The liability question deserves a real answer, not reassurance
Most content aimed at accountants about AI tools either dismisses the liability question, treating it as an edge case not worth dwelling on, or catastrophises it, implying that any AI use exposes a practice to unmanageable risk. Neither response is honest, and neither helps a practice actually think through where their exposure sits.
The useful question is narrower and more practical: when AI produces an output that feeds into a client's accounts or tax position, where exactly does that output need a human check, where does professional judgement remain squarely with the accountant, and what does that mean for how a practice frames its liability, both to itself and to its professional indemnity insurer.
Where AI outputs generally need human review
AI tools used in bookkeeping today mostly perform two functions: extraction, pulling structured data out of an unstructured document like a receipt or invoice, and categorisation, assigning that data to the correct expense or income category. Both functions carry different error profiles, and both need different levels of human oversight.
[Extraction accuracy](https://receiptflow.co/blog/receiptflow-vs-datamolino-uk-accountants) depends heavily on document quality. A clean, standard-format invoice extracts reliably in almost all cases. A crumpled receipt, an unusual layout, or a handwritten amount introduces genuine ambiguity, and a well-built extraction tool will flag low-confidence fields for review rather than silently guessing. The point of human review here is narrow and specific: confirming figures the system itself is uncertain about, not re-checking every clean extraction from scratch.
Categorisation judgement is where more genuine ambiguity lives. Mixed-use expenses, entertainment versus staff welfare, and borderline capital versus revenue classifications all require the kind of contextual judgement that depends on knowing the client's specific circumstances, not just the receipt in isolation. This is squarely a human decision, and no AI tool should be treated as making that call unsupervised, the same trust-but-verify principle we cover in why AI receipt fraud is every accountant's problem now.
Where errors typically occur in practice
The error pattern that shows up most often is not the AI misreading a clear figure. It is a low-confidence extraction being accepted without the review it was flagged for, because the review step was skipped under time pressure. The second most common pattern is a categorisation suggestion being accepted by default rather than actively confirmed, particularly for expense types with genuine ambiguity like mixed business and private use, the same default-acceptance habit that shows up in expense claim fraud at UK small businesses.
Both patterns share a root cause with the review gaps covered in AI receipt fraud and what UK accountants need to know: the human review step exists in the workflow but gets treated as optional in practice, especially during busy periods. This is worth naming plainly, because it reframes the liability question. The risk is rarely that the AI tool is unreliable. The risk is that the review step built around it gets skipped.
Framing your professional indemnity exposure honestly
Using AI tools in a bookkeeping workflow does not, on its own, change where professional liability sits. The accountant or practice remains responsible for the accuracy of the accounts and returns they submit, regardless of which tools were used to produce the underlying figures. This is consistent with the general direction of professional body guidance on AI, which treats AI as a tool subject to the same professional standards and oversight as any other, rather than a basis for reduced accountability.
For professional indemnity purposes, the practical question an insurer or regulator is likely to ask after an error is not "did you use AI" but "what review process did you have in place, and was it followed". A practice that can demonstrate a clear, documented review step for low-confidence extractions and ambiguous categorisations is in a materially stronger position than one that cannot, independent of which specific software was involved.
A few concrete steps make that position demonstrable rather than assumed:
- Document the review process, not just that one exists, but what specifically gets checked and by whom.
- Keep a record of flagged low-confidence items and their resolution, so there is an audit trail showing genuine review happened, not just automated processing, the same kind of trail our year-end receipt checklist asks practices to build anyway.
- Set explicit rules for categorisation ambiguity, so mixed-use and borderline expense decisions follow a consistent, documented standard rather than depending on whoever processed that particular receipt.
- Review AI tool vendor documentation on confidence scoring and flagging, since not every tool surfaces uncertainty the same way (see how this plays out in practice in our look at a Dext alternative for receipt scanning), and understanding exactly what a vendor's output does and does not guarantee is part of due diligence.
Where Receiptflow sits on this
Receiptflow's extraction is built to be transparent about confidence, not to present every field with false certainty. Low-confidence extractions are flagged for review rather than silently accepted, and the review step is a visible part of the workflow rather than something a practice has to build separately. This does not remove the need for professional judgement on categorisation ambiguity, and it should not be treated as though it does. What it does is make the human review step something a practice can actually point to and demonstrate, rather than something assumed to be happening informally, including for practices running that review alongside an integration with their accounting software.
See how Receiptflow keeps every extracted record reviewable and start a free trial.
The honest position on AI and liability is neither alarmist nor dismissive. AI tools change where manual effort goes, not where professional accountability sits, and the practices best positioned are the ones that can show, clearly and specifically, what their review process actually catches.
The same question comes up from the other direction with AI-generated receipts themselves: if a convincing fake gets through unchallenged, where the liability actually sits is worth knowing before it comes up with a client.
