Customer personal data and AI tools: a working guide

Consider a support agent with a difficult complaint. The customer is angry, the email thread is long, and the agent wants help writing a calm reply. The quickest path is to paste the whole thread into an AI chatbot. That thread contains the customer's name, email address, phone number, order history and, three messages down, the last digits of their card and their home address.

Nothing about the request was unreasonable. The problem was the paste. This post is about keeping customer personal data out of AI tools that are not approved for it, while still getting the help.

What counts as personal data

In data protection law, personal data is broadly any information about an identifiable person. Names and email addresses are obvious. Less obvious examples include:

• Phone numbers, home and delivery addresses.
• Government ID numbers, such as Aadhaar or PAN in India, or a Social Security number in the US.
• Order histories, account numbers and customer IDs that can be traced back to a person.
• Notes about a person: their complaint, their health condition, their family situation.
• Combinations of details that together point to one person, even without a name.

Where it slips in

Personal data rarely goes into AI tools on purpose. It arrives inside something else:

• Email threads pasted for a reply or a summary.
• Spreadsheet exports pasted for analysis: the column you wanted sits next to the column of phone numbers.
• Call notes and chat transcripts pasted for a summary.
• CRM records copied to draft a personalised message.
• Screenshots of a customer record, uploaded for help with a form.

Getting the help without the data

The useful part of most of these requests is the pattern, not the person. A few habits keep it that way.

1. Replace names with roles before pasting: the customer, the account manager, Supplier B.
2. Delete contact details. They almost never change the answer.
3. For spreadsheets, paste only the columns you need, or paste a few made-up rows that have the same shape as the real data and ask for a formula you then run yourself.
4. Summarise the situation in your own words: a customer received the wrong item twice and is asking for a refund plus compensation.
5. Use a tool your company has approved for customer data, if there is one, and only on the company account.

At Tidewater Software, a 50-person software company, the support team built a short set of prompt starters that begin with the situation in neutral words. Reply quality did not drop. The number of full threads pasted into chat tools fell sharply, because the starter was quicker than cleaning up a paste.

Screenshots and uploads

Pasted text is at least visible before it goes. A screenshot of a customer record or an uploaded spreadsheet is harder to check, because people rarely look at every cell or every corner of the image before sending it. Treat files and screenshots with more care than text, not less. Crop the image to the part that matters, or copy the few cells you need into a fresh sheet with the identifying columns removed. If a task truly needs the whole file, it probably needs an approved tool rather than a general chatbot.

When a tool is approved for customer data

Some companies do approve an AI tool for customer data, usually on a business plan with a data processing agreement, clear retention terms and a commitment about training. That is a reasonable decision, but it should be written down: which tool, which plan, which kinds of customer data, and who decided. People need to know that approval for one tool does not extend to the others. Laws such as the GDPR and the DPDP Act 2023 set out duties for anyone handling personal data, and an approved tool still has to fit within them.

In Openn Shield

The Openn Shield tool register records, for each tool, which kinds of data are allowed, including customer data and personal data as separate classes. Before pasting, people can run the text through the scanner, which looks for email addresses, Indian mobile numbers, international phone numbers, Aadhaar and PAN numbers, US Social Security numbers, card numbers and IBANs, all in the browser. Names and addresses are not patterns it can find, so the habit of replacing them still matters. See what the scanner covers.