OCR plus LLM reading
Two layers working together, so the system reads context rather than fixed positions and pulls the right field from clean PDFs, crumpled scans, and phone photos alike.
Loading…
Loading…
Insurance
A regional insurer
Walk into the back office of this regional insurer two years ago and you would have seen the same scene every morning. Trays of claims forms, policy documents, and medical reports stacked beside every desk, and a row of people quietly typing the contents into the core systems by hand. More than 30,000 documents arrived every month, and almost all of them passed through a human keyboard before anyone could act on them.
That setup worked when the book of business was smaller. By the time they came to us it had stopped working. A single complex claim could sit in a queue for days simply waiting to be keyed in. Typos crept into policy numbers and claim amounts, and every error meant a rework loop somewhere downstream. The team was good at its job, but the job itself had become a bottleneck the business could feel.
| Before | After | |
|---|---|---|
| Document intake | Keyed in by hand, document by document | Read and extracted automatically |
| Processing time | Days for a complex claim | Down 65 percent across all volume |
| Extraction accuracy | Eroded by typos and rekeying | 99 percent, confirmed by review |
| Where staff spend time | Typing trays of paper | Reviewing exceptions and real claims |
The honest problem was math. Volume was climbing faster than they could hire, and every new policy sold added more paper to the same pile. Hiring more people to type faster was expensive, slow to train, and did nothing about the error rate. Replacing the core claims and policy platforms was off the table too, because that kind of project carries real risk and the insurer could not afford to disrupt live operations to chase an efficiency gain.
So the brief we agreed on was narrow and clear. Take the manual keying out of the loop wherever it was safe to do so, keep a human in charge of anything uncertain, and do all of it without touching the systems of record. The documents themselves were the hard part. They came in as clean PDFs, as crumpled scans, and as phone photos taken in bad light, and the same field could live in a different place on every layout.
We built an AI native document pipeline that reads each incoming document, pulls out the fields that matter, and hands clean structured data to the systems the insurer already ran. OCR turns the image into text, and an LLM extraction layer makes sense of that text in context, so a claim amount is recognised as a claim amount whether it sits in a table, a paragraph, or a stamped box in the corner.
The piece that earned trust was confidence scoring. Every field the pipeline extracts carries a score, and that score decides what happens next. High confidence data flows straight through to the claims and policy systems with no human touch. Anything below the threshold is routed automatically to a reviewer, who confirms or corrects it in seconds rather than rekeying a whole document. A monitoring dashboard sits over the top so the operations team can watch throughput, accuracy, and exceptions in real time.
Two layers working together, so the system reads context rather than fixed positions and pulls the right field from clean PDFs, crumpled scans, and phone photos alike.
Every extracted field carries a score. Clear cases run straight through, and only the genuinely uncertain ones reach a reviewer.
Clean structured data flows directly into the existing claims and policy systems, with no rekeying and no new platform to learn.
Nothing uncertain ships unseen. A reviewer confirms or corrects flagged fields in seconds, which is what holds accuracy at 99 percent.
A dashboard sits over the pipeline so the operations team can watch throughput, accuracy, and exceptions in real time.
The whole pipeline sits in front of the systems of record, so the insurer captured the gains without a risky platform replacement.
The threshold that decides what goes straight through and what gets reviewed was not a number we set in a meeting room. We spent a week sitting with the operations team, watching what they flagged and why, and tuning it together. One afternoon a senior reviewer pointed at a stamped claim amount the early pipeline had read with high confidence, and she was right to be wary, the stamp had smudged a digit. That single catch reshaped how we scored stamped fields, and it was the moment the team stopped seeing the pipeline as a black box and started treating it as a colleague they could correct.
Within a couple of months the morning trays were mostly gone. Processing time across the full volume fell by 65 percent, and field level extraction accuracy settled at 99 percent because the human review was now aimed only at the cases that needed it. The same back office team stopped typing and started reviewing exceptions and working the more valuable claims, which is the work they were hired to do in the first place. The build paid for itself in operational savings in about two months.
documents a month
faster processing
extraction accuracy
payback period
Paid for itself in about two months
Most of the savings showed up as time the back office got back. The build covered its own cost in operational savings inside roughly two months, and the gains kept compounding as volume grew.
Humans where they matter
Nothing uncertain ships unseen. The pipeline automates the clear cases and routes the rest to a reviewer, so accuracy stays high while the bulk of the volume runs on its own.
We used to measure the backlog in trays on people's desks. Now my team reviews the handful of documents the system is unsure about and gets on with real claims work. It paid for itself before I had finished explaining it to the board.— Head of Operations, a regional insurer
FAQ
No. The automation layer sits in front of the existing claims and policy systems and writes to them straight through, so the insurer captured the efficiency gains without a risky core platform replacement.
Every extracted field carries a confidence score. High confidence data flows straight through, and anything below the threshold is routed to a human reviewer. That approach held accuracy at 99 percent while still automating the large majority of the volume.
The OCR and LLM layers work together so the system reads context rather than fixed positions, which lets it pull the right fields from crumpled scans and badly lit photos. When an image is genuinely unreadable, the confidence score drops and the document is sent to a reviewer instead of being guessed at.
Book a free consultation with our engineers.