The invoice audit that found real money

The challenge

A small agency team ran weekly invoicing entirely by hand — one person re-typed the same data three times across a form, a logging spreadsheet, and a project board, with client tags regularly corrected from memory.

What we built

An invoice and timesheet automation designed around trust: AI acts only as a quality-check on human-entered data (never the decision-maker), a hard human-approval gate in Slack sits before anything client-facing, and the re-run-safe design means the pipeline can never double-bill. Rollout runs in parallel with the manual process for two full cycles before cutover.

The production workflow

The result

The discovery audit alone paid for itself — reading the live billing files caught a spreadsheet formula silently dropping a line item from a subtotal (17% understated) and two “source of truth” documents disagreeing on the same invoice total. Real errors humans had missed for weeks, found before go-live.

Anonymized, illustrative recreation of a real build.