300 invoices a month. That's not an AI problem.

At 5–10 minutes per invoice, 300 invoices a month is somewhere between 25 and 50 hours of manual entry.

Every month.

The workflow to automate this is not complicated. Gmail trigger. PDF parsing. AI extraction. GL mapping lookup. Import file generation. The pieces exist and they work.

The part people underestimate is the GL mapping table.

This is the file that tells the system: when you see an invoice from this vendor, it goes to this GL account. When the line item contains this description, it maps to this cost category. Without it, the AI has to guess the account coding every time. Sometimes it gets it right. Often enough it doesn’t, and a human has to correct it — which, at scale, is slower than not automating at all.

The mapping table is not exciting to build. It’s a spreadsheet with vendor names on one side and GL account codes on the other. But it is the foundation the rest of the workflow depends on.

I’ve built invoice processing integrations on top of e-Boekhouden and QuickBooks. The AI extraction model is usually solid within a week. The mapping table takes longer — because it requires someone who actually knows the chart of accounts to sit down and make decisions. That step cannot be automated away.

What you end up with, once the mapping is clean: incoming invoices process without anyone touching them, edge cases route to a short review queue, and the finance team spends their time on the exceptions instead of the routine.

The exception list gets shorter every month.

That’s what 25–50 hours of manual entry looks like when you stop doing it.