Claims Automation - FNOL to Settlement

What this is

Most claims workflows either hand everything to a model or keep everything manual. This template does neither. AI reads the claimant's description and drafts summaries for adjusters, but it never decides coverage and never calculates a payout. Those calls come from deterministic rules and plain code, so every dollar amount and every coverage decision stays explainable and auditable.

Simple, low-risk claims settle automatically with no human involved. Anything flagged, high value, or uncertain routes to a human adjuster along with an AI-drafted briefing, and the adjuster has the final say. Every decision gets logged before the workflow acts on it.

AI drafts and summarizes. Rules decide. Code calculates. Humans approve the exceptions. That split is what makes an insurer comfortable letting a workflow like this touch real money.

How it runs

The workflow accepts a First Notice of Loss submission from a customer portal, a mobile app, or a transcribed phone call, and normalizes it into one shape. If a policy number, date of loss, or claim type is missing, the workflow stops immediately and sends the claim back to intake, so no AI or API calls run against an incomplete claim. Once the claim passes that check, the workflow looks up the policy's live status, product, and coverage terms from the insurer's policy system.

An AI step reads the claimant's description and turns it into structured data: an assessed loss amount, a loss category, a confidence level, and any evidence that looks missing. It's told to extract, not judge. A coverage rules table takes that data plus the policy terms and returns whether the claim is covered, and at what limit and deductible. A plain arithmetic step then calculates the payout, the smaller of the assessed loss or the coverage limit, minus the deductible, floored at zero.

A second AI step screens for fraud signals separately, flagging inconsistencies without ever denying a claim on its own. A routing rules table then takes the coverage outcome, the payout, the fraud risk, and the extraction confidence, and decides whether the claim settles straight through or goes to a human adjuster, and which tier. A high fraud signal always forces human review, no matter how clean the rest of the claim looks.

Before anything moves, the workflow writes a full audit record. Claims that clear settle automatically with a payment and a confirmation email. Claims that need review get an AI-drafted summary sent to the adjuster, clearly labeled as a draft, alongside the real calculated figures. The workflow checks in periodically until the adjuster approves or declines, then issues payment for whatever amount the adjuster confirms.

The steps

• Code (receive_claim): Accepts the FNOL submission from any source, normalizes it into one shape, and generates a claim ID if one wasn't supplied.
• Code (reject_incomplete_claims): Checks for a policy number, date of loss, and claim type. Missing any of them ends the run and sends the claim back to intake before any paid call fires.
• HTTP (policy_lookup): Calls the insurer's policy system for live status, product, and coverage terms. This is always a real lookup, never an assumption.
• Claude (extract_loss_details): Reads the claimant's free text description and extracts an assessed loss amount, loss category, confidence level, and missing evidence notes. Scoped to extraction only, not judgment.
• Decision Engine (coverage_rules): Business-owned rule table that takes policy status, product, peril, and assessed loss, and returns whether the claim is covered, along with the limit and deductible.
• JavaScript (calculate_payout): Plain arithmetic. Takes the smaller of assessed loss or coverage limit, subtracts the deductible, and floors at zero. No AI involved, so the number reproduces exactly every time.
• Claude (fraud_screen): A separate AI step that flags inconsistencies between the claim and the evidence and returns a fraud risk level with specific indicators. It flags; it never decides.
• Decision Engine (routing_rules): Business-owned rule table that takes the coverage outcome, payout, fraud risk, and extraction confidence, and decides straight-through settlement versus adjuster review, and which tier. High fraud risk always wins and forces review.
• HTTP (log_decision_trail): Writes the coverage decision, payout calculation, fraud signal, and routing decision to the audit log before anything is paid or sent, tied to the claim ID.
• Switch (route_outcome): Sends straight-through claims to automatic settlement and flagged claims to adjuster review.
• HTTP (issue_payment) and Email (send_settlement_confirmation): For straight-through claims, issues the payment and emails the claimant a breakdown of the loss, deductible, and payout.
• Claude (draft_adjuster_summary): Writes a concise case summary for the adjuster, instructed to summarize only and never recommend approval or denial.
• Email (notify_adjuster): Sends the adjuster the AI-drafted summary clearly labeled as a draft, plus the real calculated figures and fraud signal.
• Wait (await_adjuster_decision): Polls the claims system until the adjuster records an approve or decline decision, holding the claim open rather than timing out.
• HTTP (issue_adjuster_approved_payment): Issues payment for the amount the adjuster confirms, which can differ from the original calculation.
• Exit (exit_declined): Ends the run cleanly when the adjuster declines. The decision is already logged.

Design notes

AI never has authority to approve, deny, or price a claim. Every AI step in this workflow is scoped narrowly to extract, flag, or summarize, and is explicitly instructed not to make judgment calls that belong to rules or humans.

Coverage and payout math are fully deterministic. They come from rule tables the insurer's own team configures, and from plain arithmetic, not from a model's reasoning. The numbers stay reproducible and explainable to a regulator, an auditor, or a customer asking why they were paid a specific amount.

Fraud escalation is enforced, not optional. A high-risk signal always routes to a human, placed so it can't be overridden by an otherwise clean routing rule.

Every claim leaves a trail. The workflow logs the reasoning behind each decision before it acts on it, so nothing gets paid or routed without a record of why.

Humans stay in control of judgment calls. Anything uncertain, high value, or flagged goes to a real adjuster, who sees the AI's draft summary clearly marked as a draft next to the real numbers, and who has to actively approve or decline before money moves.

Setup

1. Connect a policy lookup API that can return a policy's status, product, and terms when queried. The workflow calls this once per claim, before any AI or rules step runs.
2. Connect a claims backend API that can accept audit log entries, accept a payment instruction, and report whether an adjuster has approved or declined a given claim.
3. Create the coverage rules decision table, owned by the underwriting team, defining what's covered and at what limit and deductible based on policy status, product, and peril.
4. Create the routing rules decision table, defining which claims settle automatically versus which go to a human, which adjuster tier handles what, and the rule that a high fraud signal always forces human review.
5. Set up adjuster tooling so adjusters can see the case in their claims console and record an approve or decline decision the workflow can poll for.
6. Configure an email sending connection, for example Gmail, with a sending address the insurer controls, for settlement confirmations and adjuster review requests.
7. Connect Claude. The extraction, fraud screening, and summary drafting steps run out of the box on the platform with no separate API key needed.

Known limitations of this version

• The AI reads the claimant's written description but doesn't directly open attached photos or PDFs yet. It reasons from text and file names, not image or document content.
• If the policy lookup fails or the policy isn't found, this version falls through to a default not covered rule rather than flagging the failure explicitly.
• The audit trail captures the machine's decisions but doesn't yet log the adjuster's final approve or decline decision as a separate event.

When to use it

• You process claims where most cases are simple and low risk, and you want those settled without a person touching them.
• You need coverage and payout decisions to be deterministic and auditable, not left to a model.
• You want fraud screening built in, with high-risk cases always reaching a human no matter how clean the rest of the claim looks.
• You want adjusters working from an AI-drafted summary and real numbers side by side, not making decisions blind.