Insurance claims agent
We design, build and deploy an AI agent that reads a claim end to end — the policy, the incident report and every receipt — reconciles them against each other, and hands your adjuster a decision-ready file with the leakage and fraud signals flagged, while the pay-out decision stays with a human.
At a glance
Every over-payment it catches and every clean claim it fast-tracks is money — claims leakage typically runs 5–10% of paid claims, and consistent scrutiny is what a BNM conduct review looks for.
An orchestrator decomposing the claim into parallel checks, vision/OCR on reports and receipts, policy-clause reasoning, fraud signals, and BNM/PIDM/PDPA-grade controls and audit trail.
What changes when it works
Time per application
Manual triage today
~30–45 min
With the agent
~6–10 min
Time saved
≈ 25–35 minutes saved per claim — about 165–230 hours a month at 400 claims, roughly a full-time adjuster. And on leakage: on a portfolio paying RM 2m/month, catching even 2 points of a 5–10% leakage rate is ~RM 40,000/month recovered — the pay-back on the build is measured in weeks, not years.
Indicative; actual time and recovery depend on claim type, document count and your current leakage rate.
Our agentic development process
Every use case follows the same seven stages — from framing the problem to production.
Our process, built on Anthropic's agent guidance and the Agent GPA eval framework: Building Effective Agents · Agent GPA
Stage 1 · Frame
The business problem
A Malaysian insurer or takaful operator receives each motor, medical or property claim as a pile — the claim form, the policy schedule, an incident or police report, workshop or hospital invoices, and photos. Most of the adjuster's day goes to the mechanical part: checking that the claim is actually covered, that the amounts match the receipts, and that nothing is padded or duplicated — not the judgement call. It is slow, it drifts as fatigue sets in, two adjusters apply different scrutiny to the same file, and the padding that slips through becomes claims leakage the business quietly eats.
Stage 2 · Map
The manual workflow today
- 01Open the claim file and each document
- 02Read the policy schedule to see what's actually covered
- 03Read the incident report and the receipts by eye
- 04Cross-check amounts, dates and cover against each other
- 05Chase missing documents or duplicate line items
- 06Only then make the settle / query / investigate call
Where it breaks: The bottleneck is reconciling the claim against the policy and the receipts — not the judgement. It's exactly the part that fatigues, varies between adjusters, and lets padding slip through as leakage.
Stage 3 · Design
The agentic workflow
We build it as an orchestrator-workers workflow: a lead model reads the claim, decomposes it into the checks this claim actually needs, and delegates each to a specialised worker — policy cover, receipt totals, report consistency, duplicate/fraud signals — then assembles one evidenced summary and escalates anything it can't clear to a human.
The orchestrator spins up only the workers this claim needs and gathers their findings — the orchestrator-workers pattern.
Watch a case flow through the agent — it does the reading and matching; a human still makes the decision.
See it in action
In the claim file
What the agent flags
- Claim within own-damage cover (RM 8,400 vs RM 15,000 limit)
- Duplicate line item: front bumper replacement billed twice (line 4 & line 9)
- Timeline inconsistency: invoice dated 8 Mar, 4 days before the 12 Mar incident
Every flag cites the exact document, line and figure · your adjuster makes the pay-out call · ~RM 0.80–1.60 per claim.
Step by step
- 1
Read the claim file
The agent ingests every document in the file — claim form, policy schedule, incident or police report, workshop or hospital invoices, and photos — reading text and, for scans and phone photos, using vision to extract the content.
- 2
Decompose the claim (the orchestrator)
A lead model works out which checks this specific claim needs — a motor own-damage claim, a medical panel claim and a property claim don't need the same tests — and plans the workers to run.
- 3
Run the workers
Specialised workers run in parallel: a policy-cover worker (is this peril, item and amount actually covered, within limits and sub-limits?), an amounts worker (do the receipt totals add up and match the claimed sum?), a consistency worker (do dates, parties and the report line up?), and a fraud-signal worker (duplicates, padding, edited documents, prior-claim overlap).
