A fuel retail group checks every station’s sales from till to bank, every day

A pile of files, and an afternoon per gap

Every station produces a pile of files each day: the till journal, card and e-wallet reports, fleet card and pay-at-pump records, subsidy claims and, a day or two later, the bank deposit. Each payment channel has its own rules. Some settle only on business days.

Someone in accounts matched them by hand, in spreadsheets, station by station. A small gap on one shift could take an afternoon to trace, because the answer might sit in any one of those files, or in a file that had not arrived yet.

That last case matters most. A file that has not arrived is not the same as a day with nothing to report, and a missing file must never be mistaken for a zero. The group needed every station checked every day, with the exceptions brought to the top and every number traceable back to the receipt behind it.

How we worked with them

We work as forward-deployed engineers: our engineers joined the team’s workflow in a shared channel and followed real work from start to finish. We traced one real station-day from till to bank with the accounts team, and learnt each payment channel’s rules, such as which ones settle only on business days.

When a rule was unclear, such as how to round or where a shift or a day ends, the owner decided and we wrote the decision down. Each issue got its own ticket. We built in short loops: every change went to a test copy first, behind a team-only login, and went live only when the owner said yes.

What we built

• Files that collect themselves. Duplicates and unknown files are turned away.
• A daily check for every station, from the till to the bank and the fuel subsidy, down to the shift.
• Gaps you can click. Each gap opens to the receipts behind it: amber if the money is still on its way, red if it does not add up.
• A completeness calendar and a watchdog that flag missing, late or failing files.
• A read-only MCP connector, so approved AI assistants can look up results safely.

AI coding agents (e.g. Claude, OpenAI) write and test each change, and an independent AI review checks it before release. Before a release, results are rebuilt from the raw files and compared.

Every gap traced to its shift and its receipts

The system is in production across the group’s stations, every day. Files arrive by themselves, and the completeness calendar and watchdog flag anything missing or late.

Each gap opens to the receipts behind it, and its colour says whether the money is simply on its way or genuinely does not add up. A gap on one shift can be traced to that shift and its receipts.

Channels that settle only on business days show amber until the money lands, so the team is not chasing gaps that are only timing.

All money totals come from one set of rules on the server, so the figures on every screen agree. A missing file shows as missing, never as a quiet zero, so nobody mistakes it for a clean day.