Claims document automation with OCR and AI

Real company context

GNP Seguros is Mexico's largest insurer by premium volume, with 12.2% market share and over 120 years of history, part of Grupo BAL. On June 4, 2026, GNP announced expanding a multi-year agreement with Palantir Technologies to deploy its Foundry and AIP (Artificial Intelligence Platform) platforms across health, life, auto and property — aiming to identify anomalous patterns in claims, detect possible fraud attempts before payments are made, and strengthen risk management, always under human oversight. With this agreement, GNP became Palantir's first publicly announced commercial client in Latin America.

According to Community of Insurance's report on the agreement, settlement teams get immediate access to full context on the insured, the policy, and claims history, resulting in faster processing and fewer judgment errors. GNP is also building an intelligent decision engine with AIP that replaces systems where a single pricing-rule change could take weeks or months to approve, per that same report.

What's publicly confirmed: data unification, preventive fraud detection, underwriting support, and faster context available to the claims team. What isn't publicly detailed: the specific mechanism for capturing and extracting data from claim documents (OCR) — a typical technical component of this kind of architecture, not a capability explicitly announced by GNP or Palantir in available sources.

The challenge (characteristic of claims operations at this scale)

• Massive document volume per claim: medical invoices, expert reports, official records, damage photos, handwritten or scanned claim forms.
• Manual capture of these documents as the classic bottleneck before any decision engine (like AIP) can reason over the case — the system is only as fast as the quality and speed of the data it receives.
• Need for regulatory traceability and auditability (CNSF) at every automation step.

Illustrative breakdown: how an OCR + document-AI component would fit

Proposed architecture, not confirmed by GNP — the type of layer that typically feeds a system like Foundry/AIP in a claims flow.

• Multichannel capture: the insured uploads photos/PDFs of the claim from the app or portal; OCR extracts text from invoices, records, and reports automatically.
• Automatic document classification: a model distinguishes document type (invoice, record, ID, expert report) without manual intervention, and verifies the file is complete before moving to review.
• Structured extraction: key data (amounts, dates, folios, diagnoses, policy number) get extracted from the document and loaded directly into the digital file — no manual capture by an adjuster.
• Cross-validation: the system compares extracted data against the policy and the insured's history, flagging inconsistencies for human review before any payment.
• Human in the loop: no payment is released automatically; the adjuster reviews alerts and approves, consistent with the "always under human oversight" approach GNP has publicly stated for its AI use.

Typical metrics this type of component targets

Illustrative industry-reference values — not figures reported by GNP.

Metric	Typical industry reference
File capture time	From days to under 48 hours
Manual capture errors	Significant reduction by eliminating manual typing
Fraud detection before payment	Shift from reactive to preventive detection