AI and Data Protection: What Not to Paste In
Share
Most workplace AI incidents are not sophisticated. They are somebody pasting something into a browser tab because it was quicker.
Direct answer
Do not put personal data, confidential business information or anything under an NDA into a public AI tool. Tools operating inside your organisation's own tenant under existing controls are treated differently. The practical test is whether you would email it to a supplier.
Based on the AI and Automation Practitioner standard (ST1512), a Level 4 apprenticeship for people already in work. Funding rules described here are the England position and are set by the funding body.
Potentially free for you
The learner never pays. Employers potentially pay nothing either, and where the apprentice is aged 16 to 24 the government funds the whole cost even if the employer does not pay the levy.
Closing any gap. A levy transfer from another employer covers 100% of the cost. Unused transfer allowance is not carried forward at the end of the tax year, and it must be agreed before the apprentice starts.
From 1 August 2026 the government stopped adding the 10% top-up to new funds entering levy accounts, so there is less in the system than there was.
Part of our guide to AI Explained in Plain English for Work.
The three pastes that cause incidents
Customer records, to get a faster reply drafted. This is the commonest and it involves personal data by definition.
Employee information, usually in an HR or performance context. Higher sensitivity and frequently special category data.
Contract or commercial terms, to get a summary. Not personal data, and a confidentiality breach that can carry its own consequences.
In all three cases the intent is reasonable and the outcome is that information left your organisation.
Why a settings toggle does not fix it
Opting out of training is a promise about future use rather than a control over what already happened.
The data was transmitted, processed and probably logged. Those are facts regardless of the setting.
Under UK data protection law the relevant question is whether you had a lawful basis and appropriate safeguards, not whether a toggle was set.
That is why the answer is a boundary on what goes in rather than a configuration on what happens after.
What is genuinely fine
Anything with no personal or confidential information in it. Drafting generic text, explaining a concept, structuring an argument.
Anonymised or synthetic examples. Rewriting a real query with invented names and details usually gets you the same result.
Your own work product that contains nothing about anyone else.
Tools operating inside your organisation's tenant under existing permissions, subject to your own policy.
There is more on this in our guide to AI explained in plain English.
The practical rules to give staff
Name the categories specifically. Customer records, employee data, contract terms, anything under an NDA, anything with bank details.
Give the test alongside. If you would not email it to a supplier, do not paste it in.
Name the approved tools and say which are inside your tenant and which are not. Staff cannot follow a rule they cannot apply.
Name a person to ask. Policies without a named contact produce guessing, and guessing produces incidents.
If it has already happened
Report it quickly. Speed matters more than blame and organisations that punish reporting stop hearing about incidents.
Record what was shared, with which tool, when, and by whom. That record is what any assessment depends on.
Assess whether it meets the threshold for reporting to the regulator. That is a decision for whoever holds that responsibility.
Fix the cause rather than the instance. Someone pasted it because the approved route was slower, and that is the real finding.
The specific UK GDPR points
Your employer is the controller for anything staff put into a tool, and the tool provider is usually a processor. That relationship needs a contract.
International transfers need appropriate safeguards. Many tools process outside the UK and the free tiers say less about this than the business tiers.
Special category data, which includes health, biometric and trade union information, has a higher bar and should never go into a public tool.
The Information Commissioner's Office publishes free guidance for employers on all of this, written for practitioners rather than lawyers.
What to build into your policy
A clear list of what must never go in, in specific terms rather than as a principle.
A statement that reporting a mistake quickly will not be treated as misconduct, which is what gives you visibility.
A route to request a new tool, because without one people just sign up.
A review date, because tooling and licensing change faster than policy documents usually do.
Key takeaways
- Customer records, employee data and contract terms are the three risky pastes.
- An opt-out toggle is a promise about future use, not a control over what happened.
- Anonymised or invented examples usually get you the same result safely.
- If you would not email it to a supplier, do not paste it in.
- Say that reporting a mistake quickly will not be treated as misconduct.
Frequently asked questions
What can I not put into AI at work?
Personal data, confidential business information and anything under an NDA, unless the tool operates inside your organisation's own tenant under existing controls.
Does opting out of training make it safe?
No. The data was still transmitted, processed and probably logged. An opt-out is a promise about future use rather than a control over what happened.
Is it safe to paste customer emails in?
Not into a public tool. Rewrite the query with invented names and details instead, which usually produces the same result safely.
What if my organisation has no policy?
Apply the practical test yourself. If you would not email it to a supplier, do not paste it into a public tool.
What should I do if I already pasted something?
Report it quickly. Record what was shared, with which tool and when. Speed matters more than blame.
Are Copilot and similar tools different?
Tools operating inside your own tenant under existing permissions are treated differently, subject to your organisation's own policy.
Can I use anonymised data?
Usually yes, provided it is genuinely anonymised rather than lightly disguised. Invented examples are safer still.
What makes a good AI data policy?
Specific banned categories, a named person to ask, a route to request tools, and a clear statement that reporting mistakes quickly is not misconduct.
The funded route, in figures
Figures are the England position. Levy-paying employers usually draw the whole cost from the levy they already contribute, so they pay nothing extra. Employers below the £3 million pay bill threshold pay nothing for an apprentice aged 16 to 24. For an older apprentice the government funds 95 per cent, and a levy transfer from a larger employer can cover the rest. Eligibility and funding rules are set by the funding body and can change.
Where to go next
If you want data handling and governance assessed as part of a recognised qualification, the funded route covers both.
We help UK employers and employees access government-funded AI training. Training is delivered by third-party providers. We do not guarantee a place, a qualification, a job or any particular outcome. Eligibility and funding rules are set by the funding body and can change. Always check the current position before you commit. See our disclaimer.