- 4
Reconcile and flag
The orchestrator gathers the workers' findings, reconciles them, and flags the two or three things that matter — an uncovered item, a duplicated line, a mismatched date — each with the figure and the document it came from.
- 5
Assemble a decision-ready file
It produces a one-page summary: the claim, the checks that passed, the recommended action (settle / query / investigate), and every flag with its evidence — ready for the adjuster to act on.
- 6
Hand to a human
The adjuster reviews the flagged items and makes the settle, query or investigate decision. Nothing is auto-paid.
Stage 4 · Build
How we build it with Claude
It's wired straight into your policy admin and claims system through connected tools, and it reads your live cover rules — limits, sub-limits and exclusions — per claim, so it checks against the policy as it actually stands, not a cached copy. It can recommend but never settles or moves money, and every flag it raises traces back to the exact document, line and figure it came from — a full audit trail built in from day one for a BNM/PIDM review.
Integrations
- Policy admin / claims management system
- Document store or claims intake inbox
- Optional: workshop / medical-panel rate references, anti-fraud / prior-claim lookups
Under the hood
- Pattern: an orchestrator-workers workflow on the Claude Agent SDK — the lead model decomposes the claim at runtime and delegates to workers, each tool exposed via an in-process MCP server.
- Context: the policy schedule, cover rules and the claim bundle are pulled into the window — text plus vision for scanned reports and receipt photos — with the live policy data (limits, sub-limits, exclusions) read fresh per claim, not from a stale copy, and long files compacted so nothing is silently dropped.
- Tools: read-only lookups (policy/cover reader, a claims-history reader for prior-claim overlap, a receipt/OCR extractor, a workshop/panel-rate reference) are kept separate from any write — the agent writes only a structured recommendation into the claims queue, never a pay-out.
- Guardrails: no autonomous money movement — it can recommend but never settle; schema-locked output; least-privilege with no write access to the policy admin or payment system; low-confidence or high-value claims route to a senior queue.
- Untrusted input: every document — invoice PDFs, report text, receipt images — is treated as untrusted and prompt-injection resistant, so a doctored invoice can't steer the recommendation; PII (IC, medical, vehicle, bank details) is handled PDPA-aligned.
- Recovery: where a worker's confidence is low or a document is missing, the orchestrator re-runs that worker or requests the piece, then escalates to the adjuster — it never guesses a settlement; every path ends at a human.
- Observability & audit: every tool call, worker finding and the final recommendation are traced and logged, and the adjuster's decision is captured against the file — a machine-readable audit trail a BNM/PIDM review can reconstruct, not a screenshot.
- Eval harness: a golden set of historical claims including known leakage and fraud, with deterministic checks (cover limits, arithmetic, dates) plus an LLM-as-judge on the fuzzy fraud calls, regression-tested on every change before it touches live claims.
Stage 5 · Architect
Single model or multi-agent?
A claim genuinely splits into independent checks — cover, amounts, consistency, fraud — so an orchestrator delegating to workers fits, but the workers are model calls within one process sharing the claim context, not separate autonomous agents — reliable, auditable, and far cheaper.
We keep the workers as tiered in-process calls (Haiku for simple checks, Sonnet to reconcile, Opus only for hard adjudication) rather than a full multi-agent fleet (≈15× the tokens) — reserved for genuinely open-ended, breadth-first work, which claim triage is not.
Which model does what
Read & extract from documents (incl. scans & photos)
Reliable vision and structured extraction on messy invoices, reports and receipt photos.
Orchestrate & reconcile across the claim
The workhorse — decomposes the claim, runs the workers and reconciles findings at moderate cost.
High-volume simple checks (duplicates, arithmetic, date sanity)
Cheapest and fastest for the bulk yes/no worker checks that don't need full reasoning.
