# Can AI Process Your Invoices? What to Check First

Source: https://cplt.tech/blog/can-ai-process-invoices-small-business
Author: Stéphane Lepain
Published: 2026-11-18T07:12:12+00:00
Tags: invoices, small-business, human-review

---

Yes, AI can read an invoice and fill in the supplier, invoice number, dates, totals and line items. Microsoft, Amazon and Google each document an invoice service for it. The reading is the easy part to buy. What stays with you is the checking: whether the numbers add up, whether the invoice was already paid, whether it matches an order, and who approves the payment.

## What the invoice services read

Microsoft's [AI Builder invoice model](https://learn.microsoft.com/en-us/ai-builder/prebuilt-invoice-processing) returns the purchase order reference, invoice ID, dates, subtotal, tax, total, vendor tax ID, bank details such as IBAN and SWIFT, and each line item with its quantity and unit price. Its [Document Intelligence invoice model](https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/prebuilt/invoice) extracts "key fields and line items from sales invoices, utility bills, and purchase orders". Google's [Invoice Parser](https://cloud.google.com/document-ai/docs/processors-list) "extracts both header and line item fields". Amazon's [Textract](https://docs.aws.amazon.com/textract/latest/dg/invoices-receipts.html) extracts "invoice or receipt date, invoice or receipt number, item prices, total amount, and payment terms".

The input matters. The AI Builder page asks for "one clear photo or scan per invoice" as a JPEG, PNG or PDF, up to 20 MB. A crumpled phone photo of three invoices on a desk is a poor start for any of them.

## What a confidence score tells you

Microsoft's and Amazon's services can return a confidence score with a field. Microsoft's [accuracy and confidence page](https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/concept/accuracy-confidence) describes it as an estimated probability between 0 and 1: a score of 0.95 "indicates that the prediction is likely correct 19 out of 20 times". It adds that confidence can be used "to determine whether to automatically accept the prediction or flag it for human review", and that "not all document fields return a confidence score".

Two cautions come with that. The scores of different services use different scales: Microsoft's run from 0 to 1, while Amazon's sample output shows a receipt total with a confidence of 97.1. Set any threshold from your own results, never from a number copied from another page.

The second caution is arithmetic. Take ten fields that are each right 95% of the time. If they were independent, which real invoices are not, an invoice with all ten right would turn up about 60% of the time (0.95 to the power of 10 is about 0.6). This is an illustration, not a measured result. It shows why I look at whole invoices and at the wrong fields that matter most, such as the total and the bank details.

An accuracy figure on an AI provider's page describes the invoices that provider tested. Your own suppliers' invoices are the test that counts.

## Five checks before anything is paid

This is my own checklist, built from how the extraction fields are laid out. It is not a validated standard.

1. Do the numbers add up? Recompute the lines, the subtotal, the tax and the total with a short script, not with the AI model. A wrong total then stops the invoice before a person reads it.
2. Is it a duplicate? Look for the same supplier, invoice number and total in your records.
3. Does it match an order? Several of these services read a purchase order reference. Compare it with what you ordered, and send invoices without a match to a person.
4. Are the bank details the ones you hold for this supplier? The services can read bank details such as the IBAN or the account number off the page. If they differ from your supplier record, a person phones the supplier on a number you already hold before anyone changes anything.
5. Who approves the payment? Keep approval separate from reading. At the start, a named person sees every invoice. Later, the exceptions go to that person: a low score, no order match, new bank details, or an amount above a limit you set.

## Test on your own invoices first

Take a batch of past invoices where you already know the right answer. Include the awkward ones: scans, phone photos, credit notes and each supplier layout you see often. Compare the result field by field, count the misses per field and set your review rules from that count. Microsoft's own advice for low read confidence is to improve the quality of the input documents, so the scan quality is part of the test.

That is the same method as the [cost per accepted result](https://cplt.tech/blog/ai-pilot-cost-per-accepted-result): count the time a person spends checking, not only the AI's price. When a result is wrong, trace it as described in [what to do when your AI gives a wrong answer](https://cplt.tech/blog/what-to-do-when-ai-gives-a-wrong-answer).

## Where the invoice files go

An invoice carries supplier bank details and sometimes personal names. With the services above, the provider you choose processes the file. Read that provider's data-handling terms for the service and decide whether this type of document may go there. If it may not, the reading step can run on your own systems; [use AI on your documents without a third party](https://cplt.tech/blog/sovereign-ocr-document-ai) explains which steps to check.

## Fixed steps, with a person at the end

Invoice handling fits the sequence from [AI agent vs chatbot](https://cplt.tech/blog/ai-agent-vs-chatbot-small-business): the invoice arrives, the AI reads it, your checks run, the entry is drafted and a person approves it. I would build it as fixed steps. The AI reads and drafts, and the order of the work stays under your control.

If invoices reach your inbox and get retyped into your accounting software, look at how I approach [business process automation](https://cplt.tech/services/business-process-automation). Then bring ten real invoices to a [free 45-minute call](https://cplt.tech/contact). I will show you which fields can be read and checked automatically and which stay with a person.

## Frequently asked questions

### Can AI be used for invoice processing?

Yes. Microsoft, Amazon and Google each document an invoice service that reads fields such as the supplier, invoice number, dates, totals and line items. Plan the checks and the approval step as carefully as the reading.

### How do you check that invoice data is accurate?

I would recompute the totals in code, look for a duplicate, match the invoice to an order, compare the bank details with the ones you hold and name the person who approves payment. Then test the whole sequence on your own past invoices.

### Should AI approve payments on its own?

I would begin with AI preparing the entry and a person approving the payment. Once tests on your own invoices show a type of invoice is read and checked correctly, you can decide whether that type may go through without a person.
