Sina Tadayon
CEO
Regulation
November 19, 2026

Can accountants use ChatGPT with client data?

What ICAEW and PCRT guidance say, and how to use AI in a practice safely

Yes, but not by pasting client-identifiable data into a public AI tool. The seven bodies behind PCRT, ICAEW among them, say doing that without the client's consent is likely to breach confidentiality. Generic questions are a different matter. Client work needs consent or properly anonymised inputs, a written AI policy, and a tool inside a secure environment.

Last checked against the sources: 17 November 2026. This is information, not legal or ethics advice. For a specific case, ask your professional body or a lawyer.

The AI policy pack for accounting practices.

What do ICAEW and the PCRT bodies actually say?

On 19 January 2026 the seven professional bodies that write Professional Conduct in Relation to Taxation (PCRT) published guidance on using AI in tax work (PCRT topical guidance, 19 Jan 2026). An updated PCRT applies from 1 January 2026 (ICAEW Tax News, 19 Jan 2026).

On confidentiality, the guidance says that putting client data into a publicly available AI tool is "likely to constitute a breach of client confidentiality", unless the client has consented (para 4.2).

Anything you do enter into a public tool should be anonymised and generic, so the client cannot be identified (para 4.4). And you stay ultimately responsible for the work you produce, whether or not AI helped.

ICAEW came back to the subject on 9 November 2026 (ICAEW Regulatory News, 9 Nov 2026). Its message: your confidentiality duties apply in full when you use AI, and entering client data into public AI tools risks unauthorised disclosure.

Sarah Brook, ICAEW's Director of Investigations, put it plainly. If a tool is not within a secure environment, entering information into it "can be akin to posting it publicly".

ICAEW also expects firms to have policies on generative AI that deal with the risk to confidentiality. If your practice has no written AI policy yet, that gap is now visible to your regulator.

The rule underneath all of this is R114.1 of the ICAEW Code of Ethics. It requires an accountant to respect the confidentiality of information acquired through professional and business relationships.

What counts as client data?

More than the client's name. ICAEW's own example is an accountant researching the tax effect of a client's planned share sale, who types the client's name and the sale into an open AI tool (ICAEW, 9 Nov 2026). Carrie Langridge, of ICAEW, compares that to pinning it on a virtual notice board.

When ICAEW assesses a slip, it asks whether the client's name, project names or details that could identify the entity were entered. Use the same test before anything is typed.

Taking the name out is often not enough. A turnover figure, a sector and a town can point to one small company as clearly as its name. The PCRT standard is that the client cannot be identified.

ICAEW's last check: would you be comfortable with this becoming more widely available? If not, keep it out of public tools.

What is the difference between a public AI tool and a secure environment?

Neither ICAEW's article nor the PCRT guidance lists approved products. In practice, four questions separate the two. They match what an article in ACCA's In Practice, written by the law firm Kingsley Napley, tells firms to ask a vendor (ACCA In Practice, 2026):

  1. Is what we type used to train the provider's models?
  2. How long is it kept, and where?
  3. Does the practice control the account, or does the employee?
  4. Is there a data processing agreement, with the provider acting as our processor?

The fourth question is a UK GDPR one. The ICO says you must be able to identify who acts as controller and who as processor, and that overall accountability stays with the controller, which is your practice (ICO, AI and data protection).

The ICO also says that in most cases, using AI involves processing likely to be high risk. High-risk processing requires a data protection impact assessment (DPIA) before you start.

What the two vendors say about their own plans, checked on 17 November 2026:

PlanIs what you type used to train models?Data processing agreementRetention controls for the firm
ChatGPT Free, Plus, ProIt may be. You can opt out in Settings > Data controls (OpenAI Help Centre)Not stated on the pages checkedNot stated on the pages checked
ChatGPT Business, EnterpriseNot by default (OpenAI business data page)OpenAI says it supports customers in signing a DPAQualifying organisations can configure retention
Microsoft Copilot, personal Microsoft accountIt may be. You can opt out in privacy settings (Microsoft Support)Not stated on the pages checkedNot stated on the pages checked
Microsoft 365 Copilot Chat, work accountNot used to train foundation models (Microsoft Learn, updated 3 Nov 2026)Covered by Microsoft's DPA and Product Terms, with Microsoft as processorThe organisation's Copilot retention policies apply

Sources for the table: OpenAI Help Centre, OpenAI business data, Microsoft Support, Microsoft Learn.

A business plan with training off and a DPA is a big step up from a free personal account. But it only deals with vendor risk. It does not give you the client's authority PCRT talks about, or replace your DPIA and your policy.

