unsupported point feel stronger. Strip adjectives and keep the facts.
• Silence on AI. A client asks if you use AI and you say no, when you have been using it for letters. Be honest and use the paragraph above.

Fraud risk: why inflated figures are not a grey area

Insurance fraud laws generally cover knowingly presenting false or misleading information in support of a claim, and adjusters can be held responsible for what they sign. An AI tool does not change that. If a model produces a plausible figure and you submit it without a source, the intent question may become awkward even if you did not set out to mislead. The safe position is simple: you can show where every number came from.

Some warning signs in your own workflow:

• A letter figure you cannot find in the file.
• A scope description with more items than the photographs show.
• A total that moved between drafts without a document that moved with it.
• A "reasonable estimate" for items nobody has priced.

If a client asks you to include something you cannot support, say no in writing. That protects both of you, and the figure ledger makes the conversation easier because the rule applies to everything.

A weekly check that the routine is holding

1. Pull two closed or active files at random. Re-check three figures in each against the source.
2. Look for any letter that went out without a signed ledger.
3. Note any AI error caught this week and add it to a short list of "known failure modes" the team reads.
4. Confirm that nothing identifiable was pasted into a consumer account. If it was, tell the client and stop that habit.

If your own cover might be affected by AI-related errors, the tutorial on insurance and AI mistakes explains what to ask your insurer. And for the daily handling of damage photos, the roof-damage reporting tutorial shows how to document without adjusting. If you would like help setting the workflow up around your own files, that is what my AI implementation consultation covers.

Public adjusters and AI: practical questions

Can AI write the estimate of loss for me?

It can format and total figures you supply, but the scope, quantities and prices must come from your inspection, contractor bids, invoices or a pricing database you are licensed to use. A figure a model invents or rounds upward is not evidence, and presenting it as one can be misrepresentation.

Is it a problem to put policyholder documents into an AI tool?

It can be. A claim file holds personal, financial and sometimes medical information, plus the policy. Use a business plan with written no-training terms, tell the client in your contract or intake paperwork, and keep what you paste to the minimum the task needs.

How do I know a carrier letter is accurate before it goes out?

Compare every date, policy number, amount and quoted policy sentence with the source document, and tick them off on a checklist. Read the letter aloud once for tone. If a clause is quoted, open the policy and confirm the wording is word for word.

Does my licence or contract limit how I use AI?

Licensing and contract rules differ by jurisdiction and cover things such as written contracts, fee limits and what you may say to carriers. Read your licence conditions and ask your regulator or adviser whether anything addresses automated tools. Where nothing does, the usual duties of honesty and competence still apply.

Further reads

• AI Receptionist for Roofing Companies: Storm-Day Calls and Claim Questions — Handling the surge of storm-day calls and claim questions.
• How Brokers Use AI to Compare Policy Wordings at Renewal — Comparing policy wording with AI, with checks.
• How Insurance Brokers Use AI to Handle Claims Enquiries — How another claims-side profession handles enquiries.
• How to Handle Warranty Claims Faster With AI Triage — Triage a claim queue by type and urgency.
• Which Legal Tasks Should a Small Firm Never Hand to AI? — Tasks that should stay with a qualified person.
• AI Submission Intake: Stop Re-Keying Data Into Insurer Portals — Stop re-keying details into carrier portals.
• 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.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: general public-adjusting practice; vendor privacy and plan pages for ChatGPT Business and Claude Team (checked 4 November 2026); your own licence conditions and contract rules, which you must read yourself.