AI in accounting and finance: ethics, evidence and the client

Cue 1 Five principles as five pillars, with a new technology panel fixed across them dated July 2025

EMMA 00:00 George, half my practice is using AI tools now and I've no idea whether we're allowed. Is there a rule?

GEORGE 00:08 There's a Code, and it already covers it. Five fundamental principles. Integrity, objectivity, professional competence and due care, confidentiality, and professional behaviour.

EMMA 00:16 Those are decades old.

GEORGE 00:18 They are, and in July 2025 ICAEW's Code was revised to add material on technology. It didn't invent a sixth principle. It made explicit that the five apply to the tools you use.

Cue 2 A payroll file leaving a practice into a public tool, with identifying details still visible after the name is removed

EMMA 00:31 So which one bites first?

GEORGE 00:33 Confidentiality. It's the one most easily broken by a conscientious person in a hurry. A trial balance, a payroll file, a client's letter from HMRC. Paste it into a tool your firm hasn't approved and it's held on terms you never negotiated.

EMMA 00:50 Has anyone official said that?

GEORGE 00:52 ICAEW has warned that putting client data of any kind into a public AI tool is a potential breach of confidentiality requirements.

Cue 3 A tribunal decision listing nine cases, each stamped does not exist

EMMA 01:01 What if I strip the client's name out?

GEORGE 01:04 Often not enough. A turnover, a sector and a town can identify a client. And where it includes employees or directors it's personal data too, so data protection law applies on top of the Code.

EMMA 01:18 Fine. Approved tools only. What about what comes back out?

GEORGE 01:22 That's competence and due care. These tools produce plausible text. Ask for case law and you can get cases that look right in every respect and don't exist.

Cue 4 A professional checking a generated statement against the legislation itself

EMMA 01:33 That's actually happened?

GEORGE 01:34 Harber against HMRC, First-tier Tribunal, December 2023. A taxpayer appealing a penalty relied on nine tribunal decisions. All nine had been generated by an AI tool. None was real. The appeal was dismissed.

EMMA 01:48 She wasn't a professional, though.

GEORGE 01:50 No. But in October 2025 a large professional services firm agreed to refund part of its fee to the Australian government, after a report was found to contain AI generated errors, including references nobody could trace.

Cue 5 A signed set of accounts with the question why this figure and a person answering it

EMMA 02:04 So what's the standard?

GEORGE 02:06 Understand the tool well enough to rely on it appropriately. For tax and legal statements, that means going to the legislation, the guidance or the decision itself. Every time.

EMMA 02:17 You mentioned objectivity. How does a tool threaten that?

GEORGE 02:21 Two ways. Automation bias, accepting an output because the tool produced it. And using a tool whose working you don't understand well enough to explain its result.

EMMA 02:32 Is there a test?

GEORGE 02:33 Yes. If someone asked why this figure, this classification, this conclusion, could you answer without saying the system produced it? If not, you're not yet in a position to sign.

EMMA 02:45 So the tool does the arithmetic.

GEORGE 02:48 And you hold the responsibility. That stays with the person whose name is on the work.

Cue 1 A ladder of review from reading a draft, to sampling categorised transactions, to tracing a figure, to checking the legislation

EMMA 02:54 George, last time was the principles. Now the practical bit. How much do I have to check what these tools produce?

GEORGE 03:02 In proportion to what depends on it. A draft email to a client needs a read. Transactions categorised by a tool need a sample checked, more where it was unsure or the amounts are large.

EMMA 03:16 And figures?

GEORGE 03:17 A figure that goes into accounts or a return gets traced to its source. A statement of tax or law gets checked against the legislation or guidance.

Cue 2 A working paper showing what was asked, what was produced, what it was checked against and what was changed

EMMA 03:28 Any warning signs?

GEORGE 03:29 Round, tidy, convenient answers. And anything the tool couldn't have known. A client's circumstances it was never told. A rule that changed after it was built.

EMMA 03:40 Say a client disputes something next year. What do I show them?

GEORGE 03:44 How the conclusion was reached. And a tool's output alone isn't evidence, because the same question may give a different answer tomorrow. So record what you asked, what it produced, what you checked it against, and what you changed.

Cue 3 Two guidance documents dated June 2025 and 30 March 2026 with the phrase appropriate confidence

EMMA 04:00 Has the regulator said anything?

GEORGE 04:02 The Financial Reporting Council has, for audit firms. Guidance on AI in audit in June 2025, with documentation guidance. And on 30 March 2026, guidance on generative and agentic AI.

EMMA 04:14 What's the gist?

GEORGE 04:15 That how much you validate and oversee should depend on the tool and how it's used. And that you need appropriate confidence in an output before you rely on it. It's audit guidance, but it's a sensible standard for anyone whose work may be reviewed.

Cue 4 An engagement letter with a short paragraph on AI tools, client data and review by a qualified person

EMMA 04:33 I've seen working papers that just say per AI tool.

GEORGE 04:37 That isn't a working paper. Say what was checked, against what, and by whom.

EMMA 04:43 Do I have to tell clients?

GEORGE 04:45 There's no single rule that says always. But integrity and professional behaviour point towards openness, especially where AI did research or analysis they're relying on.

Cue 5 Four duties unchanged by AI: money laundering, tax returns, data protection and decisions about people

EMMA 04:55 Some of mine would object.

GEORGE 04:57 Which is why you deal with it in advance. Put it in the engagement letter. Whether and how you use AI tools, what happens to their data, and that the work's reviewed by a qualified person. And check your indemnity cover while you're at it.

EMMA 05:15 Is there anything AI doesn't change at all?

GEORGE 05:18 Plenty. Anti money laundering is still yours. A tool may help screen a client, but due diligence, monitoring and the decision to report a suspicion stay human. Tipping off is still an offence.

EMMA 05:32 And tax returns?

GEORGE 05:33 The taxpayer's, and you remain responsible for the care you take as agent. Data protection applies as before. And where a tool shapes a decision about a person, like credit, UK rules revised in February 2026 give them a right to human intervention.

EMMA 05:50 So the short version.

GEORGE 05:51 No regulator accepts the software did it. If the duty was yours before the tool arrived, it's yours now.