I recently audited an admin team at an industrial company. Three people spent about 4 hours a day doing the same thing: opening emailed invoice PDFs, reading the data, looking up the vendor in the system, typing everything into the ERP, checking it adds up. Fifteen minutes per invoice when things went smoothly. When there was an error — price mismatch, different quantity from the delivery note — a single invoice could eat up 30 minutes.
The absurd part is that Microsoft already includes an AI in its licenses that does this in seconds. It's called AI Builder, it's been available for years, and almost nobody knows about it.
How AI Builder works with invoices
AI Builder has a pre-trained model for invoice processing. You pass it a PDF — or a photo, or a scan — and it automatically extracts invoice number, date, vendor, tax ID, detail lines with quantities and prices, taxes, total. Nothing to train to get started. The model comes ready and works with invoices in multiple languages and formats. I've processed invoices from vendors in 4 different countries with the same flow without touching a thing.
Is it perfect? No. But it gets it right 85-90% of the time. And for the rest, it flags low-confidence fields so a person can review — with data already pre-filled, not starting from scratch.
The flow I typically build
Invoice arrives by email (or gets uploaded to a SharePoint folder). Power Automate detects the attachment and sends it to AI Builder. AI Builder returns all structured fields. The flow cross-checks that data against the vendor database and payment terms. If everything matches, it registers directly in Dataverse or sends it to the ERP via connector. If something doesn't line up — unknown vendor, amount out of range — it sends it for human review with all the information ready.
The admin team goes from typing data to managing exceptions. Which is, incidentally, where they actually add value.
Real numbers
In that implementation I mentioned at the start, time per invoice dropped from 15 minutes to under 1 for those that pass automatically (most of them). The ones needing review, 3 minutes. But what had the biggest impact wasn't the time — it was that transcription errors disappeared. When someone copies an amount by hand at 5 PM after processing 40 invoices, they make mistakes. It's human. AI doesn't get tired at 5 PM.
The team recovered hours per day. Hours that used to go into mechanical work and are now spent analyzing spending trends, negotiating with vendors, or simply not leaving the office at 8 PM.
What you need to start
Power Automate Premium license. If your company already has Microsoft 365, you probably just need to activate it. No Azure, no servers, no ML team. The pre-trained model works from day one. If your invoices are very peculiar, you can train a custom model with 5-10 examples — but start with the pre-trained one, it's enough for most cases.
This goes beyond invoices
The interesting thing about this pattern — document comes in, AI extracts data, flow validates and records — is that it applies to almost any business document. Delivery notes, purchase orders, quality forms, work orders, contracts. Once you have the invoice flow working, adapting the same pattern to another document type takes days. Not months, not another project.