In June 2026 Grant Thornton published a piece on how AI in financial reporting redefines audit risk. Its useful move is to point the question at the client rather than the firm.
The Four, In Plain Terms
When a company uses AI to produce what it reports, the firm reviewing that output is relying on something new. Four risks are named, and they are worth translating out of audit language, because they apply to a compilation and a bookkeeping cleanup as much as to an audit.
Model design risk. Bias, incomplete training data and embedded assumptions can introduce blind spots that are not immediately visible. In practice: the tool was configured once, by somebody, with assumptions nobody wrote down, and those assumptions are now in every month.
Data quality risk. Risk moves from manual execution to system design and data integrity. A person miscoding an invoice makes one error. A rule miscoding a vendor makes the same error every time it sees that vendor, and it will look consistent, which is the problem.
Interpretability risk. The logic behind a conclusion may not be fully transparent. Ask why an item landed where it did and the honest answer may be that nobody can reconstruct it.
Over-reliance risk. Automated output creates a false sense of precision, particularly when results appear consistent. Consistency feels like accuracy. It is not the same thing, and a system can be reliably wrong far more easily than a person can.
Why This Reaches Smaller Firms First
It would be easy to read that as a problem for auditors of large companies. The opposite is closer to true.
A large company using AI in financial reporting has a controller, an internal audit function and a governance framework that at least names the tool. A small business owner who turned on automated categorisation in their bookkeeping software has none of those, does not describe what they did as deploying AI, and will tell you the books are done.
The firm that then produces the accounts is relying on a system nobody has evaluated, configured by nobody in particular, on assumptions nobody recorded. That is a materially different position from relying on a bookkeeper whose work you know.
The Question That Has to Enter the Conversation
The profession already asks clients how the records were kept. That question needs one more clause.
Which parts of this were produced by software making its own decisions, and has anyone checked a sample of those decisions against the source documents. Not to catch the client out, and not to argue against them using the tools, but because the answer changes what a firm can reasonably take as given.
Grant Thornton's guidance to auditors points the same way: focus on independently validating system outputs rather than accepting automated conclusions, and understand how responsibility and accountability sit within the new arrangement. That instruction survives translation to every other service line. Somebody has to own the output, and the software does not.
Where the Work Actually Goes
The practical consequence is that verification moves earlier and gets more specific.
Instead of reviewing the trial balance, somebody traces a sample of automated postings back to the invoice, the contract or the statement that should support them. That is not new work in kind. It is older work than the trial balance. What makes it feel new is that it now has to be done on a population that looks tidy.
MetaWurks is built for exactly that trace. It ingests a client's invoices, contracts, statements and correspondence and lets an accountant query the whole set in plain English, so checking what actually supports a posting is a question rather than a search. Documents ingested into the platform are not used to train models and are not exposed to other users, role based access controls decide who can open which client's records, and audit logs record who opened what and when.
The tidiness of a client's books used to be weak evidence that somebody careful had been through them. That inference has stopped holding, and the firms that notice first will price the work correctly.
Join the Conversation
For your three largest bookkeeping or compilation clients, do you know which parts of their records were categorised by software rather than by a person?