AI in accounting: The four ways it arrives in your firm

By Kenny MacAulay | November 4, 2026
AI in accounting: The four ways it arrives in your firm

AI in accounting rarely arrives through one decision. It reaches a firm through four routes: an approved rollout, features a vendor switches on inside software you already own, in-house builds, and shadow AI that staff find for themselves. Only the first goes through procurement. Since January 2026, putting client data into a public AI tool without consent is likely to breach confidentiality, so the other routes now carry conduct risk.

How is AI in accounting actually arriving in UK firms?

AI in accounting reaches most firms through four separate channels. Alistair Brisbourne, Head of Technology Research at ACCA, distinguishes them less by what each tool does than by what the firm can account for afterwards. On that test, only an approved rollout involves a decision anyone would recognise as one.

The appetite is not in doubt. Of the managing partners and chief executives ICAEW surveyed for its 2026 mid-tier research, 86% said their technology strategy explicitly includes AI adoption. That base is 35 UK firms of 11 to 249 principals, small but squarely the firms reading this. For most of them, the question is no longer whether AI is in the building. It is how it got there.

Channel How it arrived Visibility
Approved rollout Business case, procurement, contract Good. Board-level metrics exist
Features in existing software The vendor switched it on Poor. No new contract, no review
In-house development Built or commissioned internally Variable. Depends whether anyone documented it
Shadow AI Staff found it themselves None

Read down the right-hand column and the problem is plain. Three of the four routes left no record of who said yes.

The route partners think about least is already inside your software

The channel partners have thought least about is the AI feature that arrives inside software the firm already owns. The vendor switches it on, it bypasses the procurement process built to catch exactly that, and it sits outside contracts and data agreements signed before the capability existed.

Nobody signs anything, which is why it slips through. There is no business case and no review meeting. The data agreement your firm negotiated a few years ago was written for a product with no language model inside it. Nobody was asked to check whether the terms still hold, or to put the questions worth asking any provider about data residency and audit logging.

This is also where the risk compounds with the data you already hold. A feature that can read your files reaches everything in them, including client records you kept long past the point you needed them. If you are thinking about what an AI tool will find when it gets there, the steps in preparing accounting data for AI are a sensible place to start.

The practical test is an inventory of every tool in the firm with an AI feature, including the ones switched on inside software you already owned. The guide's own fail signal is blunt: if that list is shorter than your software inventory, it is not finished.

Why do controls on AI tools push the work onto personal phones?

A restriction that comes without an approved alternative moves AI use onto a device the firm cannot see. John Toon, Head of Technology at HLB International, oversees technology across roughly 350 member firms. He has watched sensible controls produce exactly that result, with client data leaving the agreement that protected it.

He describes a member firm that adopted ChatGPT, built with it quickly and hit its token limits far sooner than expected. It applied soft limits while it worked out the cost, which is the right instinct. Somebody who hits a limit halfway through a task might not want to stop, though. They open the free version of another tool and carry on.

That's probably one of the big risks, just not being aware of what your staff are using and where that data is going ... You can put website controls and you can put IP controls over certain sites, but that doesn't stop anyone from using their phone.
John Toon, Head of Technology, HLB International

Toon is clear that even a big firm with a completed security review would struggle if someone walked the floor and asked staff where client data was going. That is not carelessness. Staff are doing what firms have asked of them for two years, which is work faster.

Closing that gap needs an approved route that is easier than the workaround, and client data held somewhere the firm can see who touched it. Acting Office holds each client on one record with role-based permissions and a full audit trail, so an approved tool reads from a place the firm can account for.

What changed in January 2026 for client data and public AI tools?

Since January 2026, putting a client's figures into a public AI tool has been a professional conduct matter for anyone using AI in accounting work. Topical guidance from the seven PCRT bodies says it is likely to breach client confidentiality unless the client has consented. The obligation sits on the member, and through the member on the firm.

PCRT, Professional Conduct in Relation to Taxation, is written for tax work. Confidentiality is a fundamental principle in every body's code of ethics, though, so the same caution applies well outside tax.

Picture the ordinary version. Someone in your firm this week pasted a client's figures into a public AI tool to get a letter drafted faster. Nobody told them the tool keeps what it is given. Without the client's consent, that one act is likely to be a confidentiality breach in the firm's name, and the person who created it did not know it was one.

The controls have not kept pace, and that is measurable. DSIT's Cyber Security Breaches Survey 2025/26, a random probability sample of 2,112 UK businesses, found 31% using, adopting or considering AI. Only around a quarter of those had practices in place to manage the cyber risk that comes with it.

Four questions are worth putting to your own firm this week:

  1. Can you list every tool with an AI feature, including ones switched on by a vendor?
  2. Who approved the most recent one, and what did they see first?
  3. Do contracts stop providers and sub-processors training on client data?
  4. Who rechecks that when a product changes?

One of the four routes went past someone who could refuse it. All four are the firm's responsibility, and the answers to those questions are where a client or an insurer will start. Acting Office published Before you deploy AI to work through them on the record with contributors from ACCA, HLB International, Apteus Group and a private equity backed accountancy group. Download the full guide for the fourteen-question audit at the back. Each question carries its own fail signal, so you can see which of the four routes your firm can already account for and which it cannot.

FAQ

What is shadow AI in an accounting firm?

Shadow AI is any AI tool staff found and started using themselves, outside the firm's approval route. It has no visibility at all: no contract, and no record of which client data went in. It tends to grow where the firm restricts tools without offering an approved alternative that is easier to use.

Is it a breach to put client data into a public AI tool?

The PCRT topical guidance, published on 19 January 2026 by seven UK professional bodies, covers this directly. Putting client information into a publicly available AI tool is likely to breach client confidentiality unless the client has consented. The obligation sits on the individual member, and through the member on the firm.

How many UK accounting firms are adopting AI?

In ICAEW's 2026 research on mid-tier accountancy firms, 86% of the 35 responding firms of 11 to 249 principals said their technology strategy explicitly includes AI adoption. Respondents were managing partners and chief executives, and fieldwork ran from 23 February to 15 March 2026.

Note: Please note that the content of the above blog and the aforementioned information are solely for the purpose of awareness and are informative in nature. The content is designed with intent to ease the understanding while preserving the essence and importance of the compliance rules and shall not be considered as an ultimate replication of the rules. ActingOffice does not own any responsibility whatsoever for any unpleasant event that may arise due to the misinterpretation of a specific part or whole of the information.
View more latest Blog Post