AI Builder: Read Your Documents Automatically

Three hundred supplier invoices a month, keyed by hand into the accounts. Delivery notes photographed by drivers and retyped in the evening. Scanned paper forms whose boxes someone copies out. Identity documents to check at every hire. In a small business, reading documents eats hours that create nothing. AI Builder, the artificial intelligence layer of Power Platform, reads those documents and pulls out the data, inside your Microsoft 365. Here is what it can do, what it costs, and when it is not worth it.

What AI Builder document processing does

AI Builder offers prebuilt models and models you train on your own documents. Among the prebuilt ones, the most useful for a small business:

• Invoice processing. Supplier, number, date, amounts, VAT, line items: extracted from a PDF or a scan, with no configuration.
• Receipts and expenses. The photographed restaurant receipt becomes a line with date, merchant, amount.
• Identity documents. Reading passports and some identity documents; test on your real documents before relying on it.
• Text recognition. Basic OCR: turning an image into text, including legible handwriting.

A prebuilt model is wired up in an hour. It needs no examples, and it improves with Microsoft's updates without you doing anything.

Prebuilt model or a model trained on your documents

As soon as the document is your own (an in-house delivery note, a service report, a registration form, a quality control sheet), we move to a custom document processing model. You supply a few examples of each layout (count five as a minimum per type, around twenty to be comfortable), you mark on screen where each field sits, and AI Builder learns. It handles fixed-box forms as well as freer documents, and tables.

That training work takes half a day to a day per document type, and it is often your team that does it with us, because they are the ones who know how to read those documents. A supplier turns up with a new layout: we add its examples, retrain, done.

The confidence threshold and human control

This is the setting that separates a useful project from a scary one. For every extracted field, AI Builder returns a confidence score. We set a threshold, say 85%. Above it, the data goes straight into the flow. Below it, the document goes to a verification queue: a Power Apps screen where a person sees the scan on the left, the fields on the right, corrects and approves in thirty seconds.

Two rules we do not negotiate. An amount that triggers a payment or an accounting entry always goes through human validation, whatever the score. And we keep the original document next to the extracted data, so it can be checked later.

Fitting it into a Power Automate flow

AI Builder never works alone; it is one step in a Power Automate flow. The classic pattern for invoices: an email lands in the invoices mailbox, the attachment is saved to SharePoint, the AI Builder action analyses it, the flow creates the row in your tracking table (or in your accounting software if it has a connector), renames the file with the supplier and the number, and notifies the person who has to approve. Our invoice tracking template shows that chain end to end.

For delivery notes, the trigger is a photo dropped from a phone into a SharePoint library. For scanned forms, a watched folder on the copier. The principle is the same: a trigger, a read, a check, a write. If you are new to flows, our five Power Automate automations give you the basic reflexes.

Credits and licences: what to know beforehand

AI Builder is paid for in Copilot Credits, consumed for each page analysed: one invoice page costs 8 credits, a few cents, to be checked against the current Microsoft price list. Those credits are bought as a prepaid pack or paid as you go. The credits included in some Power Apps and Power Automate licences disappear on 1 November 2026: do not build your budget on them. At three hundred invoices a month, the cost runs to tens of euros; do the maths with your real volume of pages, a three-page invoice counting as three. Our article on what is included in Microsoft 365 sets out the framework.

When plain OCR is enough, or nothing at all

AI Builder is not always the right answer, and we would rather say so up front.

• The PDF is already text. An invoice generated by software carries its data in plain text; a flow that reads the text is enough, no AI needed.
• The invoice arrives as an electronic invoice. With business-to-business e-invoicing spreading, more and more invoices arrive with structured data (Factur-X, for instance). There is nothing left to read, only to connect.
• The form is always identical and clean. Zone-based OCR does the job at lower cost.
• The volume is low. Twenty documents a month, one person keys them in an hour. The automation would cost more than it saves.

A flow like this starts with a short scoping workshop, which checks that your documents are suited to it and estimates the Copilot Credits on your real volume, then is built for a fixed fee: the amount is agreed once scoping is done.