Claims sit in queues because something is missing, not because assessment is slow. We automate the assembly — intake, classification, completeness, verification, routing — as separable services, and leave the settlement decision where it belongs.
Every guide to this topic lists fifteen use cases. Fifteen is not a plan. Sorted by what an error costs, they collapse into three groups — and the order of work falls out of that immediately.
Notification from every channel into one record. Claim-type classification, field and table extraction from documents and photographs, completeness against the pack for that claim type, cross-document consistency, duplicate detection, policy-in-force verification.
An error costs a correction. High volume, cheap mistakes, genuinely mechanical work — and the stage where elapsed time actually collapses. We take it first, in almost every engagement.
Triage and complexity scoring, straight-through eligibility against criteria written down in advance, assignment to the right handler, exception queues, and the per-claim evidence chain: inputs, model versions, sources, and who reviewed what.
An error costs a delay. Worth automating once assembly is stable. The evidence chain is cheap to build now and painful to reconstruct when a complaint lands eight months later.
Merit assessment against policy wording, quantum, approval, decline.
An error costs the payment — or the complaint, the ombudsman referral and a remediation exercise across every claim decided the same way. We will build it. We will first tell you why most insurers should not, and what to do instead.
Typical stack:
We say no to claims automation more often than this market does, so it is worth being specific. If any of these describes you, the project fails and we would rather say it now than in month four.
If the doubt is broader than this project — whether the company should be automating anything yet — that is a different question with its own answer: the free AI Readiness Score takes three minutes, and the AI Adoption Assessment gives a company-level verdict from $20,000, up to and including “not yet”.
Your compliance function will ask this, and the honest answer is more interesting than a yes or a no. Claims handling is not on the Annex III high-risk list — but under Article 6 that is a classification you have to document, not assume, and there are three ways to lose it by accident.
| What the component does | Status | Reference |
|---|---|---|
| Claims handling by a private insurer — assess, settle, decline | Not listed | absent from Annex III; classification must be documented |
| Granting, reducing or revoking benefits on behalf of a public authority | High-risk | Annex III, 5(a) — statutory and public schemes |
| Risk assessment and pricing, life and health insurance | High-risk | Annex III, 5(c) — the underwriting side |
| Claims outcomes feeding individual re-rating or renewal pricing | High-risk | 5(c) reached through the back door |
| Fraud, merit and pricing fused into one score | High-risk | nothing can be classified separately from the rest |
| Intake, extraction, completeness, authenticity checks | Outside | provided they are genuinely separable from the decision |
Which is why we build assembly, fraud screening and the decision as separate services with their own interfaces and their own contractual boundaries. Fuse them and you pull the whole document stack into whatever the strictest component attracts. One rule applies today with no ambiguity: GDPR Article 22 has governed solely-automated decisions with significant effects since 2018, and an automated decline is squarely within it.
Outside the EU the classification question mostly disappears — no comparable regime lists claims handling as high-risk — and is replaced by a narrower one that bites sooner: is this an automated decision about a person, and can you show your working? Several of these bind earlier than the EU’s.
| Where | Binding | Instrument, and what it asks |
|---|---|---|
| DIFC, Dubai | since Jan 2026 | Regulation 10 of the DIFC Data Protection Law — a system deciding about an individual without human involvement is a High Risk Processing Activity. Assess and document first; tell the claimant, and let them object. |
| Mainland China | since 2021 | PIPL Article 24 — right to an explanation and to refuse a decision made solely by automated means. |
| Hong Kong | in force | HKMA circulars on big data analytics and AI, and on generative AI in customer-facing use. |
| Singapore | supervisory | MAS FEAT plus the AI Risk Management guidelines — insurance models explainable enough for meaningful challenge and customer recourse. |
| Saudi Arabia | in force | PDPL with the SDAIA framework — human review checkpoints for automated decisions in high-risk contexts. |
Note what none of them care about: intake, extraction and completeness checking — the group where the money is. Every one of them cares the moment a system declines a claim without a human in the loop. Which is the same architectural line the EU analysis reaches from the other direction, and why the recommendation does not change with the jurisdiction.
Full breakdown of all five regimes, with the Annex III text →
Four phases, each with what we need from you. Generic phase names are how a proposal avoids committing to anything — so these are specific to claims, and so is the list of what stalls each one.
