Illustrative example. This case study describes a typical engagement of this kind, not a specific client. Figures are representative of the work, not a guarantee of results.
AI and automation · Insurance
Quote preparation cut from five hours to under one at a commercial insurance broker
A 45-person commercial insurance broker was losing deals to slow quotes because account managers spent most of their week retyping client documents. A document AI pipeline, with people reviewing anything uncertain, gave them that time back.
Results
Measured over the first full quarter after roll-out.
The problem
New business arrived as emails with attachments: proposal forms, schedules of assets, claims histories and financial statements, in every format imaginable. Account managers retyped the details into the broker management system, then again into each insurer's submission format.
Each submission took about five hours. With around 120 a month, that was most of the account management team's week. Quotes took two to three days, and the business knew it was losing clients to faster competitors.
Earlier attempts with off-the-shelf OCR had failed on the variety of documents, and the team was sceptical that AI would do better.
What we did
Discovery sprint, three-week pilot, then a six-week roll-out.
Week 1: Discovery
We sat with account managers, timed the process end to end and collected 150 past submissions with the correct values. Together we agreed the 40 fields that mattered and an accuracy target for the pilot.
Weeks 2–4: Pilot
A pipeline using Amazon Textract to read documents and a language model on Amazon Bedrock in the Sydney region to extract and normalise the fields. Every field carried a confidence score, measured weekly against the 150 examples.
Weeks 5–10: Roll-out
Automatic intake from the new-business inbox, a review screen that highlights uncertain fields against the source document, an integration that writes approved data into the broker system, and an audit log of every change.
Handover
Runbooks, recorded walkthroughs and training for the operations lead, who now owns the evaluation set and reviews accuracy each month.
Stack
What made the difference
The evaluation set did more than measure the system. Agreeing on the correct answer for 150 real submissions exposed inconsistencies in how the team itself recorded data, and fixing those improved quality before any AI was involved.
Human review was what made adoption work. Account managers trusted the system because they could see exactly which fields it was unsure about and check them against the source document in one click.
Keeping everything in the broker's own AWS account in Sydney made the privacy conversation with their insurers short.