AI for Insurance Adjusters: Claim Notes, Photos and Carrier Updates

An AI agent can help an independent insurance adjuster by drafting claim notes from dictated observations, proposing labels for photos and preparing carrier updates for review. The adjuster keeps authority over coverage, valuation and every message that leaves the file. A compounding claims memory means the agent carries forward each claim's facts, sources and open questions, so every new note starts from the same record. The value depends less on the model than on the file structure, the carrier rules written down in advance and the sign-off steps that stay with a person.

Key takeaways

• A claims memory works per file: facts, sources and open questions stay attached to the claim they belong to.
• The agent drafts notes, photo labels and updates, and the adjuster approves anything sent to a carrier or policyholder.
• Carrier requirements belong in written templates with named owners, not only in what the model happens to remember.
• Permissions and the claim folder structure should be set before the first live file, not after a problem appears.
• Start with one role and one check you can observe, such as whether draft notes match the adjuster's confirmed observations.

What a compounding memory changes for an independent adjuster

An independent adjuster usually carries several open files at once, each with its own carrier, reporting rhythm and unresolved questions. The facts that matter move between site visits: a contractor revises a scope, a policyholder calls about access, a salvage vendor waits for a decision. Today those facts sit in a notebook, a voice memo, an email thread and a claims platform, and rebuilding the full picture before a carrier update takes real time.

A compounding memory means each file keeps its own growing record, and the agent retrieves only the relevant part when it drafts the next note or update. Anthropic's guidance on context engineering makes a useful point here: an agent works best when it receives a focused, curated set of information for the task in front of it, not everything it has ever seen. For claims, that means memory is scoped to one claim, one carrier and one policy, and nothing crosses between files. Compounding also should not mean the agent quietly changes how it writes for a carrier. Preferences should be written down, reviewed and versioned like any other procedure.

Claim notes that stay consistent across a file

Claim notes are evidence. A good note separates what the adjuster observed, what another person said, what a document contains and what remains open. An agent can turn a voice memo or rough bullets into a structured entry, but only if the structure is fixed before the first claim is opened. A workable note template includes these fields:

• Date and time of the observation, and who was present
• What the adjuster directly observed, described in plain terms
• What a third party said or provided, with the document name
• Open items, each with the person responsible for it
• Anything the agent could not confirm, marked as unverified

The rule that matters most is that the agent never upgrades a statement. If a contractor said a subfloor was dry, the note records that the contractor said so, and it does not write that the subfloor is dry.

Illustrative scenario: an adjuster inspects a water loss in a kitchen and dictates that the subfloor felt soft near the dishwasher. A week later, a contractor's estimate arrives describing the subfloor as sound. The agent places both entries on one timeline, notes that the estimate was written before the adjuster's second visit, and drafts a question for the adjuster rather than a conclusion about which account is correct.

Photo sorting without losing the evidence trail

Photos are where an agent can save the most sorting time, and where a careless step costs the most. The agent can propose labels from image content and file metadata, such as room, elevation, damage type and capture time. It should not rename, crop or overwrite originals. Keep the original files read-only, keep agent labels in a separate photo index and record an audit entry for every change. The same discipline applies in other trades, as a moving company's damage claim workflow or an equipment rental damage record would show: label first, confirm second, and never overwrite the original.

Photos also carry data the file does not need. Metadata can include location, and images can show people, mail or addresses unrelated to the loss. Decide in advance which fields an agent may read and which are removed before any external tool processes the file. Attached documents can add a different risk. A scanned invoice or a contractor's note may contain text written to look like instructions, such as a request to ignore earlier rules. OWASP's guidance on prompt injection advises treating such content as data to be examined, not as commands. In practice, the agent reads these files for facts only, and its tools cannot take actions because of text found inside them.

Carrier updates and the rules you have to write down

Each carrier has its own expectations for timing, format and content. Those expectations come from carrier guidelines, contract terms and the regulations that apply where you are licensed, and an agent cannot reliably infer them from a few past emails. Write each carrier's requirements as a template with an owner, a review date and two explicit lists: what every update must include, and what it must never include.

The safe default is that the agent drafts and the adjuster sends. Anything touching coverage position, liability, reserves or settlement amounts stays with the adjuster until they have checked it against the file. The agent can flag that an update is overdue and can draft it from the observation log, but changing a status in the carrier's system is a separate action that needs its own approval. The NIST AI Risk Management Framework offers a practical reference for this kind of oversight, because it treats AI risk as something to map, measure and manage with defined responsibilities for people.

File structure, permissions and sign-off to set up now

These steps should be completed before the first live claim runs through the system.

• Step 1: Define one folder structure per claim with fixed sections for policy references, documents, photo index, observation log, open items and outgoing updates.
• Step 2: Write a carrier template for each carrier you work with, listing required timing, required fields and prohibited statements.
• Step 3: Give the agent read access to the claim folder and write access only to drafts, the observation log and the photo index. It holds no send permission and cannot change claim status.
• Step 4: Connect accounts through standard delegated authorization (OAuth 2.0), so access can be revoked without changing passwords. OWASP's authorization guidance supports granting only what a task requires and checking access on each request.
• Step 5: Set sign-off gates. The adjuster approves every outbound message, and a second review applies to any figure or coverage language.
• Step 6: Log every agent action with a timestamp, the source files used and the person who approved the result.
• Step 7: Test on closed files whose outcomes are known, record every draft that contradicted the adjuster's own notes, and repeat the test after any template or model change.
• Step 8: Define a stop condition. When the agent cannot match a fact to a source, it leaves the item open rather than filling the gap.

How Be Human approaches this, and where to start

Be Human designs and operates this kind of setup as a managed service. A first engagement defines one useful role for the agent, the memory and tools it needs, the approval boundaries and one observable success criterion. Ongoing work can include monitoring, recurring evaluations of draft accuracy against adjuster-confirmed facts, permission reviews whenever carrier access changes, integration maintenance and incident response. The adjuster keeps authority over every claim decision throughout.

If you want to find which part of your file work is the right first task for an agent, take the free Be Human AI diagnostic below.

Frequently asked questions

Can an AI agent write claim notes for an adjuster? Yes, within clear limits. The agent can turn dictated observations and documents into structured notes, but each fact should keep its source, and anything unconfirmed should stay marked as open. The adjuster reviews and approves every note before it becomes part of the file.

Does the agent send updates to carriers or policyholders? Not by default. It drafts carrier updates from the observation log and the carrier template, and the adjuster approves and sends them. Sending and changing claim status are separate permissions, and you can grant them later, one at a time, only after testing has shown the draft quality holds.

How should photos and personal data be handled? Keep original files read-only and store agent labels in a separate index. Decide which metadata fields the agent may read, and remove location data or unrelated personal details where your obligations require it. Confirm the handling rules against the privacy law that applies to your policyholders and to the jurisdictions where you work.

What should an independent adjuster set up first? Start with a fixed folder structure for each claim, a written template for one carrier and read-only access to the claim folder. Add the observation log and drafting permissions next, and consider any outbound step only after that. Choose one measurable check, such as whether drafts match the adjuster's confirmed observations, before expanding the role.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.