Whether a particular set-up counts as a secure environment is a judgement your practice has to make and write down. If you are unsure, ask your professional body.

ICAEW itself suggests firms consider AI systems that keep data within the firm. It also warns that staff may still reach for public tools or for AI features built into other software. A policy only works if staff have an approved tool to use instead.

What should an AI policy for a practice cover?

ICAEW expects one. ACCA's In Practice article says every firm should have a firmwide policy, so staff know when a tool may and may not be used. A workable policy covers ten points:

  1. Approved tools. Name each tool and the account type allowed for client work. Personal accounts are not allowed for anything client-related.
  2. What never goes in. Client names, identifiers, tax references, and any figures or facts that could identify a client.
  3. Consent. When you need the client's agreement, and where it is recorded. PCRT suggests reviewing engagement terms on how client data may be processed.
  4. Telling clients. ACCA's article suggests clients learn at the start which AI tools will be used on their work.
  5. Vendor checks. Training use, retention, storage location, encryption and the DPA.
  6. DPIA. Completed before an AI tool processes clients' personal data.
  7. Human review. A qualified person checks every AI output used in client work. Responsibility stays with the member.
  8. Training. Staff are trained on confidentiality risk, and training is refreshed as tools change.
  9. Incidents. Who staff tell, how fast, and who decides whether to report.
  10. Review date. ICAEW says policies should be reviewed regularly, because the tools change quickly.

Turn these ten points into your practice's policy.

Get the policy wording, a one-page summary your staff sign, and the regulator annex.

What if client data has already gone into ChatGPT?

Deal with it the same day. ICAEW sets out the questions to work through (ICAEW, 9 Nov 2026):

  • What was released? Which data, into which tool and account, and was it accidental or intentional?
  • What are the consequences for the client and for the firm?
  • Does it have to be reported to ICAEW? Breaking an internal AI policy is not automatically reportable. It becomes reportable if it involves an identifiable breach of the Code of Ethics, and then the duty to report misconduct applies.
  • Does the client need to be told? ICAEW reminds firms of their wider obligations to tell clients about confidentiality breaches.
  • Is it a personal data breach? If it is likely to put people's rights and freedoms at risk, the ICO must be told as soon as possible and, where feasible, within 72 hours. If the risk is high, the people affected must be told too (ICO, personal data breaches).

Record each decision and why. If in doubt, ask your professional body or a lawyer.

Questions practices ask

Is using ChatGPT a breach of client confidentiality?

Not in itself. The PCRT guidance says putting client data into a publicly available AI tool without the client's consent is likely to be a breach. Questions with no client details in them are a different matter.

Does ICAEW ban ChatGPT?

No. ICAEW says entering client data into public AI tools risks unauthorised disclosure, and it expects firms to have policies on generative AI. It does not list banned or approved tools.

Is ChatGPT Business safe for client data?

OpenAI says it does not train on ChatGPT Business or Enterprise data by default, and supports signing a DPA. That reduces vendor risk, but it does not replace client consent, a DPIA or your own policy. Whether it counts as a secure environment is your practice's judgement.

Can I use ChatGPT if I anonymise the data?

The PCRT guidance says data put into public AI tools should be anonymised and generic, so the client cannot be identified. For a small business client, removing the name is often not enough. Sector, location and figures can identify it too.

Does the guidance apply to ACCA members?

The PCRT AI guidance comes from seven professional bodies, and ACCA publishes it on its own website. An article in ACCA's In Practice also tells firms to adopt a firmwide AI policy.

Do I need a DPIA before using AI?

The ICO says most uses of AI involve processing likely to be high risk, and high-risk processing requires a DPIA. Complete one before an AI tool processes clients' personal data.

Who is responsible if the AI gets it wrong?

You are. The PCRT guidance says members remain ultimately responsible for any work they produce, whether or not they used AI.

Where does Letro fit?

If your policy bans public tools, your staff need somewhere approved to go. That is the gap Letro is built for.

Private AI runs on models Letro hosts in Europe, in an environment assigned to your organisation. No model provider sees your files. Nothing trains anyone's model.

Client conversations sit on the same platform. Messages are end-to-end encrypted; the server stores ciphertext. Hosted in Switzerland.

It does not replace your policy, your clients' consent or your own review. It gives your policy an approved tool to name. See how private AI for accountants works in a practice.

In preview. Switched on for founding firms first.

Give your team the rules before the first prompt.

Policy wording, staff one-pager and regulator annex, in one email.

Want to see the tool itself? Request Private AI early access →

Related articles

No items found.