Does my accounting software use my data to train AI? How to check in ten minutes
Short answer
It may only if you instructed it or the data is truly anonymous. Your accounting software is normally your processor, and a processor may handle your books "only on documented instructions from the controller" (UK GDPR, Article 28(3)(a)). A provider that decides by itself to train its AI on your data is treated as a controller for that processing (Article 28(10)) and needs its own lawful basis. Anonymous data is outside GDPR; pseudonymised data is not (Recital 26). So the answer sits in four places: the DPA, the privacy notice, the AI feature terms and a switch in the settings.
The rule in one table
| What the provider does with your books | Under UK GDPR and EU GDPR | What you check |
|---|---|---|
| Keeps and processes them only to run the service | processor, Article 28 | the DPA names the purpose as delivering the service |
| Trains or improves its AI on them because your settings or contract say so | processor on your instruction | the instruction is written down: a switch you set, a clause you accepted |
| Trains its AI on them for its own product, by its own choice | controller for that processing, Article 28(10) | the privacy notice names the purpose and a lawful basis; your privacy notice should mention it |
| Uses only anonymous totals (no person identifiable) | outside GDPR, Recital 26 | the text says anonymous, not "de-identified" or "pseudonymised" |
| Uses pseudonymised data (names swapped for codes) | still personal data, Recital 26 | treat as training on your data |
All read on legislation.gov.uk on 5 November 2026: Article 28 and the recitals. The words to watch are "improve our services", "machine learning", "de-identified" and "aggregated": each can be fine or not, depending on whether a person can still be identified.
The ten minute check
- DPA (data processing addendum or agreement). Search for "train", "improve" and "model". A clean DPA says the provider processes your data only to deliver the service and binds any AI sub-processor to no training. (What else a DPA must contain.)
- Privacy notice of the provider. Search the same words. If it says it uses customer content to develop or train models, ask under which basis and whether you can opt out.
- Terms of the AI features (assistant, auto categorise, receipt reading). These sometimes have their own terms, separate from the main ones.
- Settings. Look for a switch about sharing data, improving the product or benchmarks. Note its default and set it the way you want.
Write the result in one line in your own records, with the date. If the provider adds or replaces a sub-processor, such as an AI provider, it must tell you first so you can object (Article 28(2)).
Worked example: we ran the check on our own terms
Octobooks, read on octobooks.ai on 5 November 2026:
| Place | What it says | Result |
|---|---|---|
| DPA, section 3.1 | we process the data "only on your documented instructions" | processor |
| DPA, Annex 1, point 9 | "If we use an AI provider, only under terms that forbid training on your data." | no training by a sub-processor |
| Sub-processors | no AI provider of our own; the AI you connect is yours | nothing goes to an AI provider from us |
| Help: data protection and your AI | "We do not use your data to train AI models, and we do not sell it." | no training |
| DPA, section 3.2, privacy note and Settings | a switch, *Share anonymous vendor categories*: only a vendor, a standard category and a count leave your books, no amounts, customers, people or free text, and a vendor counts only once three accounts use it | an instruction you can turn off; it improves booking suggestions, it trains no AI model |
Time taken: eight minutes, four pages. That is what a pass looks like: every place answers, and the one exception is a named switch.
Your own AI is a separate question
If Claude or ChatGPT reads your books through a connector, your accounting software's terms do not cover what the AI reads. That goes to your AI provider under your plan with them. On Claude, the consumer plans (Free, Pro, Max) train only if you allow it under Model Improvement, and keep data up to 5 years with training on, 30 days with it off; the business plans' Commercial Terms say Anthropic "may not train models on Customer Content" (Commercial Terms, seen 5 November 2026). Plan by plan, with ChatGPT: is it GDPR compliant to put my invoices into ChatGPT or Claude?
In Octobooks you also decide how much your AI sees: Standard, Careful (people as pseudonyms, contact and bank details left out) or Private (no personal data to the AI). How to set it: data protection and your AI. A general summary, not legal advice.
More questions
- Is it safe and GDPR compliant to put my invoices and bank statements into ChatGPT or Claude?
- Do I need a data processing agreement (DPA) with my accounting software?
- Where should my accounting data be stored under UK GDPR? Does it have to be in the UK?
Sources
- UK GDPR, Article 28: processor only on documented instructions; 28(10) processor becomes controller, seen 5 November 2026
- UK GDPR, Recital 26: anonymous and pseudonymised data, seen 5 November 2026
- Anthropic: Commercial Terms, seen 5 November 2026
- Anthropic: is my data used for model training? (consumer plans)
- Octobooks: Data Processing Agreement