Field Test 001 · Two-week live report
Can an AI agent reconcile business receipts?
Marc replaced the recurring Friday receipt chase with an agent that checks card transactions each morning, finds supporting documents and sends routine matches to one approved internal accounts destination.
The human problem
The email that made Friday feel heavier.
One missing-receipts email usually meant another would follow. A chase could take around an hour, cover several weeks and still leave an invoice or attachment behind.
The real cost was not just the hour. It was the recurring feeling that the work was accumulating and never fully under control.
The goal was not “automate finance”. It was narrower: keep the receipt queue moving and return uncertain work to Marc.
Verified historical baseline
These figures describe the historic problem baseline. They are not transactions processed by the current agent.
The live workflow
Routine work moves. Exceptions come back.
Sammi, the operational agent inside Grok Bot, checks the company-card system proactively every morning rather than waiting for accounts to report a problem.
Review current card transactions and reconciliation state.
Search approved email sources for the invoice or receipt.
If it is absent, use an approved supplier portal.
Validate the document against the transaction and avoid duplicates.
Send routine matches to one approved internal accounts destination.
Return unidentified, blocked or unauthorised items to Marc.
Report when the actionable queue is clear.
The handbrake
Automation with a declared boundary.
A live exception showed the control model clearly. When accounts could not open a PDF, the agent prepared a replacement image but stopped for Marc's approval before sending it.
The agent may
- View approved transaction information.
- Search approved email and supplier portals.
- Download likely invoices and receipts.
- Deliver routine matches internally.
- Record its action and report status.
The agent must never
- Make payments or approve expenditure.
- Alter accounting or card records.
- Code transactions or make tax judgements.
- Send to arbitrary or external recipients.
- Guess when evidence is missing.
Exceptions returned to Marc
Unknown transactions, unavailable logins, documents held by colleagues, new access requests and any judgement outside the trained routine.
Result after two weeks
Everything actionable cleared. One external hold.
As of 2026-08-25, every item the agent could currently action had been completed. One item remained open because a colleague still needed to provide the source document.
The workflow still interrupts Marc for a short exception, but it has replaced a larger recurring batch chase with smaller, bounded decisions. Marc reports less stress and more confidence that the work is in hand.
Early result: routine receipt work is automated; uncertainty and authority decisions still belong to the owner.
This is a two-week field report, not a completed monthly close. No percentage or annual time-saving claim is made.
Provisional verdict
Useful enough to continue. Not proven enough to disappear.
The agent has earned continued use inside its narrow boundary. The next proof point is a complete statement period: receipts found without Marc, exceptions, owner minutes, missing items at close, wrong matches and duplicate sends avoided.
Build the controlled version