← Back to DocProcPicture a finance inbox on a Monday morning: invoices, receipts, bank statements, all PDFs, all waiting for someone to read them, retype them and check the numbers. We taught Snowflake to do that part itself. The platform that already holds the data now reads the documents too, so nothing sensitive ever leaves it.
The part we didn't expect: the biggest improvement came from removing a step. Each document now costs 7 cents instead of 90 and after we trained the system on our own documents, it became more accurate, not less.
How it works
A document arrives. The AI first identifies what it is (an invoice, a receipt or something else) and renames the file accordingly. Then it reads it and pulls out the details: supplier, amounts, dates, IBAN. It also tells us how sure it is. If it's confident, the document goes straight to the books. If there's any doubt, a person takes a look, with everything already pre-filled.
For the auto-approved documents the process runs end to end with no human touch: no reading, no retyping, no checking. A document that used to wait in a queue for someone's attention is in the books about a minute after it arrives. We also keep the confidence bar deliberately high while we test; that automatic share will climb.
From inbox to audit-ready
DocProc sits in the middle of the tools you already use. Documents flow in from email and SharePoint. The extracted values and the document go straight into your ERP, and the file itself is stored back in SharePoint in a clean, structured folder system, so every document is findable and your records are audit-ready.
Here's what that looks like in SharePoint. Every file follows the same naming pattern (date, supplier, reference, type) and carries the extracted values as metadata columns, so you can sort, filter and retrieve any document in seconds:
Behind these columns sit further metadata fields written by DocProc: the ERP record it was booked to and a unique document key, the full trail an auditor needs. Suppliers shown are fictitious examples.
What it means in people's time
Plug in your own numbers. The defaults reflect a typical Swiss mid-size finance team and a fully loaded personnel cost of CHF 65 per hour.
Every assumption above is editable, so the result reflects your own operation. We are aiming for 92% fully touchless, all the way to payment in e-banking. Net savings are after DocProc running costs.
What it costs
Our first version worked. But most of the money went into one preparation step: converting each document to text before the AI even looked at it. Then newer Snowflake AI learned to read the original file directly. So we deleted the step. Same input, same output, 92% less money.
How accurate is it?
We trained the AI on documents our own reviewers had corrected, then tested it on ones it had never seen. Field by field:
And remember the safety net: anything the AI isn't sure about goes to a person anyway. These percentages decide how much work is left for people, not whether errors reach your books.
Why it matters
Everything, AI included, runs inside Snowflake. No documents sent to outside services, no new vendor to assess.
Every run records exactly what it cost, per document. Budgeting is a query, not an estimate.
Your team stops retyping documents and only checks the ones that genuinely need judgment.
Ready to transform data into a force for your business?
We're a Snowflake Select Partner delivering data-management and financial-risk solutions across Switzerland. Let's talk about what a native, evaluation-driven pipeline could do for your team.