AI gives accounting firms wrong answers mainly because of AI data quality. Every tool sold into this market was built and demonstrated against a complete client record, and most firms do not have one. The answer that comes back is plausible and built from real figures, so you cannot spot it by looking. You catch it by checking against an authoritative source and being able to rerun the work.
Why does AI give confident wrong answers from real figures?
AI gives confident wrong answers in accounting firms because it works with whatever data it reaches and has no way of knowing that data is partial. The demo worked because the demo data was complete. Inside a real firm, the same tool meets a client record split across systems, and it answers anyway.
Asked what software vendors wrongly assume about the data inside accounting firms, Evan Jones answered in three words. That it's complete. Jones is specifying the data architecture for a private equity backed group of accountancy businesses, and he has seen the gap from both sides. The distance between a complete record and one reliable version of each client is where most wrong answers start.
The two sides do not always mean the same thing by "client" either.
I've seen that where one firm will say a client is whoever we have an engagement with, and someone else says, well, a client is whoever we bill.
Evan Jones
Two definitions produce two sets of stored data, and neither firm has done anything wrong. The tool, though, was built against one idea of what a client is, and most firms hold something else. Point it at a partial dataset and it still produces opportunities worth chasing, along with clients whose profitability looks low against others in the file. Nothing flags the gap. From where the tool sits, nothing went wrong: it answered the question you asked, using the data you had.
Where AI data quality breaks: what the tool assumes, and what your firm has
The gap between what an AI tool expects and what a firm holds is specific, and it can be written down. Jones's account of vendor assumptions breaks into four mismatches, and each one produces an answer that looks reasonable while resting on an incomplete picture of the client.
| What the tool was built to expect | What it finds |
| One client record, and one answer to who this is | Four versions under four names, none of them wrong |
| An agreed definition of a client | Whoever the firm has an engagement with, or whoever it bills |
| A complete service history for each client | Whatever each system needed to do its own job |
| Fees and costs granular enough to compare one job with another | A total on an invoice, split after the fact if at all |
This is where AI data quality stops being an IT phrase. A profitability analysis built on invoice totals split after the fact will rank clients, confidently, on numbers nobody would sign off by hand. A client history assembled from whichever system answered first will miss the work recorded somewhere else. Firms with a single client record across every service line have one less mismatch to manage.
The profession has sensed it. ACCA and Chartered Accountants Australia and New Zealand surveyed 1,600 finance professionals globally in 2026. They found 93% with at least a moderate level of concern about the integrity and verifiability of AI-generated insights, and 67% with a high level of concern.
How do you verify an AI answer in an accounting firm?
You verify an AI answer by checking it against something authoritative, and the hard part is that for many uses nothing authoritative exists. Summarise a contract and the source document is sitting there. Analyse a large volume of data, or run a workflow across several steps, and the only thing left to check against is the process itself.
Alistair Brisbourne, Head of Technology Research at ACCA, puts the profession's standard plainly.
The big part of our profession isn't just getting the right answer. It's having traceable results according to a strict process and methodology.
Alistair Brisbourne, Head of Technology Research, ACCA
That gives a partner a test to apply immediately: whether you, as the person accountable, could rerun that work by hand and reach the same result. Brisbourne's position is that this has to be possible. Mostly it is not, because nobody recorded what went in.
He reads the profession's concern as a gap in governance and a gap in training, more than scepticism. He also notes that concern falls as proficiency rises. Falling concern is not the same as capability keeping pace with confidence.
Adoption is massively outpacing the development of clear guardrails and policies ... upskilling is not just a capability enhancer, it's also a risk mitigator.
Alistair Brisbourne, Head of Technology Research, ACCA
Rerunning work needs a record of the inputs. Where client work sits in Acting Office, the platform keeps a full audit trail of every user action against the client record, which gives that test somewhere to start.
Who is responsible when AI output is wrong?
Responsibility for wrong AI output sits with the accountant who signed the work, whatever tool produced it. The PCRT topical guidance from seven professional bodies says members are "ultimately responsible for any work they produce ... irrespective of the use of AI tools in its creation". The regulator reached a similar place by a different route.
The first of three risk categories in the FRC's March 2026 AI guidance for audit is deficient output, defined as the risk that issues in the system's inputs or design result in a deficient output. The regulator describes a failure that starts in the inputs. The profession puts it on whoever signed the output. Between those two statements sits your client file.
It changes again when a tool stops drafting and starts acting. The Digital Regulation Cooperation Forum, the joint body of the ICO, FCA, Ofcom and CMA, reported in March 2026 that most real deployments still sit at assistant or operator level, under close supervision. The useful question for your firm is narrower. It is which of your tools can already write to a client record, and who granted that.
If it's got the capability to act and update client records and do submissions, and we give it the permission without planning it properly, it's probably going to run off and create a mess for us.
George Llewellyn, Head of Technology, Apteus Group
In most industries AI is a tool. In an accounting firm it behaves more like a member of staff whose work you are signing. The checks belong before the client sees it, in the same way maker-checker approvals already sit on human work. AI data quality is where that work starts. Acting Office built Before you deploy AI on the record with contributors from ACCA, HLB International, Apteus Group and a private equity backed accountancy group. It follows the question through to what human oversight has to mean. Download the full guide for the questions that show whether your firm could reconstruct an AI-assisted piece of work today, and what to fix first if it could not.
FAQ
Deficient output is the first of three risk categories in the FRC's Generative and Agentic AI Guidance, published on 30 March 2026. The FRC defines it as the risk that issues in an AI system's inputs or design result in a deficient output, which places the failure in the data and setup rather than only in the tool.
You can use it, but responsibility stays with you. Under the January 2026 PCRT guidance, members remain ultimately responsible for work they produce irrespective of AI use. A practical standard is whether you could rerun the work by hand and reach the same result, which requires a record of what went in.
In most accounting firms the bigger risk is plausible output built from incomplete data, rather than invented figures. The tool assumes a complete client record and finds several partial versions. Its answer uses real numbers from a real system, so it looks right, and nothing in the output flags what was missing.