Healthcare · 2026

An AI agent that cleared a healthcare provider's document backlog

Insurance pre-authorisation documents were being read and keyed by hand. An agent with human review in the loop now handles the first pass.

AI AgentsAI/MLA multi-city diagnostics network · named under NDA

26h to 4h
Average turnaround
Measured from document receipt to verified data in the hospital system.
91%
Fields auto-extracted
Cleared the confidence threshold with no human correction required.
99.2%
Extraction accuracy
On the held-out evaluation set, after human review of escalated fields.
8 of 11
Team redeployed
Moved from data entry to patient coordination roles.

The challenge

What was actually going wrong

1
A team of eleven spent most of each day extracting fields from scanned pre-authorisation forms.
2
Average turnaround was 26 hours, which delayed patient procedures and drew complaints.
3
Formats varied across nineteen insurers, so rule-based extraction had already failed twice.
4
Patient data could not leave the provider's own Azure tenancy under any circumstances.

Our approach

Sequence mattered here as much as the individual decisions. Each step below made the next one safe.

01
Deploy entirely inside the client's tenancy
The agent runs on Azure OpenAI inside the provider's own subscription, so no patient data crosses a boundary the compliance team had not already approved.
02
Extract with citations, not confidence alone
Every extracted field carries a bounding-box reference to its source in the scan, so a reviewer verifies in seconds instead of re-reading the document.
03
Route by confidence to a human
Fields below the confidence threshold are escalated rather than guessed. The agent is explicitly permitted to say it does not know.
04
Build an evaluation set before building the agent
Six hundred manually labelled documents across all nineteen insurers became the regression suite. No prompt or model change ships without scoring against it.

The results

Each figure below was measured after the engagement, not projected during it.

6x

faster document turnaround

Average turnaround
Measured from document receipt to verified data in the hospital system.
26h to 4h
Fields auto-extracted
Cleared the confidence threshold with no human correction required.
91%
Extraction accuracy
On the held-out evaluation set, after human review of escalated fields.
99.2%
Team redeployed
Moved from data entry to patient coordination roles.
8 of 11

PLACEHOLDER: client quote pending written approval. The team stopped being typists and went back to being coordinators.

Chief Operating Officer · Diagnostics network

Services used

Related work

We will tell you what we would repeat from this engagement, what we would change, and what it is likely to take.