Staff AI training: what to cover so people stop pasting client data

Good staff AI training for data protection takes about 30 minutes and covers five things: what never goes in, how to take details out, which tools are approved, a few real-looking examples, and a record of who completed it. A practice manager can run it without a lawyer. The aim is not to stop people using AI. It is to keep client data out of it.

Why a short session is worth it

Staff already use AI. Most of it is harmless. Some of it is not.

The Register, reporting on LayerX's 2025 report, found that 77% of enterprise AI users have been copying and pasting data into chatbot queries. About a fifth (22%) of those pastes include PII or payment card data. PII means personal details, such as names, addresses and dates of birth.

In the 2025 NCA and CybSafe survey, people who admitted sharing company data with AI most often named internal documents (50%) and customer data (44%), according to jazz.security. The same source reports a Gartner survey where over 57% used personal AI accounts for work.

The law is plain too. The ICO has stated there is no 'AI exemption'. If you process personal data, data protection law applies.

The session plan

Thirty minutes. One room or one call. Five parts.

1. Why it matters (3 minutes). AI is welcome. Client data stays out.
2. What never goes in (7 minutes). The list below.
3. How to take details out (8 minutes). Practise on one example together.
4. Approved tools (5 minutes). Which accounts to use. Which not to.
5. Quick check and sign-off (7 minutes). Three questions, then a dated agreement to the policy.

If you do not have a policy yet, start with our UK AI policy template. The training should match it line for line.

What never goes into an AI tool

Keep the list short. People remember short lists.

• Client names, addresses, emails and phone numbers.
• Bank details, card numbers and sort codes.
• National Insurance and NHS numbers.
• Passwords and secret keys.
• Lists of people, such as payroll or client exports.
• Whole client files, contracts or CVs.

How to redact before pasting

Redact means take out. Staff swap real details for labels. The question still goes through.

Teach three moves.

• Swap names for stand-ins. "Fenwick Joinery" becomes "Client A".
• Remove numbers. Account numbers, NI numbers and invoice numbers go.
• Paste the part, not the file. Only the paragraph the question needs.

For a fuller method, see how to redact sensitive data before pasting into ChatGPT.

Worked example

Use a fictional firm in the room. Ashcombe works well.

Priya wants help with a chasing email. Her first draft prompt reads: "Write a polite reminder to Fenwick Joinery about invoice 4471, account 12345678, contact Dave Fenwick."

Ask the group what comes out. The answer: the company name, the contact, the invoice and the account number.

The safe version reads: "Write a polite reminder to Client A about an overdue invoice from last month." The answer is just as useful. Priya adds the real details back in her own email.

Approved tools

Name the tools. Do not leave it to guesswork.

• Approved: the business accounts you pay for, signed in with work email.
• Not approved: personal accounts, free tools nobody has checked, browser add-ons that read pages.
• Not sure: ask before using it.

The ICO has also noted, in a January 2026 report summarised by Inside Privacy, that ad hoc experimentation with agentic AI can complicate oversight. Agentic AI means tools that act on your behalf, such as sending emails. A "not sure, ask first" rule covers them.

Track who completed it

Training only counts if you can show it happened. Measured Collective, summarising ICO guidance, notes that organisations should keep copies of training material and details of who received it. They should also use assessments, completion monitoring and follow-up for people who do not complete.

A simple tracker does the job.

Keep the slides and the quiz with the tracker. That is your evidence. Our guide to evidencing AI controls shows how to present it.

Spot who needs a refresher

Some signs are easy to see.

• A wrong answer on the quiz.
• New starters who missed the session.
• Someone asking about a tool that is not on the list.
• Policy changes, such as a new approved tool.

Training sticks better with a check in the background. Helio Guard checks each paste and upload on listed AI sites, on the PC. It takes out or warns about sensitive items and records the kind of data, site, person, outcome and time. We never see what your team types. It also records each person's dated policy agreement and flags staff who need training. It costs £1 a person a month, from 5 people, with a 14-day free trial and no card needed. You can see the details on our pricing page.

Common questions about staff AI training

How often should staff AI training be repeated?

Staff AI training should be repeated whenever the AI policy or approved tool list changes, and for every new starter at induction. Many firms add a short yearly refresher. Anyone who gets quiz answers wrong or misses the session should be followed up sooner, and the follow-up should be recorded in the same tracker.

Does AI training need to be done by a data protection specialist?

AI training does not need a data protection specialist for a small firm. A practice manager can deliver a short session using the firm's own policy, one worked example and a quick quiz. What matters is that the content matches the policy and that you keep the material and a dated record of who completed it.

Is training alone enough to stop client data going into AI?

Training alone rarely stops every mistake, because people paste in a hurry. Most firms pair staff AI training with approved business accounts and a check on what gets pasted. Training explains the rules, the check catches slips, and the records show clients, insurers or auditors that both are in place.

Related: Is it safe to paste payroll data into ChatGPT? What to strip out first