AI in finance operations: invoice processing that actually works
Easy Insight Team ·
AI invoice processing works for a UK SME when it does one job: reading supplier invoices into draft bills that a person approves. Extraction is cheap; the hard parts are exceptions, HMRC evidence and fraud. Start with your bookkeeping software's own capture, measure the exception rate, and never let AI approve payments or change bank details.
This is not tax advice. It is the setup we would want in place before AI touched a purchase ledger, built from HMRC and Microsoft documentation as of October 2026. If your VAT position is unusual, ask your accountant before changing how you keep records.
Why start with invoices?
Supplier invoices are the textbook first automation. They arrive constantly, the fields are the same every time, and the process is usually written down already, even if only in someone's head. In our guide to business process automation for UK SMEs, invoice and receipt capture sits at the top of the list for exactly that reason: high volume, clear rules, and a right answer you can check.
That last point matters for AI. If you are choosing your first AI project, our guide to implementing AI in a small business explains why a task with a checkable answer beats an ambitious one.
What does AI invoice processing actually do?
It covers a chain of steps, and AI is good at some of them and should be kept away from others.
| Step | What happens | Who should do it |
|---|---|---|
| Capture | Invoices arrive by email, upload or phone photo | Automated: a dedicated inbox forwarding into the tool |
| Extract | Supplier, date, invoice number, net, VAT, total and line items are read off the page | AI, with low-confidence fields flagged |
| Check | Duplicate check, VAT arithmetic, supplier matched to an existing record | Rules first; AI only to suggest a match |
| Code | Nominal code and cost centre suggested from the supplier's history | AI suggests, a person accepts |
| Approve | Someone confirms the goods or services were received | A person, always |
| Pay | The bill is paid to the bank details on file | A person, from details verified outside the invoice |
The extraction step is where AI earns its keep. Microsoft's invoice model documentation (updated August 2026) describes a prebuilt model that reads phone photos, scans and digital PDFs, extracts key fields and line items, returns structured data, and supports invoices in 27 languages. Many cloud bookkeeping tools offer their own bill or receipt capture built on the same idea; check what yours includes.
How much does the extraction cost?
Less than people expect. As of October 2026, the Azure Retail Prices API lists Microsoft's prebuilt Document Intelligence models at £7.5475 per 1,000 pages pay-as-you-go in UK South. A business receiving 300 two-page invoices a month sends 600 pages, which comes to roughly £4.53 a month for the model itself.
So the model is not the budget line. The costs that matter are:
- The wrapper. Most SMEs will use extraction through their bookkeeping software or a capture add-on, priced per user or per document. Compare those on your own volume.
- Integration. If invoices need matching to purchase orders in a separate system, someone has to build and maintain that link.
- Exceptions. Every invoice the tool cannot read confidently lands back with a person. As an illustration, ten minutes on each of 30 exceptions is five hours a month.
Our view: if you use Xero, QuickBooks, Sage or similar, try the capture feature you may already be paying for before buying or building anything. A bespoke pipeline earns its place only when you have purchase-order matching, several entities or an ERP the off-the-shelf tools cannot reach. Our piece on build vs buy for SMEs sets out that decision in general.
What does HMRC need you to keep?
This is the part most vendor blogs skip, and it decides how your pipeline must store things.
A VAT invoice you receive is your evidence for reclaiming input tax. HMRC's VAT guide (Notice 700), section 16.3, lists what a full VAT invoice must show, including a sequential invoice number, the time of supply, the supplier's name, address and VAT registration number, the customer's name and address, a description, and the VAT rate and amounts. Your check step should flag an invoice missing the supplier's VAT number before anyone reclaims VAT on it.
On storage, HMRC's record-keeping notice (VAT Notice 700/21) gives two examples that map directly onto AI extraction:
- If you only type selected data from an invoice into your software, you must keep the original invoice, because that data is not a copy of it.
- If your software scans the invoice into your ledger and the image is retained and contains all the detail needed for VAT, you do not need to keep the original, unless it is needed for another purpose.
The same notice says records for VAT should generally be kept for at least six years. The practical rule: whatever tool you choose, it must keep the image alongside the extracted data, for that long, and you must be able to find it when HMRC asks.
Where does AI invoice processing go wrong?
Our view: the failures are rarely about reading the text. They are about what happens next.
Duplicates. The same invoice arrives as an email attachment and then as a posted copy, or a supplier sends a reminder with the original attached. Check supplier plus invoice number plus amount before anything reaches the approval queue.
Credit notes and statements. A model trained on invoices will happily read a credit note as an invoice, or a statement as a very large one. Classify the document type first.
