How AI bookkeeping actually works
Categorisation, reconciliation, reporting — under the hood of an AI bookkeeper, and why it's more reliable than a rules-based one.
“AI bookkeeping” sounds like magic and is often sold like it. The reality is more reassuring: it's a pipeline of well-understood steps, each made smarter by models that learn from context.
Step 1 — Categorisation
Every transaction needs a category. Rules-based systems match on keywords and break the moment a supplier name changes. A model looks at the whole picture — amount, counterparty, history, your past corrections — and assigns the right category, improving as it learns your business.
Step 2 — Reconciliation
Reconciliation matches money that moved against the records that explain it: invoices, bills, receipts. This is fuzzy matching at its core — amounts that nearly agree, dates that are close, references that are messy — exactly the kind of judgement models handle well.
Step 3 — Reporting
Once the data is clean and current, reporting is the easy part: cash position, income and expenses, VAT due, runway. Because it's live, you can ask questions in plain English and get answers grounded in your actual numbers.
Why the human still matters
Good AI bookkeeping doesn't remove the human — it removes the drudgery. You review, approve and handle the genuine exceptions, while the system does the repetitive 95%. That's the right division of labour between people and models.