ChatGPT tax advice: the hidden cost hitting UK accountancy practices

AI is producing measurable productivity gains in UK accountancy practices. That much is reasonably well established. Xero and KPMG have both published data suggesting that AI use across the profession is now majority practice rather than early adoption, with productivity improvements reported at a significant share of the firms that have deployed it.

But there is a second trend running alongside the good news, one that gets much less attention. According to a warning from Dext - the receipt and document capture platform - a large proportion of UK accountants report that clients are now using public AI tools for tax and financial advice, and that a meaningful share of those accountants are spending several hours a month correcting the errors that come out of it. Dext's figures, based on a survey of UK accountants, pointed to 77% reporting clients using ChatGPT-style tools for this purpose, with 93% of respondents spending up to 10 hours a month on correction work as a result.

I cannot independently verify those specific figures against the underlying study. They come from secondary reporting, and the methodology behind large survey numbers is often more conditional than the headline suggests. But the direction of travel is consistent with what the broader sector data supports: AI use by clients is widespread, largely unguided, and landing correction work on practices that had no hand in creating it.

That is a governance problem, not a technology one.

How AI is actually performing in practices

Before getting into the correction-work side, it is worth being clear that the practice-side picture is genuinely positive.

Industry data suggests that a majority of UK accountancy firms are now using AI in some form - Startups.co.uk, citing the Xero/KPMG research, reported 53% adoption, with 46% of those firms reporting productivity gains and, at an industry level, an uplift in the hundreds of millions of pounds attributed to AI-driven efficiency. The same research suggested a significant reduction in error rates for routine processing work.

These are plausible outcomes. Accounting involves substantial amounts of repetitive document processing, data extraction, and reconciliation. These are tasks where well-deployed AI performs consistently and where errors in a human-operated process are often caused by volume and fatigue. For firms that have set up AI well in back-office workflows, the gains are real and compounding. The post on where AI actually saves UK accountants time covers this in more detail - including the three workflows that consistently move the needle and the three where AI adds risk rather than removing it.

The problem is not that AI does not work in accountancy. The problem is that unguided client use creates a cost that the practice bears but did not cause.

The hidden bill

When a client uses ChatGPT to prepare their own tax position, classify a transaction, or work out whether something is deductible, they are often working from a tool that has no visibility of their specific situation, no access to their records, and no professional obligation to be accurate.

The ICAEW, in its guidance on AI and the PCRT (Professional Conduct in Relation to Taxation), is clear that professional judgement, human oversight, and transparency to HMRC are not optional extras - they are required elements of tax advice. A public AI tool is not bound by any of those standards. It cannot be.

The result, for practices on the receiving end, is a predictable pattern. A client arrives with figures that look plausible but are wrong. Or a transaction has been coded on a classification that does not hold up on inspection. Or a self-assessment return has been drafted using AI that confidently applied a rule that does not apply to this client's circumstances.

Each of those cases requires a professional to review, correct, and often have a conversation that is longer than the one they would have had if the client had come to them first. The correction work is unbillable more often than it should be, because the political difficulty of charging for fixing something the client already "did themselves" is real.

At the scale suggested by Dext's survey figures, that represents a material unrecognised cost - hours that absorb capacity without generating revenue, in a profession where margin is already under structural pressure.

Why it is a governance problem

The instinct in many practices is to treat this as a client education issue. Tell clients not to use AI for tax advice, or at least not without checking it with the practice first.

That is reasonable, but it is not sufficient on its own. The behaviour is not going to stop. AI tools are free, accessible, and getting better. A client who used ChatGPT to draft a reply to a supplier email last week is going to use it for tax questions too, because the tool does not distinguish. The question is not whether to prevent it - it is how to reduce the cost it creates for the practice.

The practices that handle this most effectively tend to do two things together.

First, they set a clear position with clients on where AI is and is not appropriate without professional review. Not a ban - a boundary. "Use it to help you gather information or draft questions to bring to us, but not to make tax decisions." That is a one-paragraph addition to the engagement letter or client welcome pack, not a lengthy policy document.

Second, and more important for the practice's own economics, they make sure their own AI deployment is focused on the work that generates the most value per hour - not on the work that sounds most impressive.

