Is it safe to put my business accounts into ChatGPT or Claude?
Short answer
It can be, if you choose the plan and settings with care and keep payroll and customer details out of a plain chat. Business plans of Claude and ChatGPT do not train on your content by default. Consumer plans depend on a setting. Under UK GDPR, you stay responsible for the personal data in your accounts.
Which plans use your content for training?
As their own help pages say on 4 November 2026:
| Plan | Used to train models? | Setting |
|---|---|---|
| Claude Free, Pro, Max | Only if you allow it, or if a chat is flagged for safety review | Model Improvement, in Privacy Settings |
| Claude for Work, Anthropic API | Not by default | Explicit feedback or opting in can change this |
| ChatGPT Free, Plus, Pro | Yes by default, until you turn it off | Improve the model for everyone, in Data controls |
| ChatGPT Business, Enterprise, API | Not by default | The organisation can opt in |
Thumbs up or down matters on both. Anthropic may keep a rated conversation for up to 5 years, and OpenAI says a rated conversation may be used even after you opt out. Do not rate a chat that holds accounts.
What in my accounts is personal data?
A company's own figures are not personal data, but much of what sits around them is. One made up bank line shows the difference:
| Item | Personal data? | In a plain chat? |
|---|---|---|
| Total sales for March: £18,400 | No | Fine, under the right plan |
| Card fee to a payment company | No | Fine, under the right plan |
| Payment from J Smith, a sole trader customer | Yes | Replace the name with a customer number |
| Salary paid to a named employee | Yes | Keep out |
| National Insurance number, home address | Yes | Keep out |
If you are a sole trader, your own bank statement is personal data about you too. You can choose to share it, but customers' and staff's details are not yours to give away lightly.
What does UK GDPR expect from me?
The Information Commissioner's Office says that buying an AI service from someone else does not take away your responsibility for the data. You decide what goes in and why, so for that you are normally the controller. A provider that only follows your instructions is likely to be a processor, which needs a written contract. A provider that uses the data to train its own models for itself becomes a separate controller for that.
In practice, for a small business:
- Choose a plan whose terms include data processing terms, and keep a copy.
- Say in your privacy notice that an AI service helps with your bookkeeping.
- Decide what never goes in. Write it down in one line for anyone who uses the tool.
This is a general summary, not legal advice. The ICO's guidance is general too: we found nothing written just for small firms using a chatbot for bookkeeping.
Does a connector change anything?
It changes how the data travels, not who is responsible. With a connector, the AI fetches what the task needs and does not get a file you pasted, so you can limit what it may do. What the AI reads still goes to your AI provider under your plan's terms, so the table above still applies.
In Octobooks, you tick what a connection may do, and you can end it at any time on the page Connections. Text written by banks, suppliers and customers reaches the AI marked as data, not as an instruction. Our data processing terms and sub-processors are public. To try this with one closed month, see run one month and check it.
Read next
More questions
- Can ChatGPT do my bookkeeping?
- Is it safe to let AI book entries in my books?
- Who pays if AI gets my VAT return wrong?
Sources
- Anthropic: is my data used for model training? (commercial plans, updated 18 August 2026)
- Anthropic: is my data used for model training? (consumer plans, updated 16 March 2026)
- OpenAI: how your data is used to improve model performance
- ICO: allocating controllership across the generative AI supply chain
- ICO: accountability and governance implications of AI