Why 'AI for accountants' must be approval-gated — PCRT is not optional

It's 8pm on a Tuesday. A client sends a WhatsApp: "Quick one — am I better off paying dividends or salary this year?" You know the shape of the answer. But you haven't looked at their 2025-26 numbers yet, and you don't know if they changed their pension contributions in January.

Now imagine an AI tool you use has already replied on your behalf.

That's not a hypothetical. It's the direction several "AI for accountants" products are heading — AI that connects directly to your client-facing channels and responds to queries without a human in the loop. Fast. Convenient. Terrifying.

PCRT is the reason this matters

If you're a practising accountant in the UK, the Professional Conduct in Relation to Taxation (PCRT) standards aren't optional reading. Whether you're a member of ICAEW, ACCA, CIOT, ATT, or AAT, your professional body has adopted PCRT as the ethical framework for tax work. It covers everything from confidentiality to the accuracy of returns — and it is explicit that you, the practitioner, are responsible for the advice you give.

Not the software. Not the AI. You.

PCRT requires that tax advice is based on a proper understanding of the client's affairs. It requires that you do not mislead HMRC. It requires that you take care with estimates and judgements. Those are standards that apply to a human professional exercising professional judgement — they cannot be delegated to an algorithm, and they certainly cannot be satisfied by an AI that fires off a reply before you've seen it.

When you tell a client something about their tax position, you are putting your professional reputation and your firm's registration behind that answer. That's what professional membership means.

Two shapes of AI risk

Most AI tools aimed at accountants sit in one of two categories.

The first category is the genuinely dangerous one: AI that talks directly to your clients. It sends emails, posts replies, populates client portals — all without your review. The vendors frame this as efficiency. What they are actually selling you is a way to outsource professional judgement to a model that has no professional indemnity insurance and no idea what that client said to you on the phone last month.

The second category is subtler and more common: AI that drafts responses for you to copy and send. You do press "send" — so technically you've reviewed it, right? Maybe. But if there's no audit trail showing that you read the draft, that you considered whether it was accurate, that you actively approved or amended it before it reached the client — then in practice you have no defensible record. Under a complaint or an HMRC inquiry, "the AI drafted it and I sent it" is not a strong position.

Neither of these shapes is acceptable once you think about what PCRT actually asks of you.

The audit trail is not bureaucracy

When something goes wrong in an accountancy firm — a client disputes advice, HMRC opens an enquiry, a complaint reaches the professional body — the question that follows is always the same: what did the accountant know, when did they know it, and what did they do with that information?

A written audit trail answers those questions. It shows that a question came in, that someone in the firm reviewed the relevant facts, that an answer was prepared, that a qualified person signed off on that answer, and that the client received only the approved version. That's not box-ticking. That's the chain of custody that protects your licence to practise.

AI makes the drafting part of this chain faster. That's genuinely useful. But it only helps you if the rest of the chain is intact — if the approval step is real, documented, and attributable to a specific person at a specific time.

Without that, AI is not a productivity tool. It's a way to produce more undefended advice, faster.

Why most products skip the hard bit

Building approval gates into software is more work than building a chat interface. It requires you to think carefully about the workflow: how does a question arrive, who can answer it, what happens when the accountant disagrees with the AI draft, what does the client portal show before approval happens, what timestamp and user ID goes into the log?

These are not glamorous product decisions. They don't make for exciting demos. But they are the decisions that determine whether a firm is protected or exposed.

Most products skip them because their buyers — until recently — haven't been asking the right questions. The pitch is "save time answering client queries." The response is "great, how do I sign up?" Nobody asks what happens when the AI is wrong about a client's marginal rate. Nobody asks where the audit log lives.

That's changing. HMRC's Making Tax Digital programme is driving more firms to think seriously about digital record-keeping and accountability. Firms that are already operating under PCRT scrutiny — firms that have been through a professional body complaint, or advised on an IR35 dispute, or dealt with an HMRC compliance check — those firms understand immediately why approval-gating matters. They've seen what it looks like when the paper trail is thin.

What responsible AI looks like in practice

The shape that works is simple to describe, even if it takes care to implement.

A client asks a question through a secure portal — not WhatsApp, not email, not a chatbot that pipes directly to a reply. The question is received by the firm, not the AI. The AI then drafts a response, drawing on whatever context the accountant has provided. The draft sits in a queue. An accountant reads it, edits it if needed, adds anything the AI missed, and then — explicitly, with their name attached — approves it. Only after that approval does the client see anything.

The log records every step: question received, draft generated, reviewed by whom, approved at what time, sent. If the accountant rewrote the entire answer, the log shows that too.

This isn't slower than the alternative. With a well-designed tool, it takes seconds to approve a straightforward answer. But it is fundamentally different from an AI that bypasses the professional entirely.

How Approvox handles this

I built Approvox because I couldn't find a tool that took the approval step seriously.

Every question a client asks lands in the accountant's dashboard — not in the client's inbox. The AI drafts an answer, but the client sees nothing until the accountant has reviewed and approved it. The accountant can approve as-is, amend the draft, or write their own answer from scratch. Whatever they do, it's logged: who reviewed it, what they approved, when it was sent.

That log is not hidden in a database somewhere. It's visible to the accountant and exportable. If a professional body inquiry or an HMRC compliance check ever asks what happened on a particular client query, the answer is there — time-stamped, named, complete.

We also mark every AI-generated draft clearly as a draft. The accountant is never in any doubt that they are approving AI output — which means they approach it with the appropriate scepticism rather than treating it as a finished answer.

None of this is complicated. It's just a decision to put the professional back in the loop rather than treating them as optional.

If you want to see how it works in practice, you can start a 14-day trial — no card needed — at https://approvox.co.uk/register.