Is my bookkeeping data ever used to train AI models?

By default, yes, in anonymised form only, and you can opt out at any time, free. Your books may be worked on by human dedicated bookkeepers, AI tools, or a combination of both. And that mix may shift over time as systems improve.

Commitment: Either way, the commitment stays the same: your books are closed and your reports are delivered each month, regardless of how the work is produced.

Payroll only: If you use our Payroll & HR without Monthly Bookkeeping, none of your data is ever used to train anything of ours.

Verification: 1st Money uses AI tools to assist with tasks like transaction categorisation. Like any software, AI can occasionally produce inaccurate or incomplete outputs. High-confidence suggestions are coded automatically during the month. Anything less certain goes to a review queue, and your dedicated bookkeeper goes over the month's work before closing your books. No month closes without a person having checked it.

What the AI sees: Only transaction data is sent out for categorisation: the date, amount, currency, account name, the bank's own description, and a payment reference. No payslip data, no employee record, and no bank account details.

Opting out: If you'd prefer to opt out of AI training entirely, you can do it by asking your dedicated bookkeeper. It's free, and opting out never changes your price or your service.

The AI provider: That data goes to a U.S.-hosted AI provider under a written contract, and nothing sent for categorisation is ever used to train that provider's models.

Training data: By default, anonymised bookkeeping and payroll records may be used to help improve AI models over time. Identifying info is replaced with a fully imaginary identifier before anything reaches the training data: names, addresses, employee National Insurance numbers, bank account numbers, sort codes, VAT, UTR, and company numbers, geolocation data, counterparties, and the free-text description. We never attempt to re-identify anonymised data, and we contractually bar anyone we share it with from doing so.

If you opt out: We take your contribution out of the training corpus right away. It goes into no further training of any model, new or existing. Models we've already released aren't retrained or withdrawn. Every model is tested for memorisation before release, and one that can be made to emit anything identifiable isn't released.

Was this page helpful?

Ideas? Add your vote Tell us what to build next :)