Automating invoice processing without breaking the audit trail

The problem

Someone opens a PDF, reads a supplier name, an invoice number, a date, a net figure, a VAT figure and a total, then types all six into a finance system. They do this a few hundred times a month. It is slow, it is boring, and the errors it produces are the expensive kind — a transposed figure in a VAT box is not caught by anybody until it is a problem. Meanwhile the invoices themselves arrive by email, by post, as photographs from site, and occasionally embedded in the body of a message.

What automation can and cannot do reliably

Modern document AI is genuinely good at pulling structured fields out of unstructured documents, including scans and photographs. It is not good at knowing whether an invoice should be paid.

Draw the line there and these projects go well. Cross it — approve automatically because extraction confidence was high — and you have automated the wrong step.

• Reliable: supplier, invoice number, dates, line totals, net, VAT, gross, currency, purchase order references where present
• Reliable with tuning: line-item breakdowns, nominal coding based on supplier history, duplicate detection
• Not reliable: deciding whether the goods arrived, whether the price is what was agreed, or whether this supplier should be paid at all

The audit trail is the design constraint

For anything touching finance, the question is not just "did it work" but "can you show what happened". Every automated step should leave the original document, the extracted values, a confidence indicator, and a record of who approved it and when. If a figure was corrected by a human, that correction should be visible rather than silently overwriting the machine output.

This matters twice over for accountancy practices processing client documents, where the trail may need to satisfy a client, HMRC and a professional body — and where the practice's own PI position depends on being able to reconstruct a decision.

Design the exception path first

The happy path is easy. The value evaporates in the exceptions: the invoice that is a credit note, the supplier who changed their layout, the scan that is upside down, the duplicate submitted twice under different filenames.

Decide up front where low-confidence extractions go and who owns that queue. A common and workable rule is that anything below a confidence threshold, anything over a value threshold, and anything from a new supplier goes to a person regardless. The rest posts as a draft for approval. Over time the thresholds move as you see what the system gets wrong — but they move deliberately, not by drift.

A realistic implementation sequence

We would generally run this over a small number of weeks rather than as a big programme, because the useful information comes from real invoices rather than from planning.

The technical build is rarely the long pole. Agreeing the coding rules with whoever owns the ledger usually is.

• Collect a sample of real invoices, including the awkward ones, and measure current handling time honestly
• Run extraction over the sample and score field-level accuracy — this sets expectations before anyone commits
• Build the intake route: a monitored mailbox, a folder, or a scan destination
• Post drafts into the finance system rather than final entries, with the source document attached
• Run parallel for a month, tracking corrections to find the systematic errors
• Tighten the rules, then automate the coding for suppliers where accuracy has proven out

How to tell whether it paid for itself

Measure three things before you start, or you will not be able to answer this later: minutes per invoice, error rate found downstream, and days between receipt and posting. Volume matters too — automation on eighty invoices a month is hard to justify; on eight hundred it usually justifies itself in a quarter.

The benefit that surprises people is the third one. Getting invoices posted within a day rather than a fortnight changes the quality of every cash flow conversation the business has, and that tends to matter more to a finance director than the labour saving.

The build sits under AI document processing, usually alongside a small amount of integration work to get data into the ledger you already use.