Fraud. This is the serious one. The National Business Crime Centre describes mandate fraud as a fraudster posing as a supplier and claiming to have changed their bank details, and advises verifying payments and supplier details directly with the company on a known phone number or in person first. An automated pipeline that updates bank details from whatever is printed on an invoice does the fraudster's work for them. Bank details should only ever change through a separate, verified process, and that is a design rule, not a setting.
Personal data. Invoices from sole traders carry a named individual's details, so a cloud extraction service is processing personal data. Check the vendor's terms and data location as you would for any AI tool; our GDPR and AI checklist covers the questions.
How do you know it is working?
Measure before you switch anything on. Time a fortnight of the current process: minutes per invoice, invoices handled, and how many needed chasing or correcting. Then track three numbers after go-live:
- Exception rate: the share of invoices a person had to correct or complete.
- Minutes per invoice, measured the same way as before.
- Errors that reached the ledger: duplicates paid, wrong VAT, wrong supplier.
If the exception rate stays high after a month, the problem is usually the inputs (poor photos, unusual supplier layouts) rather than the model, and fixing the inputs is cheaper than switching tools.
What changes when e-invoicing becomes mandatory?
The ground is moving. The government's consultation response on e-invoicing (updated 26 November 2025) says the UK will introduce mandatory e-invoicing for all VAT invoices from 2029, covering business-to-business and business-to-government invoices but not sales to consumers. It is explicit that PDFs, Word documents and images are not e-invoices for this purpose. The government's tax update of June 2026 adds that Peppol will be the core interoperability network, and an implementation roadmap is due at Budget 2026.
When that lands, VAT invoices will arrive as structured data and there will be nothing to extract. Our view on what that means now:
- Extraction remains useful until 2029, and afterwards for receipts, non-VAT documents and overseas suppliers outside the mandate.
- Do not sign a multi-year contract for a tool that only reads PDFs. Ask any vendor what their Peppol plans are.
- The steps after extraction (checks, matching, approval, fraud controls) carry over unchanged. Effort spent there is not wasted.
Where to start
Pick the bookkeeping tool you already have, turn on its capture feature for one supplier category, keep the images, and measure the exception rate for a month. If you then need matching across systems or a pipeline built around your ERP, that is the kind of work our AI automation practice scopes, and the AI practice overview sets out how we approach the rest.
Sources (checked 6 October 2026): Microsoft Learn, Document Intelligence invoice model (updated 15 August 2026); Azure Retail Prices API, Azure Document Intelligence, UK South, GBP; HMRC VAT guide (Notice 700, updated 25 June 2026); HMRC Record keeping (VAT Notice 700/21); National Business Crime Centre, mandate fraud; GOV.UK e-invoicing consultation response (updated 26 November 2025); GOV.UK Tax update 2026 summary (updated 23 June 2026).
Frequently asked questions
How accurate is AI invoice extraction?
Accurate enough on clean, typed invoices to be worth using, but the figure that matters is your own exception rate on your own suppliers' invoices, not a vendor's headline. Run a month of real invoices through any tool before committing, and route every low-confidence field to a person rather than trusting it.
Can we throw away paper invoices once they are scanned?
For VAT, often yes. HMRC's record-keeping notice says that if software scans an invoice into your ledger and the image is retained and contains all the detail required for VAT purposes, you do not need to keep the original, unless it is needed for another purpose. Typed-in data alone is not a copy, so keep the image.
How much does AI invoice extraction cost?
The extraction itself is cheap. As of October 2026, Microsoft's prebuilt invoice model costs £7.5475 per 1,000 pages pay-as-you-go in UK South, so 300 two-page invoices a month is about £4.53. The real costs are the integration, the review time on exceptions and the subscription of whatever tool wraps the model.
Does mandatory e-invoicing in 2029 make AI extraction pointless?
Not yet. The government says all VAT invoices will have to be e-invoices from 2029, with Peppol as the core network, and PDFs will not count. Until then, and for receipts and non-VAT documents after it, extraction still does the work. Avoid long contracts for a tool that only reads PDFs.
Should AI approve or pay invoices on its own?
No. Let it read, match and draft. Approval and payment stay with a person, and a change of supplier bank details should never be taken from an invoice at all: the National Business Crime Centre advises verifying supplier details on a known phone number or in person first.
Easy Insight is a UK consultancy for AI, web, apps and data — senior specialists only, no juniors.
Next step
Wondering where AI would actually pay in your business?
An AI strategy and readiness review (from £2,500) tells you where AI pays, where it doesn't, and what to leave alone. The price is fixed in writing before we start, and the advice comes from a team that runs AI in its own products.