A small operations team inside a professional-services firm was responsible for turning incoming case files into structured coverage records. The work was careful, important, and almost entirely manual. The same figures were read off a PDF, typed into a case system, and then typed again into a reporting sheet, several times a day, by people far too capable for the task.
The brief
Stop the rekeying without replacing the systems the team trusted, and without turning a transparent process into a black box. The team had been burned before by a tool nobody could see inside, so legibility was a requirement, not a nicety. Whatever we built had to be something they could inspect, correct, and explain to an auditor.
The build
We built a pipeline that reads each incoming file, extracts the fields that matter, and writes them into both downstream systems in one pass. Every extraction is shown to a person before it commits, with the source document beside it, so a reviewer confirms rather than retypes. Low-confidence cases are flagged and queued; clean ones move straight through.
Nothing was hidden. Each step logs what it did in plain language, and the rules that decide what counts as low confidence are written where the team can read and adjust them. We worked in small visible increments, shipping the first useful slice in weeks and tightening it alongside the people using it.
What the team owns now
The team runs the pipeline themselves. They adjust the confidence thresholds when a new file format appears, read the logs when something looks off, and onboard new colleagues to it without us in the room. The rekeying is gone, and the hours it took came back. The reviewers spend their attention on the cases that genuinely need judgement, which was the point all along.
We're not on a retainer for this. That's the part we're proudest of.