The Most Common AI Mistakes Accounting and Professional Services Make

KPMG Clara and similar tools produce files that tick every compliance box but contain no trace of professional scepticism. Partners review the file checklist as complete and sign it off without asking whether the auditor actually challenged management's estimates or identified a going concern risk that the tool missed.

The fix

Before sign-off, require the engagement team to document in writing the three audit judgements where they overruled or questioned the AI output, and why.

The tool calculates materiality thresholds based on financial metrics, but it does not know whether a listed client announced a strategic restructuring last month or whether a regulatory body issued new guidance on a key revenue stream. Audit files then proceed with benchmarks that are technically correct but contextually blind.

The fix

Require the engagement partner to review external news, regulatory filings, and board minutes for each client before accepting Halo's materiality output, and adjust in writing if warranted.

Deloitte AI and Thomson Reuters tools select samples by statistical methods without understanding which transactions are unusual or high-risk in your client's business model. A tool might sample low-risk routine items while missing the one-off adjustment that needed audit attention.

The fix

Have a senior auditor with client knowledge override the AI sample to include specific high-risk or unusual transactions, and document the override reason.

Junior staff ask ChatGPT about lease accounting or revenue recognition and get a coherent answer that sounds authoritative. The answer lacks the nuance of IFRS Board interpretations, ICAEW guidance, and the specific fact patterns your firm has seen in disputes with regulators.

The fix

Require junior staff to cite IFRS standards and relevant ICAEW or professional body bulletins by name before accepting ChatGPT as a starting point, not an ending one.

Files generated by KPMG Clara or similar tools are audited and signed, but the audit file record does not say whether substantive procedures were performed by the tool, the auditor, or both. When regulators ask about procedure design, you cannot trace the chain of judgement.

The fix

Require a separate section in every audit file that lists each AI tool used, what it did, and what human review was performed before accepting its output.

The tool retrieves case law and statute but does not know that HMRC has soft-positioned a tax issue in practice notes or that your region's local tax office handles a ruling differently. You give advice that is legally sound but commercially wrong because it ignores how HMRC actually administers the tax.

The fix

Before finalising any tax position, have a tax partner check the relevant HMRC practice note, Tax Bulletin, and HMRC Manuals to see if real-world practice differs from the statute the AI cited.

ChatGPT or generic AI tools give answers about tax treatment of employee benefits or corporate restructuring that work in England but not Scotland, or that miss Northern Ireland-specific rules. A client implements the advice in the wrong jurisdiction and creates a compliance failure.

The fix

Before giving any cross-border tax advice, explicitly identify which jurisdictions apply and search professional databases for country-specific guidance, not just general principles.

Tools like Deloitte AI or ChatGPT can generate technically correct tax explanations, but they do not know your client's tolerance for risk, the relationship history with that CFO, or the way you usually communicate bad news. A poorly toned email can trigger client frustration or trigger an unnecessary query back.

The fix

Before sending any substantive client communication that originated from an AI draft, have a partner read it for tone and consider whether it could damage the relationship if misunderstood.

AI summarises case law on business purpose or substance over form but skips the fact-specific distinctions that have made similar cases fail or succeed. Your advice looks solid until a client implements it and a dispute arises where the distinguishing facts matter more than the rule the AI emphasised.

The fix

For any tax position involving mixed motives, require a partner to read the full text of the three most relevant cases and document why the client's facts are closer to the win than the loss.

Thomson Reuters or other tools calculate uncertain tax positions by applying probability percentages to exposure amounts. The AI does not know whether your tax team actually believes those percentages are realistic given your client's risk profile or past audit history. The provision ends up too aggressive or too conservative, and when auditors ask why, the answer is just the AI's assumptions.

The fix

Require your tax director to sign off on the probability and exposure assumptions in every AI-generated provision calculation, and require written reasoning if they differ from what the tool suggested.

A client emails a question about a compliance deadline or a tax deadline, and an AI chatbot or ChatGPT answers it on your website. The answer is factually correct but does not know your firm's policy on how conservative to be, which interpretations you prefer, or which clients you advise differently. The client gets advice that conflicts with what your partner would give.

The fix

Disable or remove public-facing AI chatbots that answer substantive questions unless a qualified adviser reviews every answer before the client sees it.

A new prospect inquiry comes in and is routed to a chatbot or an automated response instead of a person. In professional services, the first exchange builds confidence that you take the client seriously. Automation signals that you do not, and the prospect goes to a competitor who picks up the phone.

The fix

Route prospect inquiries and new client questions directly to a named adviser, not a tool, and use that conversation to decide whether the engagement fits your firm before you send any substantive response.

A junior accountant asks ChatGPT to explain how to test a revenue cycle instead of asking a senior or studying the firm's methodology. The junior gets an answer fast, understands nothing about your firm's approach, and will never develop the judgement to spot problems that the AI would miss. The profession loses capability at scale.

The fix

Require juniors to document what they learned from a senior adviser or a training resource before they start work, and prohibit using AI to answer methodological questions until they have done the task three times without it.

A tax summary or advisory report is drafted by Deloitte AI or a similar tool based on data tables. It is accurate but generic. A client reads it and sees no evidence that you understand their business, their risks, or the conversations you have had with them. They do not renew because the output could have come from anyone.

The fix

Require every client-facing report to include a summary paragraph written by the engagement partner that names specific business challenges discussed with the client and explains how the findings address them.

A partner reviews an AI-generated tax analysis or audit section and checks it for obvious errors, but does not know what information the tool was not fed and what judgements it could not make. The output goes to the client with gaps that experienced advisers would have spotted. The client finds them first or a regulator raises them later.

The fix

Train all partners to ask themselves three questions before accepting any AI output: What client-specific facts would an experienced adviser know that this tool was not told? What related guidance or precedent might it have missed? What does this output assume about our risk appetite?