What can I put into ChatGPT at work?

The honest answer is that it depends on two things: what the information is, and which version of the tool you're using. Both take seconds to check once you know how.

This guide explains how to decide. It uses ChatGPT in the title because that's what most people ask about, but everything here applies equally to Microsoft Copilot, Claude, Gemini and any other AI tool.

It's practical guidance, not legal advice. Your organisation's policy and approved tools still take priority. Where the law is involved, your data protection officer (DPO) or legal adviser has the final say.

Two questions before you paste

What is this information? Is it public? Is it about a person? Is it confidential to your employer, a customer or a supplier? Is it especially sensitive?

Which version of the tool am I in? A personal account on a free or paid consumer plan, a business version your employer has licensed, or an internal tool your IT team runs?

The same piece of information can be fine in one version and wrong in another. A customer's complaint email might be allowed in your employer's licensed Copilot and not in your personal ChatGPT app, even though the task is identical.

Public or confidential?

Public information is already published or meant to be: your company's website, a published price list, a press release, a live job advert, a general question like "how do I write a polite chaser email?". It can go into any tool. You still check the answer.

Confidential information would cause harm or loss if it got out: unreleased financial figures, pricing strategy, tender bids, contract terms, legal advice, product plans, anything marked confidential, and anything another organisation shared with you in confidence. It doesn't go into a personal account. It may go into an approved work tool if your policy allows that kind of material there.

In between sits ordinary internal material: meeting agendas, process notes, draft newsletters. Low harm if seen, but not yours to publish. An approved tool is the right home for it; in a personal account, strip anything specific first.

Personal data

Personal data is any information about an identified or identifiable living person. It includes obvious things like names, email addresses, phone numbers and account numbers. It also includes details that identify someone indirectly, when combined with other information. The ICO explains this in What is personal data?

At work, personal data turns up everywhere: customer emails, CVs, staff rotas, meeting notes, complaint logs, email signatures, screenshots and transcripts.

Putting personal data into an AI tool counts as using ("processing") it under UK data protection law. Your organisation needs a proper basis for that and has to keep the data secure. That's why the practical rule is: personal data goes only into tools your organisation has approved for it, and only as much as the task needs. In a personal account, remove it first or describe the situation instead.

Special category data

Some personal data is more sensitive. The ICO lists nine types of special category data: racial or ethnic origin, political opinions, religious or philosophical beliefs, trade union membership, genetic data, biometric data used for identification, health, sex life, and sexual orientation. Criminal offence data is a separate category with similar safeguards.

Add to that safeguarding information about children or adults at risk, and legally privileged material.

This is high-risk information. It never goes into a consumer AI tool, with or without names, and it goes into a work tool only if your organisation has approved that tool for that kind of information in writing. If you work in a school, note that the Department for Education's guidance on generative AI in education recommends that personal data is not used in generative AI tools at all.

The bands and the decision tool's results

The free AI-Safe Desk course sorts information into four bands. The decision tool gives three results. They line up like this.

Most everyday work is green. The mistake is assuming everything is.

What "approved" means

An approved tool is one your organisation has agreed you can use for work, usually because it has a contract with the provider covering how data is handled. Three versions of the same product can sit behind one familiar name.

• Consumer: a free or personal paid account. Consumer terms, between you and the provider. No agreement with your employer.
• Business or enterprise: licensed by your organisation, with a contract that usually limits training and retention and includes a data processing agreement.
• Internal: run inside your organisation's own systems under its own rules.

The screens can look almost identical. Check which account you are signed in with every time. Microsoft's work and personal Copilot, for example, can appear in the same places, and the protections depend on signing in with your work account.

Being approved doesn't mean anything goes. Your policy may approve a tool for internal documents and customer emails but not for health information or HR cases. Read what it's approved for.

At Fernway Group, a fictional office equipment company used across this site, staff have Microsoft 365 Copilot Chat through their work accounts. Fernway's policy allows customer emails and internal documents there. It doesn't allow staff health information or anything about disciplinary cases in any AI tool. When Alex Doyle, an account manager, wants help replying to a tricky customer, he uses Copilot with his work account. On his personal phone app, he'd describe the situation without the customer's details.

Anonymisation and pseudonymisation

These two words sound similar and mean very different things.

Anonymised information can no longer identify anyone, taking into account everything reasonably likely to be used to identify them. Truly anonymous information isn't personal data.

Pseudonymised information has had direct identifiers replaced, for example names swapped for codes like "Customer A" or "Employee 14", but someone could still link it back to the person, because the key exists or other details give it away. The ICO's anonymisation guidance is clear that pseudonymised data is still personal data.

Most "redaction" people do before using AI is pseudonymisation. That's useful: it reduces risk, and for amber information it's often enough. It isn't the same as making the information anonymous, and it doesn't turn red information into green.

