AI inside an existing workflow: automating document intake for an insurance back office

The client

An insurance services firm in Paris processing claims at volume. Every claim arrived as a bundle of PDFs and photos by email. Staff read each document, typed fields into the claims system, and routed the file to the right handler.

The problem

The bottleneck was intake, not judgment. Skilled handlers spent most of their day transcribing, and the backlog peaked exactly when customers were most anxious, right after storms and accidents. The firm did not want a new system. It wanted its existing workflow to stop eating its staff's time.

The constraints

French insurance data. Claims bundles contain health and financial data under GDPR and sector rules. The pipeline had to run on the firm's existing infrastructure with documents never leaving their environment.

Regulated accuracy expectations. A wrong field in a claim is not a UX bug, it is a compliance event. The system had to prove accuracy per document type before being trusted, and stay proven after.

Staff adoption, not staff replacement. The handlers' process, screens, and routing rules were non-negotiable. Anything that required retraining would have failed politically before it failed technically.

AI where it helps, humans where it matters

A pipeline classifies each incoming document, extracts the fields the claims system needs, and scores its own confidence. High-confidence extractions go to a one-glance review screen. Low-confidence ones fall back to exactly the manual process staff already knew. Nothing enters the claims system without a human approving it or a sampling audit covering it.

No rip-and-replace

The claims system, the email intake, and the handlers' routing rules all stayed. The AI layer sits between the inbox and the back office, writing through the same interfaces staff used. Adoption took days, not months, because nobody had to learn a new system.

Measured before trusted

The pipeline ran in shadow mode first: extracting everything, entering nothing, and being scored against what staff typed. It went live per document type only after beating a set accuracy bar on that type. Accuracy is monitored continuously, and any drift below threshold pulls that document type back to manual.

Documents never leave the building

Classification and extraction run on the firm's own infrastructure. The model is not called as an external API carrying customer documents; the data stays inside the compliance boundary that already governed the claims process.

An audit trail for every field

Each extracted value carries its source document, its confidence score, and who reviewed it. When an auditor asks why a claim holds a value, the answer is one lookup, not an investigation.

Handlers review, not re-type

The review screen shows the extracted field beside the source text so approval is a glance, not a re-entry. Handlers were already checking work; the pipeline just gives them less of it to check.

Intake in hours, not days

Documents move from inbox to claims system the same day they arrive. The storm-day backlog that used to build up now clears while customers are still waiting on it.

Handlers back to judgment

Staff hours moved from transcription to claim handling. Adoption took days because the only change in their workflow was that less typing appeared in front of them.

Trust that is measured, not asserted

Every field is reviewable to its source, accuracy is monitored per document type, and drift pulls work back to manual automatically. The compliance team signed off on exactly those properties.