Representative engagement
An agent workflow reads counterparty confirmations, matches them against booked trades and drafts replies to discrepancies. Operations now handle the exceptions and no longer rekey.

- Client
- A Geneva commodities trader
- Sector
- Commodities Trading & Shipping
- Duration
- 6 weeks to production
- Services
- Agentic Systems · AI Operations
01Challenge
The operations team received several hundred confirmations a week by email, as PDFs, scans and free text, in English, French and occasionally Spanish. Each was compared field by field with the trade booked in the ETRM. Quantity, tolerance, price formula, delivery window, incoterm and payment terms.
Matching was manual, done under time pressure and by a few experienced people. Discrepancies were sometimes found late, after documents had gone to the bank. Hiring more operators was the default plan.
02Approach
A two-person forward-deployed pod sat with the operations team for the first week, collecting 240 historical confirmations with their outcomes. From these we wrote the eval set before building anything. For each document it holds the correct fields and the correct match decision.
We shipped to two operators in week three and to the whole team in week five. Agents handle extraction and matching. Uncertain or mismatched items go to a review queue with the evidence highlighted. Operators approve or correct in one click. Each correction is added to the eval set.
03The system
Agents watch the confirmations inbox, extract terms, query the ETRM through a read-only connector and classify each confirmation as matched, mismatched or uncertain. For mismatches, a reply to the counterparty is drafted for an operator to send.
Components
- 01Inbox watcher and document parser for PDFs, scans and email bodies
- 02Extraction agent with field-level confidence scores
- 03Read-only ETRM connector and matching rules agreed with operations
- 04Review queue with highlighted evidence and one-click approve or correct
- 05Eval suite of 240+ real cases run on every release
- 06Audit log of every extraction, decision and approval
Stack
- Hosted LLM (Swiss region)
- Python
- PostgreSQL
- ETRM read-only API
- Document OCR
- Eval harness
04Results
A Geneva commodities trader
- −62%
- Handling time per confirmation
- 81%
- Confirmations matched without manual rekeying
- 3 days → same day
- Median time to flag a discrepancy
- 0
- Additional hires planned for the next volume increase
Representative engagement. Client withheld under NDA. Details are anonymised and combined so that no client can be identified; metrics are those measured on the engagement.
Related work
Next step
Tell us which process takes your team the most time. In a thirty-minute call, we can tell you if the sprint is a good start.