Three deliveries, each covering one of the three groups above. Client names withheld under NDA; sectors shown to indicate context. See full case studies →
The closest analogue to a claim file we have built: large, messy document sets read by a retrieval-augmented pipeline that drafts a structured memorandum with citations back to source, handling the routine bulk so specialists spend their time on the hard remainder. Plus LLM-assisted triage and next-step suggestions wired into the existing case-management system. Read the case →
A real-time classification pipeline with confidence-based routing: clear cases handled automatically, ambiguous ones escalated to a person. The mechanics of an exception queue that actually holds up under volume. Read the case →
Document authenticity, face match and liveness, built as its own service that hands a verified applicant to the process that decides. Exactly the separation a claims stack needs between “is this genuine” and “is this owed”. Read the case →
The whole pipeline can run on-premises or air-gapped — no claim document, medical report or photograph leaves your network. GPU hardware is quoted in the same contract, so residency is a deployment choice rather than a vendor negotiation. Details in security and compliance and private AI infrastructure.
Policy wording, deductibles, limits and exclusions are implemented as versioned deterministic rules, not learned from historical settlements. The same facts produce the same outcome, permanently — which is what you need when a complaint arrives months later and the model has been retrained twice since.
Fixed price against a one-page scope agreed before the build starts. No time and materials, no discovery that bills indefinitely, no scope that grows between the proposal and the invoice. We do not quote a range before that scope exists — a number produced before anyone has counted your claim types or seen your intake channels is a guess.
Scope first, price second, and both numbers are published. The specification is a product: the AI Solution Blueprint, one system, 5–6 weeks, from $30,000, credited in full against the build if implementation starts within 90 days. It walks your claim from notification to payment and ends with a build-ready spec, a fixed price and a timeline. Builds of this kind typically start around $150,000 over 4–6 months for a document layer covering one line of business. The exact figure comes out of the Blueprint rather than before it, because a number produced before anyone has counted your claim types or seen your intake channels is a guess, and guesses get repriced in month three.
Most engagements reach first working results in 2–4 weeks after a discovery and data-audit phase, then iterate to production. For claims that first result is the intake and completeness layer running on real historical files — including the night-time photographs and the scans of faxes, not a curated sample.
Claims handling is not listed in Annex III: point 5(c) covers risk assessment and pricing in life and health insurance, which is underwriting, and point 5(a) covers granting or reducing benefits only where the system acts by or on behalf of public authorities. For a private insurer the usual answer is not high-risk, but Article 6 requires that classification to be documented rather than assumed. We build so that it stays documentable — which mainly means keeping fraud, assembly and the settlement decision separable. Full breakdown here.
The classification question mostly disappears — no comparable regime lists claims handling as high-risk — and a narrower one replaces it that binds sooner. DIFC Regulation 10 has been in full enforcement since January 2026 and treats a system deciding about an individual without human involvement as a High Risk Processing Activity. China's PIPL Article 24 has given a right to refuse a solely automated decision since 2021. None of these regimes care about intake, extraction or completeness checking; all of them care the moment a claim is declined without a human in the loop.
Yes. The full stack can run on-premises or air-gapped, with no claim document or photograph leaving your network, and we can quote the GPU hardware in the same contract. For insurers with data-residency constraints this is usually the only configuration that clears legal review.
Usually we advise against it, and we say so before the proposal. Once you price the wrongful-decline tail properly — the complaint, the ombudsman referral, the remediation across every claim decided the same way — most of the remaining value is speed, and speed is cheaper to buy upstream by handing a complete verified file to a human. If you do want the decision automated we will scope it explicitly, not fold it into a document project.
That is the normal case, and it is the reason pilots on curated data do not survive production. We combine OCR with layout-aware models and pre-processing for photographs, scans and multi-column forms, and route low-confidence extractions to a human rather than guessing. The first phase runs on your real historical files precisely so the quality of the worst input is known before anything is committed.
The layer underneath: a capture method chosen per document type, not one method for the whole estate.
Explore →The same separation problem on the pricing side, where the high-risk classification is not optional.
Explore →The wider IDP practice: extraction, classification and drafting across any document-heavy process.
Explore →The long version: economics, rules versus models, and the full classification analysis.
Read →An engineer replies, not an account manager. You get back a one-page scope and a fixed price.
Want the spec first? AI Solution Blueprint — from $30,000, credited against the build.