AI in finance: what it can do for a finance team, and what it can't

Much of what is written about AI in finance is about banks: trading, fraud detection, credit scoring. This guide is about the finance function inside an ordinary business, where a finance director or financial controller wants to know which jobs AI can take on, and what that does to approvals, records and the audit.

The short version: AI is good at reading messy documents and writing first drafts, and poor at being the final word on money. The pattern that works is that AI prepares, rules check, a named person approves, and every step is recorded. That is how we build AI automation for finance teams, and it is a fair test for any tool you are offered.

Five jobs where AI helps a finance team

Invoice capture and coding

Supplier invoices arrive as PDFs and photos in many layouts, but carry the same facts: supplier, invoice number, date, lines, VAT and total. A language model can read those, draft the bill and suggest a nominal code from how similar invoices were coded before. Accounting packages now do some of this themselves: Xero's bills page says "Xero's AI extracts the key details to create draft bills ready for approval". Note the words draft and approval. Matching invoices to purchase orders and paying them is covered in our guide to accounts payable automation.

Matching and reconciliation

Much bank matching is rules, not AI: the same amount, the same reference, a date within a few days. AI earns its place on the awkward lines: one payment covering three invoices, a mistyped reference, a remittance advice in an email attachment. Xero says its AI, JAX, reconciles automatically only when it's highly confident (on Grow plans and above, or as an add-on), otherwise suggesting a match for a person, and shows which method it used: one of your bank rules, a match to an existing record, how you reconciled similar lines, or how other Xero users did. QuickBooks says it suggests a category for each bank transaction and matches payments to invoices "for your review". Seeing why a line was matched is what lets someone check the work later.

Credit control drafting

Chasing debt is mostly writing: reminders, call notes, replies to a disputed line. AI can draft each one from the account's history for a person to check and send. QuickBooks describes its assistant, Intuit Intelligence, drafting invoices in bulk (a feature it labels beta) and sending prioritised reminders "with your approval". A draft has to be right on the facts: amounts, invoice numbers and any claim for statutory interest, which for business-to-business debts is 8% plus the Bank of England base rate unless a contract sets a different rate. Those figures should come from the ledger and a calculation, never from the model. Putting an account on stop or passing it to a solicitor stays a person's decision.

Forecasting support

A short-term cash forecast is mostly arithmetic on what you already hold: invoices due in, bills due out, payroll and VAT. Xero's cash flow forecast runs from 30 days on its Ignite plan to 180 days on Ultimate, showing how upcoming quotes, invoices and bills will affect your bank balance. AI helps around the forecast rather than inside it: pulling payment terms out of contracts, explaining why last month missed the plan, drafting board-pack commentary. Keep the forecast itself where a person can follow every formula.

Month-end

AI can compare accruals and prepayments with last month's and list what looks missing, flag journals unusual in size, timing or account, chase outstanding items, and draft variance commentary for a person to edit. It should not post journals on its own. The auditing standard ISA (UK) 315 lists controls over journal entries among those expected to be identified in every audit (paragraph A160), so a journal an AI prepares should still have an approver your auditors can see.

Where rules do better than AI

Language models have a failure that NIST's Generative AI Profile calls confabulation: "confidently stated but erroneous or false content". A wrong supplier name gets spotted; a wrong VAT amount in a tidy draft may not. Our rule: use AI where the input is messy and the output can be checked, and plain code wherever there is a right answer that can be calculated.

Our own self-storage software, StoreBay, shows the second half of that rule. Its billing engine decides what is owed and when, with amounts kept in whole pence and VAT rounded line by line. The card and Direct Debit providers only move the money. A Direct Debit counts as paid when the provider confirms it, and if it comes back later the payment is reversed and the arrears reopen. Reminders, a late-fee notice and a final notice follow a schedule the business sets, and declaring a default is always a person's decision, never a timer's. Those are rules written as code, because a billing run must give the same answer every time.

What HMRC expects

VAT Notice 700/22 requires VAT-registered businesses, unless exempt, to keep records digitally and file VAT returns using software. Section 3.2.1 matters for AI. Typing an invoice into your software is allowed, but once data is in the software that keeps your VAT records, any further transfer or change must go by a digital link, such as an API or a file import; HMRC does not count cut and paste. So an AI tool that reads invoices before they reach the ledger is outside that rule, though writing to the ledger through its API saves the retyping. One that works on data already in your records, such as recoding transactions, must send its changes back by a digital link, or be a feature of the package itself.

Then keep the evidence. VAT records must be kept for at least 6 years, including the invoices you receive, as originals or electronic copies. VAT Notice 700/22 adds that data typed from an invoice is not a copy of it: keep the invoice, or a scanned image carrying all the VAT detail. A limited company must keep its accounting records for 6 years from the end of the last company financial year they relate to, and you can be fined £3,000 by HMRC, or disqualified as a director, for not keeping them. For an AI step, keep the source document beside what was read from it, and who approved it.

One change is coming. The government's e-invoicing roadmap, published on 28 October 2026, says VAT-registered businesses will have to issue and receive VAT invoices as e-invoices from 1 April 2029, unless an exemption applies, with the detailed requirements due between October 2027 and April 2028. PDFs, Word documents and image files do not count. As suppliers move to e-invoices, fewer invoices will need reading by AI at all, so judge an invoice reading project on what it saves before then.

What your auditors will look for

If your accounts are audited, your auditors work to the Financial Reporting Council's ISA (UK) 315, which already describes the kind of control an AI step should sit inside. Its appendix on internal control gives an example of an automated approval: an invoice's unit cost compared with the purchase order's within a set tolerance, approved if inside it and flagged for investigation if not. It adds that the more a business relies on automated controls, the more important general IT controls may become: who can access the system, and how changes to it are made.

In practice, every AI step in a finance process should have:

• A named approver. One person or role signs off each output that reaches the ledger, and the system records who and when.
• Separate duties. Whoever can change the AI step's instructions should not approve its output, and nobody should be able both to change a supplier's bank details and to approve a payment to them.
• A record of what went in and what came out. The source document, the data read from it, the model used, and anything the approver changed.
• Changes treated as changes. A new prompt or model version is tested against real examples before it goes live, and recorded.

Regulated firms and personal data

For firms it regulates, the Financial Conduct Authority says it does not plan to introduce extra regulations for AI; it relies instead on existing frameworks, including the Consumer Duty and the Senior Managers and Certification Regime.

Invoices from sole traders and debts owed by individuals involve personal data, so data protection law applies to any AI that reads them. Start with the ICO's guidance on AI and data protection, which the ICO says is under review after the Data (Use and Access) Act. Ask any supplier which AI providers your data goes to, and on what terms.

How to start

1. Map one process first. Pick a job that runs often, such as supplier invoices or the debtors chase, and map it as it really runs, exceptions included.
2. Try what you already pay for. If your accounting package already captures bills or suggests matches well enough, use it: a feature inside the ledger has no hand-off between systems to get wrong.
3. Test on your own documents. Take a few dozen real examples, awkward ones included, and check every field the AI returns.
4. Build only for the gap. If the work spans systems the package can't reach, such as a shared inbox, a purchasing system and the ledger, that is where a custom build pays. Our guide to intelligent document processing covers the reading part.

If you want a hand

We build AI steps like these into the systems finance teams already run, with tests on real documents and a person approving anything that reaches the ledger. If you have a job in mind, start a project and tell us about it.