Reducing a four-hour statement review to minutes with automation
A workflow extracting, classifying, validating and analysing financial statements while keeping analyst sign-off in place. This is an example workflow design; the exact approach and results depend on the organisation’s process, data and systems.
- INDUSTRY
- Financial services
- SERVICES
- AI & Agentic Workflows, Internal Tools
The problem
Financial services teams often receive statements as PDFs and spreadsheets in inconsistent formats. Each review pack can mean re-keying figures into a standard template, checking totals, and writing a short commentary.
Most of that time can go on transcription and checking, while the judgment work takes only a fraction of the review.
Typical manual workflow
- 01Statement arrives by emailmanual
- 02Analyst downloads and renames itmanual
- 03Figures re-keyed into templatemanual
- 04Totals checked by handmanual
- 05Commentary writtenmanual
- 06Senior reviewmanual
Dashed steps are typically done by hand.
The approach
A pipeline extracts line items from each statement, maps them to the firm’s chart of accounts, and checks that totals reconcile. A language model drafts commentary from the structured figures rather than the raw document.
Analysts work from a review queue where each item shows the source page beside the extracted value, making checks faster without removing final approval.
Example system design
Design decisions
Why final approval stays human
The review may be a regulated output. The system prepares and checks; an analyst signs. Corrections are logged and can improve extraction rules.
Why the model never sees the decision
Commentary is generated from validated numbers only. If validation fails, no draft is written, removing a class of confident but wrong output.
Why keep the existing template
Clients and auditors may already know the format. Keeping it avoids unnecessary change management while the underlying process improves.
Failure handling
Example outcomes
| MEASURE | MANUAL PROCESS | AUTOMATED PROCESS |
|---|---|---|
| Time per pack | 4h 00m | 0h 31m |
| Re-keyed figures | ~240 | 0 |
| Analyst touchpoints | 6 | 1 |
Technology
- Python
- PostgreSQL
- Claude
- n8n
- Google Workspace