Invoices that handle themselves.
A Nordic distributor processed invoices by hand across three systems that didn't talk to each other. We automated extraction and validation — and processing moved from days to minutes.
Three systems, none of them talking.
Finance re-keyed every invoice by hand between the ERP, the scanning tool, and a spreadsheet that had quietly become the source of truth. A single invoice could take days to clear, and month‑end was a scramble.
Worse, nobody could say with confidence what had been processed and what hadn't. Errors were caught late, if at all, and the process lived in a few people's heads.
- Manual data entry across three tools that didn't integrate
- No audit trail — exceptions surfaced days later
- Processing time grew with volume, not with headcount
Clarity first, then automation.
We started with a paid diagnostic: mapping the real workflow, exceptions included, until everyone described it the same way. Only then did we automate.
AI now extracts every invoice's data; business rules validate it before anything moves; and the ERP, scanner and ledger are connected so data moves once and matches everywhere. People still approve anything unusual — the model never decides alone.
- AI extraction with rule‑based validation on every document
- One integration layer across ERP, scanning and ledger
- Human‑in‑the‑loop approvals and a full audit trail
Days became minutes — and stayed there.
Invoice processing dropped from days to minutes, freeing roughly two people's full weeks — about 300 hours a month back (estimate) — and finance got something it never had: a process it can see, trust, and explain.
“We stopped firefighting invoices. Now we only look at the ones that actually need us.”
Why it had grown this messy.
The company had grown fast through acquisition. Each business unit brought its own tools, and finance ended up bridging them by hand because no single system covered the whole flow. The spreadsheet started as a stopgap and never left.
By the time they called us, two people spent most of their week moving numbers between screens — and any time off meant the queue backed up.
How the nine weeks ran.
Weeks 1–3 — Diagnostic. We sat with the team, mapped the workflow including every exception, and quantified where time actually went. We delivered findings and a prioritised roadmap — not a pitch.
Weeks 4–7 — Build. We built extraction and validation first, then the integration layer, shipping each flow into production only once it was validated against real invoices.
Weeks 8–9 — Handover. We instrumented monitoring, documented the system, and trained the team to own it. Exceptions now route to a person; everything else runs untouched.
The layer stays clean as they grow.
On a light retainer, we monitor what we built and take on the next workflow as it surfaces — purchase orders are already in scope. Because the process is documented and observable, adding to it is incremental, not another rebuild.