As of 11 August 2026, AI can only partly categorise a year of business transactions.
Most people should hand this to a purpose-built tool.
Can you do it?
30 minutesto a draft.
4 hoursto something you’d act on.
Cost, all in£20/month
Skill neededpower-user
Who has to check ita colleague
What the alternative costsA bookkeeper charges roughly £45/hour; a year of backlog is typically 6–10 hours (checked 2026-08-11).
If this goes wrong: a year of mis-categorised transactions propagates into your accounts and you find out at filing time, when unpicking it costs more than the bookkeeper would have.
What to actually do
Use a tool built for this
The route this page recommends
Purpose-built software does the mechanical share of this — see what to use below.
Hand it to a person
Second choiceA person who owns the outcome does this end to end — worth it when the failure is dear.
Do it yourself
The distant thirdA chat interface, power-user skill, and roughly 4 hours until you can act on the result.
The prompt
Here is a CSV export of 12 months of business bank transactions. Categorise each row against the attached chart of accounts. Flag every row you are less than certain about rather than guessing — a wrong category is worse than an unanswered one.
Run it in ChatGPT, Claude or Gemini (paste it there).
Tested with claude-haiku-4-5-20251001 (claude-haiku-4-5-20251001) on 2026-08-11.See the receipt — the raw output, stored.
What it gets wrong
- Recurring payments to ambiguous payees (a director’s card at a supermarket) get confidently mis-filed — judgement calls need your context.
- Transfers between your own accounts are routinely categorised as income or expense.
- It will not ask about VAT treatment unless told; UK VAT edge cases need a human pass.
Why AI falls short here
What to use
FreeAgent
£19 (Sole Trader, checked 2026-08-11)
Its bank-feed rules engine does the recurring 80% mechanically, which is exactly the part AI does confidently and you still have to check — here it is deterministic instead.
GnuCash
Free
Free and local: if the year of data is sensitive enough that uploading it anywhere feels wrong, this is the credible no-cloud route.
How we scored this
| Axis | Score (0–2) |
|---|---|
| Output | 2 |
| Inputs | 1 |
| Verification | 1 |
| Liability | 1 |
| Effort delta | 2 |
| Total | 7 / 10 |
The rubric, the thresholds and the routing rule are published in full at/method.
Handed to AI by 0 people.
Questions people ask
The four questions readers actually bring to this task are hand-written per ruling and land with the next corpus update.