Where AI pays off first in a practice

The correction work created by unguided client AI use is a tax on practice time. The most direct offset is freeing up equivalent hours on the practice side.

The back-office workflows in most accountancy practices have more low-hanging fruit than the fee-earning workflows. Not because AI is more capable there, but because the volume is higher and the human cost is clearer.

Document ingestion and data extraction - client bank statements, receipts, supplier invoices - is the most frequently cited area of real time saving. Tools like Dext, AutoEntry, and Hubdoc have been doing AI-powered extraction for several years now, and the time savings are well documented in practice management data. A firm processing 200 client sets a year that can halve the manual handling on document intake has recovered a substantial number of hours without changing a single client-facing process.

Billing and time recording - specifically, reducing the gap between work done and invoices raised - is a less obvious AI target but a high-value one. Draft bill generation based on time entries, automated reminders when billing falls behind actual time recorded, and exception flagging when write-offs are accumulating all address the revenue leakage that most practices know exists but underquantify.

Meeting preparation and follow-up - drafting agendas from prior meeting notes, summarising what was agreed, and populating task lists - is where AI tools like Fireflies.ai or Otter.ai recover real associate and admin time without touching anything that requires professional judgement. For client-facing meetings, this might be five or ten minutes per meeting - modest individually but consistent across a full week. The post on shaving four hours off every new client onboarding covers how this sequencing works in practice for bookkeeping practices - the same back-office logic applies across the broader profession.

A small practice running 15 client meetings a week, processing documents for 200 active clients, and managing its own billing could realistically recover five to eight hours of productive capacity weekly by deploying AI in these three areas alone. That figure comes from workflow diagnostic conversations, not a published study. The actual number depends on what the practice's current process looks like.

Start with a workflow map, not a tool

There is a reason the correction-work problem persists even in firms that have adopted AI. It is not that those firms are doing anything wrong. It is that practice AI deployment tends to follow enthusiasm and availability rather than a structured assessment of where time is actually going.

When AI is deployed first in the places that feel interesting - early advisory work, client-facing tools, exploratory research - the back-office hours that drive actual margin stay as they are. The correction work from client AI errors then sits on top of an unimproved base.

The more effective sequence is to map where the hours go before choosing where to point the tool. Not a long exercise - a structured 60-minute conversation about workflows, then a prioritised view of where AI earns its place first and where human sign-off is mandatory.

For accountancy practices specifically, the governance question and the deployment question are really the same question. Where can AI work without professional review? And where is that review non-negotiable? Getting those two lists right - for both the practice's own work and the guidance they give clients - is the difference between AI that improves the practice economics and AI that generates a new category of unbillable work.

The ICAEW's guidance on AI and the PCRT, alongside the ICO's position on AI and personal data, provides the framework. Professional judgement, transparency, GDPR compliance, and human accountability for the advice are not areas where the profession has discretion. Any AI deployment that touches client data or tax positions needs to sit inside that framework.

What tends to be missing is not the regulatory clarity. It is a clear, practice-specific map of which workflows sit where - and what the cost of getting that wrong looks like in practice time.

---

If you want to understand where your practice hours are actually going before making any AI decisions, the free AI Readiness Quiz covers the highest-friction areas in about two minutes. The full HoursBack Assessment maps it across your whole workflow and comes back with a prioritised plan for what to deploy, in what order, and where human oversight is non-negotiable. £799, report within two working days.

---

Sources: Dext, "UK businesses losing money to ChatGPT-style tax and financial advice, accountants warn" (dext.com/en/news); Startups.co.uk reporting on Xero/KPMG AI in Accounting research (startups.co.uk/accounting/ai-in-accounting/). The 77% and 93% figures are drawn from secondary reporting of the Dext survey; the underlying methodology has not been independently verified against the primary study. Xero/KPMG figures (53% adoption, 46% productivity gains) are similarly drawn from secondary reporting and should be treated as indicative of direction of travel rather than confirmed primary data. ICAEW guidance references: PCRT (Professional Conduct in Relation to Taxation), available at icaew.com. ICO AI and data protection guidance at ico.org.uk.