Clinical document AI for a provincial health agency

Clinical staff were spending a substantial share of their week extracting structured data from documents so that it could enter provincial systems. The work was necessary, entirely manual, and among the least clinically valuable use of clinically trained people available. The agency had considered automation twice and stopped both times on the same two grounds: an extraction error on clinical data carries patient consequence, and health information could not leave the province, let alone the country.

We built an intelligent document processing capability with human review designed into the workflow rather than bolted on. Documents are classified, key clinical and administrative fields extracted, and every extraction assigned a confidence score. High confidence extractions flow through. Anything below the threshold routes to human review, and the threshold was set conservatively at the agency's direction and tuned only on evidence.

The annotation work that trained the extraction models was performed by Corpshore's medical and life sciences annotation specialists, working on data that never left the province, under agreements reviewed by the agency's privacy office. Bilingual capability was built in from the start, since documentation arrives in both official languages.

Governance was explicit: model performance is monitored continuously, drift triggers review, every automated extraction is auditable to its source document, and the agency retains a documented override path. The implementation was mapped against provincial health privacy legislation and the agency's own privacy impact assessment process at design stage.

Fixed scope implementation over twenty-six weeks followed by a managed operation covering monitoring, human review, model maintenance and continuous improvement. All processing in Canada, in the province.

Straight-through processing reached 68 percent of document volume by month four, with the remainder routed to human review as designed. Clinical staff time on document extraction reduced by a majority against baseline, and was redirected to clinical work. Extraction accuracy on automated fields exceeded the agency's acceptance threshold in every monthly audit across the first operational year. Turnaround on provider submissions improved substantially. Zero substantiated incidents of incorrect clinical data entering provincial systems through the automated path, which was the agency's primary acceptance condition.

The constraint was the design. Setting a conservative confidence threshold produced lower automation than an aggressive one would have claimed, and it produced a system a health agency could actually put into production. The agency did not want maximum automation. It wanted automation it could defend.