Client data in ChatGPT: what accountants need to know before using AI
Putting client data into ChatGPT or another general assistant is only defensible on a business plan with a data processing agreement, and even then only after you have removed what the task does not need. Of the 4 main assistants we checked, all 4 say their business plans do not train on your data by default, while the consumer versions of ChatGPT and Gemini can use your chats for training unless you change a setting.
That is the short answer for accountants asking "is it safe to use ChatGPT with client data?". The longer answer depends on which plan, which settings, what data, and whether your practice has the right contract in place. This guide sets out what each vendor's own pages say, what UK GDPR expects of you, and a checklist you can use before any client data goes near a general assistant. If you want the wider picture first, our free lesson on what AI can and cannot do in an accounting practice covers where these tools fit.
This is general information, not legal advice. Vendor terms change, so check the linked pages yourself before you rely on them. We checked every vendor claim below on 15 November 2026.
Consumer plans and business plans are different products
The single most important point is that the same brand name covers very different terms. A free or personal account is a contract between the vendor and one individual. A business plan is a contract between the vendor and your firm, usually with a data processing agreement, admin controls and different defaults on training and retention.
With a personal account, the practice has no contract with the vendor and no admin control, which is why our AI policy template for accountants bans personal accounts for client work.
What the vendors say: the comparison table
| Assistant and plan | Inputs used to train models? | Retention and controls | Processor terms |
|---|---|---|---|
| ChatGPT Free, Plus, Pro | May be used, unless you turn off "Improve the model for everyone" or opt out in the Privacy Portal. Temporary chats are not used. Feedback can bring a whole conversation into training even after opting out. | Retention period not stated on the page we checked | Consumer terms, no DPA stated |
| ChatGPT Business, Enterprise | Not by default | Business: admins control retention, deleted conversations removed within 30 days (with legal exceptions). Enterprise: you control retention | DPA available for Business, Enterprise and API |
| Microsoft 365 Copilot | Prompts, responses and Microsoft Graph data not used to train foundation LLMs | Stored in line with your Microsoft 365 contractual commitments; admins can set retention in Microsoft Purview | Microsoft Product Terms and Data Protection Addendum |
| Gemini in Google Workspace | Not without the customer's prior permission or instruction | Admins can auto-delete Gemini app conversations after 3, 18 or 36 months; default 18 months | Prompts treated as customer data under the Cloud Data Processing Addendum |
| Gemini app, personal Google account | Used for training, including generative AI models, when Keep Activity is on | Default auto-delete 18 months; with Keep Activity off, kept for 72 hours; chats seen by human reviewers kept up to 3 years | Consumer terms, no DPA stated |
| Claude Free, Pro, Max | Used only if you opt in to model improvement, or if a chat is flagged for safety review | With model improvement on, de-identified data may be kept up to 5 years in training pipelines; deleted chats removed from back-end within 30 days | Consumer terms, no DPA stated |
| Claude Team, Enterprise, API | Not by default; commercial terms say Anthropic may not train on customer content | API inputs and outputs deleted within 30 days; deleted chats removed from back-end within 30 days | Data Processing Addendum incorporated into commercial terms |
OpenAI ChatGPT
OpenAI's help article on how your data is used to improve model performance says that for its services for individuals (Free, Plus, Pro) it may use your content to train its models, and that you can opt out by turning off "Improve the model for everyone" under Settings > Data controls, or through its Privacy Portal. It also says temporary chats are not used to improve its models, and that if you give feedback with a thumbs up or down, the entire conversation may be used for training even if you opted out.
For business plans, OpenAI's enterprise privacy page says it does not train on ChatGPT Business, ChatGPT Enterprise or API data by default. It says ChatGPT Business admins can control retention, with deleted or unsaved conversations removed within 30 days unless longer retention is required by law or to protect its services. It also says OpenAI can execute a Data Processing Addendum for ChatGPT Business, Enterprise and the API.
Microsoft 365 Copilot
Microsoft's data, privacy and security page for Copilot states that prompts, responses and data accessed through Microsoft Graph are not used to train foundation LLMs. It says Copilot only surfaces organisational data the individual user already has at least view permission for, which means your SharePoint and Teams permissions matter more than ever: if a junior can open a folder, Copilot can draw on it for them.
The same page says admins can set retention through Microsoft Purview and that data residency commitments sit in the Microsoft Product Terms and Data Protection Addendum. It also notes that when Anthropic models are used inside Copilot, they are currently excluded from the EU Data Boundary. This guide does not cover Microsoft's consumer Copilot app, which sits under different terms.
Google Gemini
Google's Generative AI in Google Workspace Privacy Hub states that Workspace does not use customer data for training models without the customer's prior permission or instruction, and that content is not human reviewed or used for model training outside your domain without permission. It says prompts are customer data under the Cloud Data Processing Addendum, and that admins can choose to delete Gemini app conversations after 3, 18 or 36 months, with 18 months the default. The Privacy Hub applies only to qualifying Workspace editions.
The consumer picture is different. Google's Gemini Apps privacy hub says that with Keep Activity on, your activity is used to develop and improve its services, including training generative AI models. It says a subset of chats is reviewed by human reviewers, that those reviewed chats are kept for up to three years even if you delete your activity, and that with Keep Activity off, chats are still kept with your account for 72 hours.
Anthropic Claude
Anthropic's consumer article, is my data used for model training?, covers Free, Pro and Max. It says chats are used for training if you opt in to model improvement, if a conversation is flagged for safety review, or if you opt in another way. Incognito chats are not used. Its retention article for consumers says that with model improvement on, de-identified data may be kept for up to 5 years in training pipelines, and that deleted chats are removed from back-end storage within 30 days.
