Atomiq logoAtomiq
Financial services

Document workflow for an underwriting team

Credit memos took nine working days, almost all of it analysts retyping figures out of PDFs into a template. We built structured extraction over the source documents and a review interface that puts an analyst on every field before it lands — because in this sector an unreviewed number is not a saving, it is a liability.

Sector
Financial services
Engagement
AI Consulting, then AI Automation
Duration
3 weeks, then 8 weeks
Constraint
Full audit trail, FCA-regulated entity
Stack
Client Azure tenancy
median memo turnaround
9 days → 4 hrs

median memo turnaround

of extracted fields accepted without correction
94%

of extracted fields accepted without correction

memo throughput per analyst
3x

memo throughput per analyst

of fields traceable to a source document
100%

of fields traceable to a source document

The problem

What was actually wrong.

A credit memo drew on filed accounts, management accounts, bank statements and a covenant schedule — four document families, none of them consistently formatted, several of them scanned.

Analysts spent the majority of a nine-day turnaround on transcription rather than judgement, and transcription errors were the most common cause of a memo being sent back at committee.

As a regulated lender, every figure in a memo has to be traceable to its source document and every decision attributable to a named person. That ruled out anything that produced a number without a provenance chain.

What we did

The approach, in the order it happened.

01

Consulting first, because the bottleneck was contested

A three-week engagement established that the constraint was transcription rather than analysis — which was not what the leadership team had assumed going in. Two other candidate use cases were sized and explicitly dropped.

02

Extraction with provenance on every field

Each extracted value carries a pointer to the page and region it came from. The reviewer sees the source alongside the figure, so checking a field takes seconds instead of requiring the original document to be reopened.

03

A review UI built for the analyst, not the demo

Low-confidence fields are surfaced first and the keyboard path through a memo was designed with the two analysts who would use it most. Nothing reaches the template without a human accepting it.

04

Audit logging as a first-class output

Every extraction, every correction and every acceptance is logged with a timestamp and a named user. Internal audit signed the design off before build started rather than reviewing it afterwards.

Outcome

What changed. Measured against the baseline agreed at scoping, over at least a full quarter of production running.

Results

  • Median turnaround on a credit memo fell from nine working days to roughly four hours of analyst time.
  • Ninety-four per cent of extracted fields are accepted without correction; the remainder are flagged by confidence before a human sees them.
  • The team took on a materially larger deal pipeline without adding analysts.
  • Internal audit signed off the provenance and logging design before build, and the first regulatory review passed without a finding against the system.
We were sold document AI three times before this. The difference was that someone insisted a human stayed on every field.
Head of Underwriting
Get started

Let’s find out what AI is actually worth to you.

A 30-minute call, no deck. Bring the workflow that frustrates you most and we will tell you whether it is worth automating.

Typically replies within one business day.