From file-by-file to review-by-exception

Context

An insurance and extended warranty company with a high volume of claim settlements coming from technical service providers. Every file arrived with several attached documents: work orders, invoices, supporting documentation, sign-offs and attachments in different formats.

The documentation came in heterogeneous — scanned PDFs, images, files renamed with no standard, and supporting evidence following different criteria depending on the provider. The operations team reviewed every file manually before approving, flagging or escalating the case.

The problem

Validation depended on multiple manual controls: cross-checking amounts, verifying documentation, spotting inconsistencies and applying rules that changed with the type of service.

But the main problem was not volume. Much of the operational judgement had never been formalized. There were exceptions that depended on who reviewed the file, rules that lived in the experience of a few people, and different criteria depending on the type of technical service. As volume grew, scaling the process meant adding operational effort and keeping critical knowledge informally distributed across the team.

What we did

We built a document validation solution with AI agents: the agents cross-check work orders, invoices and supporting documentation, interpret unstructured information, validate business criteria and generate automatic observations. When they find inconsistencies or ambiguous cases, they escalate the file with context and observations already written; simple cases move forward automatically.

• Specialized agents that understand complete files. We did not automate isolated steps — the agents interpret the case end to end.
• Processing of unstructured documentation. The technical service providers were not going to change how they send information; the solution adapts to the existing reality.
• A review-by-exception model. The operator stops reviewing file after file and steps in only when the case calls for human judgement.
• Operational documentation, written as we went. During the project we identified and formalized exceptions and rules that had never been written down.

What we ruled out

An automation based only on rigid rules and traditional workflows. It would have worked for simple scenarios, but the process depended too much on variable documentation and on implicit knowledge spread across the operations team.

We also ruled out requiring structured formats from the external technical service providers: solving the problem by forcing changes on third parties would have added more friction than value.

Why it worked

• We understood complete files, not isolated steps.
• People step in only where their judgement matters, not as routine reviewers.
• During the project we formalized operational criteria that lived in the team’s heads.
• The solution adapted to how third parties already worked, without asking them to change.

Stack

• Specialized AI agents
• Unstructured document processing
• Operational rules engine
• Integration with existing systems

Outcome

• Files pre-analyzed before they reach an operator
• The team stopped reviewing simple files from scratch
• Operators focused only on exceptions and complex cases
• Operational criteria formalized and documented during the project
• Critical knowledge that no longer depends on individual people