For commercial products, Anthropic's commercial training article says it will not use inputs or outputs from commercial products by default. Its commercial terms state that Anthropic may not train models on customer content and incorporate its Data Processing Addendum by reference. Its organisation retention article says API inputs and outputs are deleted within 30 days, and deleted chats are removed from back-end storage within 30 days. That article does not name the Team plan specifically.
UK GDPR: controller, processor and the contract
When your practice decides why and how client personal data is used, the practice is the controller. A supplier that processes that data on your behalf, such as an AI vendor under a business plan, is acting as a processor.
UK GDPR requires a controller to have a written contract with any processor, and the ICO sets out what needs to be included in that contract: what is processed and why, the processor acting only on your instructions, security, confidentiality, sub-processors and deletion at the end of the service. A vendor's DPA is how that requirement is usually met.
That is why consumer plans are a poor fit for client work: without processor terms, nothing meets the contract requirement, however good the privacy settings.
A few other points apply whichever plan you choose:
- Data minimisation. Only put in what the task needs. Most drafting and summarising does not need names, UTRs, NI numbers or bank details.
- International transfers. Check where the vendor processes data and what transfer safeguards its terms rely on.
- Breaches. If client data goes somewhere it should not, the ICO's guidance says a notifiable breach must be reported within 72 hours of becoming aware of it.
Confidentiality under accountancy ethics codes
Data protection law covers personal data. Professional ethics goes further. The ICAEW Code of Ethics lists confidentiality as one of five fundamental principles, alongside integrity, objectivity, professional competence and due care, and professional behaviour. The ICAEW Code is based on the international IESBA Code; if you belong to another body, check its code for the equivalent wording.
In general terms, the duty covers information acquired through professional and business relationships, not just personal data. That includes a limited company's figures, a client's business plans and a dispute you are advising on. The duty also expects you to take reasonable steps to make sure the people and services you rely on respect that confidentiality.
The safer route: AI that already sits with your data
For many tasks you do not need a general assistant at all. The AI features inside the ledger work on data that is already held by a supplier you have terms with: see our pages on Xero's AI features, Intuit Intelligence in QuickBooks and Sage Copilot. Read each supplier's AI terms too, but you are not copying client data into a new system.
Checklist: before client data goes into a general assistant
Copy this into your practice procedures.
CLIENT DATA AND GENERAL AI ASSISTANTS - PRE-USE CHECKLIST
[ ] The tool is on the practice's approved list.
[ ] We are on a business plan (not a free or personal account).
[ ] A data processing agreement or addendum is in place with the vendor.
[ ] Training on our data is off by default, and we have checked the vendor's page.
[ ] Retention is set by an admin and we know how long data is kept.
[ ] We know where the data is processed and what transfer safeguards apply.
[ ] Our privacy notice mentions AI suppliers.
[ ] For this task: names, UTRs, NI numbers and bank details removed unless needed.
[ ] No passwords, Government Gateway details or bank logins entered.
[ ] A competent person will review the output before it reaches the client.
[ ] If something goes wrong, report internally the same day.
To train the whole team on these checks and the tools behind them, see our AI course for accountants.
This is covered in lesson 1 of our free-to-start course, Practical AI for UK accounting practices: read the free lesson.
Frequently asked questions
Is it safe to use free ChatGPT with client data?
We would not recommend it. OpenAI's own help page says content from Free, Plus and Pro may be used for training unless you opt out, and the practice has no processor contract with OpenAI for a personal account. Use a business plan with a DPA, or an in-ledger tool, instead.
If I turn off training in a consumer account, is that enough?
It reduces one risk, but it does not create a contract between your practice and the vendor, and it leaves admin control and deletion with the individual. Under UK GDPR, a processor needs a written contract with the controller. A setting is not a contract.
Is Microsoft 365 Copilot safer because it is inside our tenant?
Microsoft says Copilot respects existing permissions and does not train foundation LLMs on prompts, responses or Graph data. The catch is that it surfaces anything a user can already open, so loose SharePoint or Teams permissions become a bigger problem. Tidy access before rolling it out.
Does anonymising data solve the problem?
It helps a great deal, and it should be the default. Be aware that removing a name does not always make data anonymous: a combination of business type, town and turnover can still identify a client. Treat "reduced" data as still confidential.
Do we need client consent to use AI on their work?
Not necessarily consent as such, but you do need a lawful basis for the processing, processor terms with the vendor and transparency in your privacy notice. Many practices also mention AI in the engagement letter. Check your own position with your adviser or professional body.
Which general assistant is best for an accounting practice?
There is no single answer. The best choice is usually the one that fits the software you already run, such as Microsoft 365 or Google Workspace, on a business plan with a DPA. Compare options in the AI tool directory.
Keep reading
- Xero vs QuickBooks UK 2026: Prices, MTD and Worked CostsXero and QuickBooks UK plans priced side by side, with 12-month costs for a sole trader, a VAT-registered trader and a 10-employee company, plus MTD readiness.
- QuickBooks Pricing UK 2026: Every Plan From £10 a MonthIntuit's five UK QuickBooks Online plans list at £10 to £128 a month. Here is what each includes for MTD, what payroll adds, what accountants pay through Intuit Accountant Suite and ProPartner, and how the price compares with Xero, Sage and FreeAgent.
- Sage Copilot Review 2026: What It Does and Sage Accounting Prices UKA buyer's guide to Sage Accounting with Sage Copilot: every published UK price, three worked cost examples, what the AI assistant does today, the cheaper route through an accountant, and how it compares with Xero and QuickBooks.
Free: a practice AI policy template
A ready-to-adapt policy for using AI tools with client data, in Word and PDF. Free with the 5-minute Friday brief on AI tools for accountants.
We email you the template and the Friday brief. Unsubscribe in one click.