Generative AI · Insurtech & legal
Document validation with generative AI for claims and legal teams
The claims and legal teams of a US insurtech spent hours reading long documents to extract critical data and assess risk. We built a system that does the first read for them — with verifiable citations to the original document, not summaries you have to take on faith.
Problem
Claims and legal professionals spend hours reviewing long documents to extract critical information and assess risk.
Result
AI does the first read with verifiable citations: 98% agreement with human analysts, measured over two weeks of side-by-side evaluation.
The context
Policies, endorsements, demand letters, expert reports: long documents, inconsistent formats, often scanned. The critical information — coverages, exclusions, dates, amounts — is in there, but buried. And the person digging for it is a professional whose time is expensive.
The risk is not just the cost of those hours: it is what slips through. One unseen exclusion or one misread date can change the outcome of an entire case. The manual process was not just slow — it was hard to audit.
How it works
The system receives the document, digitizes it if needed (OCR) and runs it through language models with specialized instructions per document type. The output is not a summary: it is a structured extraction — every field with its value and, crucially, the exact reference to the passage where the document says it.
The hard part in this kind of system is not summarizing: it is not making things up. The entire architecture is designed around verifiability — the AI proposes, points to where it read it, and the professional confirms in one click. The machine does the first read; the human signs. And the final test was not a demo: the system ran for two weeks in parallel with the analysts, answering the same cases, and reached 98% agreement with the human answers before taking over the work.
Engineering decisions
Decision 01
Citations, not blind trust
Every extracted data point links to the original passage. The professional does not have to re-read everything to trust the output: they verify exactly what was flagged. That traceability turns AI into an auditable tool, not a black box.
Decision 02
More than one model, with failover
We use Claude and GPT depending on the task, with automatic fallback if a provider fails or degrades. A production system processing real cases cannot depend on a single vendor’s availability.
Decision 03
Security before speed
Legal and insurance documents are sensitive data: encryption in transit and at rest, role-based access controls, and one hard rule — client data is never used to train models.
What changes for the business
The operational shift is direct: the expert stops reading entire documents and starts validating flagged findings. Their judgment — the truly expensive part — concentrates where it matters.
- 98% agreement with human analysts’ answers — two weeks of side-by-side evaluation before going to production.
- Drastic reduction in analysis time per document: from hours to minutes.
- Lower omission risk: the machine does not get tired on page 40.
- Scales through volume peaks without hiring or sacrificing quality.
- Every data point traces to its source: the entire process is auditable.
Where else it applies
Insurance
Claims, underwriting, and analysis of policies and endorsements.
Legal
Contract review, due diligence, document discovery.
Banking
KYC, credit file and collateral analysis.
Audit & compliance
Mass document verification with traceability.
FAQ
Frequently asked questions about this case
How do you prevent the AI from inventing data that is not in the document?
Every extracted field comes with a citation to the original passage, and the workflow requires human validation before any decision. If the model cannot find a data point, the system says so — it does not fill it in.
Does it work with scanned documents or photos?
Yes. The pipeline includes OCR to digitize scans and photographs before analysis. Original quality matters, but the system is designed for the real world, not for perfect PDFs.
What about document confidentiality?
Encryption in transit and at rest, role-based access, and client data is never used to train models. The system was designed for a regulated US industry from day one.
Does it handle documents in Spanish?
Yes. Current language models work equally well in Spanish and English; the per-document-type instructions adapt to the local jurisdiction and format.
Have a similar challenge?
Tell us your problem. We'll tell you how to solve it — and how long it takes.