The ICO's guidance judges anonymisation by asking whether someone could identify a person, including a "motivated intruder": a reasonably competent person who wants to identify someone and uses information that's easy to get hold of. Colleagues, neighbours and customers often know far more than you'd expect.

Why removing a name isn't always enough

Here are three situations where the name is gone and the person is still obvious.

A small team, a job title and a date. "The finance assistant who was off sick for two weeks in March." At Fernway there's one finance assistant: Grace Lynch. Anyone at Fernway, and quite a few customers, would know who it means. The sentence also contains health information.

A rare condition in a small school. "A Year 4 pupil with a rare genetic condition who uses a wheelchair." In a school with one class per year group, that describes one child. Parents at the school gate would recognise them. It's also health data about a child, so it's high risk however it's worded.

A unique incident. "The delivery driver who reversed into the customer's gatepost in Banbury last Tuesday." There was one such incident. The customer, the driver's colleagues and the insurer would all know who it is, and the details would be easy to match to internal records.

In each case, the combination of details does the identifying. The guide How to anonymise a document before using AI shows how to generalise details like these, and when to stop trying and use a different approach.

Training and retention settings

Every major AI tool has settings that change what happens to your conversations: whether they're saved, whether they may be used to improve the model, and how long the provider keeps them. They're worth setting on purpose. They change often, so check the provider's own page rather than relying on a summary.

• ChatGPT: consumer users can turn off training in data controls. Temporary Chats don't appear in your history and aren't used for training, but OpenAI keeps them for up to 30 days for safety. OpenAI says it doesn't train on ChatGPT Business, Enterprise, Edu or ChatGPT for Healthcare workspace content by default. See OpenAI's data controls FAQ.
• Microsoft Copilot: signed in with a work account under your organisation's Microsoft 365 terms, enterprise data protection applies, prompts and responses aren't used to train foundation models, and your organisation's permissions and sensitivity labels apply. See enterprise data protection.
• Claude: on consumer plans, chats are used to improve Claude only if you allow it in privacy settings, or if flagged for safety review. Incognito chats aren't used for training. Business products are under separate commercial terms. See Anthropic's page on model training.
• Gemini: in the Gemini app, a subset of chats is reviewed by trained human reviewers and reviewed chats are kept for up to three years, disconnected from your account. "Keep Activity" controls whether chats are saved and used for improvement. Work or school accounts may be under different terms. See the Gemini Apps Privacy Hub.

The AI-Safe Desk lesson The settings that matter walks you through finding and setting them.

Settings help, and they're worth changing. They don't change which band information is in. A red document is red whatever your settings.

"ChatGPT says it's private" is not a policy

People sometimes reason like this: "The app says my chats aren't used for training, so it's fine to use it for work." Or: "I asked it to keep this confidential and it said it would."

Neither is a policy. Three reasons:

• A provider's setting is a promise to you, not your employer. Your organisation has no contract covering your personal account, no say over how long data is kept, and no way to respond if something goes wrong.
• Training is only one of the things that can happen to your text. It's also sent, processed, possibly stored, and possibly reviewed. The AI-Safe Desk lesson The four things a tool can do with your text explains all four.
• The AI can't make promises about its own provider's systems. If you ask a chatbot to keep something private, it may say it will. That reply is generated text, not a commitment. Once you've sent it, it's sent.

The only people who can decide what's allowed at your work are your organisation: its policy, its DPO, its IT and security team.

Your organisation's rules come first

Organisations differ. One might approve Copilot for everything except HR material. Another might ban all AI tools for client work while it reviews contracts. A school might allow AI for lesson planning and nothing involving pupils. A regulated firm might require AI-assisted work to be labelled.

So before relying on any general guide, including this one:

• Find your organisation's AI policy, or the AI section of its IT or data policy.
• Check which tools are approved, and approved for what.
• Check whether there are rules on labelling AI-assisted work or keeping records.
• If there's no policy, treat anything beyond green as off limits in unapproved tools until you've asked.

When to stop and ask

Stop before you paste, and ask, if:

• you can't tell which band the information is in
• it's about a person and you're not sure the tool is approved for personal data
• it includes anything about health, children, safeguarding, criminal matters or HR cases
• the document belongs to another organisation or is covered by a confidentiality agreement
• you're not sure which account you're signed into
• the policy doesn't mention the tool or the task

Who to ask: your manager for team practice; your data protection officer or data protection contact for personal data; your IT or security team for which tools and accounts are approved.

If you've already pasted something that shouldn't have gone in, tell your manager or data protection contact straight away. Your organisation may have to report some breaches to the ICO within a short time, and it can only do that if it knows, so speed matters more than blame. How to use AI safely at work covers what to do.

Practise: which result would you give?

Decide for each, then open the answer.

For your own real examples, use the decision tool.

Check the destination as well as the input

Confirm the account and workspace before pasting. Then consider attachments, connected sources and the people who will receive the output. Approval for one use or tool does not automatically cover every step of a new workflow.