Adjudicate ambiguous fraud / borderline cover
Invoked only when a worker flags low confidence on a hard call — a small fraction of claims.
What it costs — an estimate
AI usage — per claim
input documents + orchestration + parallel workers + output summary, at Claude Sonnet token rates
~RM 0.80–1.60
AI usage — at 400 claims/month
roughly half with the Batch API for overnight, non-urgent triage runs
~RM 320–640 / month
Build (one-off)
a focused single-claim-type agent is a few weeks of engineering; we quote fixed after a short discovery
scoped per integration
Ongoing
small next to the adjuster time it removes and the leakage it recovers
monitoring + support
Indicative only, shown in Ringgit at roughly RM 4.70 to the USD; per-token rates follow Anthropic's published pricing (confirm current figures). Actual AI usage depends on document sizes, claim type and monthly volume.
Stage 6 · Evaluate
How we measure success
- Deterministic checks first — cover limits and sub-limits, receipt arithmetic, date and party consistency
- Self-verification — the orchestrator reconciles and re-checks its workers' flags before surfacing them
- Business KPIs — adjuster time per claim, leakage caught, straight-through rate, consistency, and false-flag rate
We score the agent with the Agent GPA framework — Goal, Plan, Action, a current standard for agent reliability — on top of the human baseline. (reference)
Setting the baseline
First we measure the current manual triage on a sample — time per claim, how consistently two adjusters reach the same call, how much leakage slips through, and how many clean claims are needlessly delayed. That human baseline is what every agent metric is measured against, so the lift is provable rather than assumed.
What we test — Goal · Plan · Action
Agreement with a senior adjuster
Its summary, flags and recommended action match an expert's on held-out claims
Required workers dispatched
Every check the claim type mandates — cover, amounts, consistency, fraud — is run
No wasteful workers
It runs the checks the claim needs, not needless ones
Field & amount extraction accuracy
Did it read each figure — totals, limits, dates — correctly off the documents?
Leakage / fraud recall (catch rate)
Of the real over-payments and padding in the golden set, how many it flags
False-flag rate
How often it flags a clean claim as a problem
No hallucinated figures
Every number in the summary traces to a source document
The go-live gate
The agent does not touch a live claim until it beats the human baseline and clears these GPA targets on a golden set of historical claims — including known leakage and fraud. Every change is regression-tested against the same set before it ships.
Stage 7 · Deliver
How we'd deliver it
We start with a proof-of-concept on your real historical claims — including the padded and fraudulent ones — and prove it catches leakage without over-flagging clean files before it touches live claims. A focused single-claim-type rollout is a matter of weeks.
Free consultation
Is your claims team reconciling files by hand?
Tell us your claim type and volumes and we'll tell you honestly whether an insurance claims agent is worth building — and what it would recover.
Related use cases
Frequently asked questions
- Does the agent approve or pay claims?
- No. It reads, reconciles and flags, and recommends an action; a human adjuster makes the settle, query or investigate decision. It never moves money. Human-in-the-loop is essential for regulated insurance and takaful in Malaysia.
- How does it catch fraud and leakage?
- A fraud-signal worker looks for duplicated line items, padding above panel/workshop rates, edited documents, timeline inconsistencies and overlap with prior claims — each surfaced with the exact document and figure. It flags for a human to investigate; it does not accuse.
- How is it kept compliant with BNM, PIDM and PDPA?
- With a full audit trail so every recommendation can be reconstructed, and built around PDPA requirements for where claim and medical data is processed, stored and accessed. Treat this as a starting point for your compliance team, not legal advice.
- How long does it take to deploy?
- A proof-of-concept for one focused claim type is usually a matter of weeks; a full rollout takes longer, mostly because of integration and getting adjusters comfortable, not the AI.
Examples are based on real, anonymised engagements; details are generalised. Anchor Sprint is a member of the Anthropic Claude Partner Network — a deployment and rollout partner, not a reseller. This is general information, not legal or compliance advice.