Case note3 sections · 2 min readBy Dezső Mező · Published October 4, 2026.mdAutomationInvoicingCase studyOps
The manual process was four disconnected steps wearing a calendar reminder: export timesheets into a spreadsheet, type the invoice lines into the invoicing tool, hunt the right PO number and billing contact per client, send, and two weeks later discover three invoices missing or wrong. Each step was small; the compound cost was a half-day of senior time weekly and a two-week payment delay baked in.
The pipeline is deliberately unglamorous: a scheduled job pulls approved timesheets, builds invoice drafts per client with the right PO and contact pulled from a small mapping table, pushes them through the invoicing API, and posts a digest — what went out, what is blocked on what — to one channel. The only human step left is a two-minute review of flagged edge cases: a new client without a PO on file, an unusual amount, a missing contact.
01The pipeline is the boring parts, sequenced
Trigger on schedule, not on memory. Pull timesheets with their approval state — unapproved hours get flagged, not invoiced. Resolve the per-client quirks in a mapping table (PO, contact, payment terms, VAT handling) instead of encoding them in an employee. Draft, review flag, send, digest — each step logged so a failed run says what stopped.
02Design for the exceptions, not the happy path
The 90% of invoices that are routine should need nobody; the value is in the flagging, not the automation. New client without a PO? Flag, do not guess. Amount 3× the client's average? Flag. Customer record missing a billing email? Flag with the fix being one form field — so the exception fixes the data, not just the invoice.
03What actually changed for the business
Invoices go out Friday morning instead of Friday evening — and invoices sent two weeks earlier get paid two weeks earlier. The admin Friday became a review-and-coffee task. The unglamorous part is the durable part: no AI, no cleverness, just three systems connected and the exceptions routed to a human.
What to take away
- Four manual steps became one scheduled job: pull, draft, flag, send, digest.
- Design for exceptions — the automation handles routine, humans handle flagged edge cases.
- Invoices sent two weeks earlier get paid two weeks earlier.
- No AI needed — three connected systems and routed exceptions beat cleverness.
More from the lab
Browse all entries
From WhatsApp chaos to a real order pipeline
Self-hosted n8n or Zapier — where the bill actually differs
Enquiry to booked call, without the inbox
MCP servers in production, not in the demo
Want this looked at on your own system?Start a conversation
