A mortgage adviser has three meeting notes left to write up at five o'clock. She pastes her notes from the intake meeting into ChatGPT and asks for a tidy version for the file. The notes contain the client's name, her date of birth, the household income, two IBANs and a remark that her partner has been unable to work since last year.
The task was a good one: turning loose notes into a readable report is exactly what such a tool is good at. The data did not need to go in. That is the whole answer in one sentence: the task can usually go into an AI tool your firm has allowed for it, and the client should not be recognisable in it. What exactly comes out depends on what you ask the tool to do.
Why does this weigh more in financial services?
Because the text you handle on an ordinary day is almost always about a client, and because that client has told you what they tell few others. An income, a debt, an account number, the reason a mortgage does not add up after all. In a disability or claims file, medical information comes on top. That openness is the basis of the profession, and it is exactly what is at stake in a prompt.
The Dutch Data Protection Authority received several breach notifications in which employees had put personal data into an AI chatbot, and advises organisations to make clear agreements on whether it is allowed and with which data. For a firm under financial supervision, there is the added question of whether it handles client data with care, because that is part of sound business operations. What that means for a firm is in AI in financial services.
This is not about you as a risk. You use the tool because the note gets done faster and reads better. The only question is what goes into the tool with it.
Which ChatGPT are you using?
There is not one ChatGPT. A free or personal account runs under consumer terms: your input can be used to improve the model unless you switch that off yourself, and your firm has no data processing agreement with it. OpenAI states on its enterprise privacy page that the business environments do not use input for training by default. That is a different product with different agreements, even though the screen looks the same. The difference is set out in free versus business AI accounts.
A meeting note with an IBAN in a personal account is therefore a data breach, even if you delete the conversation straight afterwards. Which variant you may use is a decision for your firm. Many firms have allowed Copilot and have decided nothing about ChatGPT yet. "No decision yet" does not mean it is allowed; it means nobody has looked at it. As long as that is the case, no text with client data belongs in it.
At firms that work with BeeSensible, the decision appears in the browser the moment you open an AI tool: allowed, limited use with a condition, no decision yet, or not allowed with the alternative named. It does not stop you; it shows what has been decided.
