How to use AI safely at work

AI tools can save real time at work: drafting, summarising, reorganising, checking. They can also move information somewhere it shouldn't go, produce confident mistakes, and blur who is responsible for the result. None of that is a reason to avoid them. It is a reason to use them with a few clear habits.

This guide sets out those habits for UK workplaces. It covers what information can go where, which versions of the tools to use, how to check what comes back, and who stays accountable. It applies to ChatGPT, Microsoft Copilot, Claude, Gemini and any similar tool.

It is practical guidance, not legal advice. Where the law matters, it explains the idea in plain words and links the official source. Your organisation's policy and approved tools still take priority, and your data protection officer (DPO) or legal adviser has the final say.

The short version

Before you use AI for a work task, ask five questions. They take a few seconds once they become a habit.

1. Data: is it safe to share? What information is in this, and is this tool approved for it?
2. Task: is AI appropriate? Is this a job AI does well, and is it one AI should be doing at all?
3. Source: what should the answer be based on? Your own material, a document you can check, or the tool's general knowledge?
4. Check: how will I verify it? What could be wrong, and how will you find out before anyone relies on it?
5. Responsibility: who is accountable for the final work? Usually you. If you'd be uncomfortable putting your name to it, stop.

Most of this guide covers the first question, where the serious risks sit. The examples come from Fernway Group, a fictional office equipment company used throughout this site.

Why AI at work needs different rules

At home, if you paste your own holiday plans into a chatbot, the only person affected is you. At work, almost everything you handle belongs to someone else or is about someone else: customers, colleagues, suppliers, your employer. You are looking after it on their behalf.

Three things make workplace AI different.

Other people's information is involved. A customer email has a customer in it. A spreadsheet of staff absences has staff in it. When you paste those into an AI tool, you are sharing information about people who never agreed to it being shared that way. In the UK, that brings data protection law into play.

Your employer is responsible for what you do. If an employee puts customer data into an unapproved tool, the organisation usually carries the consequences. That's why many organisations approve specific tools and set rules for them.

Other people rely on your work. A wrong figure in a personal budget is your problem. A wrong figure in a quote, a report or a letter to a customer becomes other people's problem too. AI tools can produce mistakes that read exactly like correct answers.

The UK Government's AI Playbook, written for civil servants, sets out ten principles for using AI safely. Several suit any workplace: know what AI is and what its limitations are, keep meaningful human control at the right stages, use the right tool for the job, and follow your organisation's policies.

Four kinds of information

Most workplace information falls into one of four groups. Knowing which one you're holding answers most of the "can I paste this?" question.

Public. Already published, or meant to be. A product brochure, a published price list, a job advert that's live, a question about how an Excel formula works. Anyone could see it.

Internal. Not secret, but not meant for outsiders. Meeting agendas, internal process notes, draft wording for a staff newsletter. Low harm if it escaped, but it isn't yours to publish.

Confidential. Information that could cause harm or loss if it got out. Unreleased financial results, pricing strategy, contract terms, tender bids, legal advice, anything marked confidential or covered by a confidentiality agreement.

Personal. Anything about an identifiable living person: customers, colleagues, job applicants, pupils, patients, you. Some personal information is especially sensitive (see special category data below).

Information can be in more than one group. A customer complaint is personal (it's about the customer) and may be confidential (it names a contract dispute).

How this fits the AI-Safe Desk bands

The free AI-Safe Desk course uses four bands, and this guide uses the same ones. The labels on this site match them.

• Green (● LOWER RISK): public or general, with nothing about identifiable people or confidential matters. Any tool is fine. You still check the answer.
• Amber (▲ CAUTION): personal or commercially sensitive, but the AI could do the job without those details. Remove them first, or use an approved work tool.
• Red (■ HIGH RISK): special category data, safeguarding, criminal records, legally privileged or restricted material. Never in a consumer tool. Only in a tool your organisation has approved in writing for that kind of information.
• Black (■ HIGH RISK): passwords, keys, security settings, anything that opens a door. Never in any AI tool.

The course lesson on red and black information explains the bands in more depth.

Quick reference: what to do with each type of information

If you're not sure where something sits, the interactive tool Can I put this in ChatGPT? walks you through it in a few questions.

What not to paste into an unapproved AI tool

An "unapproved" tool is any AI tool your organisation hasn't agreed you can use for work. For most people that includes personal accounts on free or paid consumer plans, phone apps, browser extensions, and any tool you signed up to yourself.

Keep these out of unapproved tools:

• names, contact details or account numbers of customers, colleagues or anyone else
• anything about a person's health, beliefs, ethnicity, sexuality, union membership or criminal record
• anything about children or vulnerable adults
• HR matters: absence, performance, disciplinary, grievances, pay, recruitment decisions
• financial information that isn't public: results, forecasts, margins, bank details
• contracts, tenders, legal advice and anything marked confidential
• documents other organisations shared with you in confidence
• passwords, access codes, keys, system settings and security information
• screenshots or photos of any of the above

Most of the time you can still get help. Describe the situation instead of pasting the document, or remove the identifying details first. The guide How to anonymise a document before using AI explains how, and where that approach stops working.

Consumer, business and internal tools: same name, different terms

The same product name can cover very different arrangements. If you understand one thing about AI at work, make it this.

• Consumer versions are for individuals: free accounts and personal paid plans. They run on consumer terms between you and the provider. Your employer has no agreement covering them.
• Business and enterprise versions are bought by an organisation. They come with a contract, usually including limits on training and retention and a data processing agreement.
• Internal deployments run inside the organisation's own systems, set up by its IT team, under its own rules.

The screen can look almost identical in all three. Check which account you are signed in with before you paste anything that isn't public. The AI-Safe Desk lesson Consumer, business, internal covers this in more detail.

Here is what the main providers say, in brief. Details change, so follow the links for the current position.

At Fernway, the company provides Microsoft 365 Copilot Chat through staff work accounts. Leah Bennett, the operations assistant, also has a personal ChatGPT account on her phone. Same person, same task, very different position: the Copilot session is covered by Fernway's Microsoft 365 agreement; the phone app is covered by Leah's own consumer terms and nothing else.

Personal data

Personal data is any information about an identified or identifiable living person. The ICO's guide What is personal data? explains that people can be identified directly (a name, an email address) or indirectly, from other details that point to them.

That second part catches people out. A message that says "the account manager for the Midlands who started in March" contains no name, but at Fernway it can only be Alex Doyle. Removing the name didn't make it anonymous.

At work it turns up everywhere: customer contacts, staff records, CVs, complaint logs, meeting transcripts, photos, and almost every email thread, because of the names and signatures.

Using personal data in an AI tool is a form of "processing" under UK data protection law. Your organisation needs a lawful reason for it and has to keep it secure. The ICO's guidance on AI and data protection sets out what organisations must consider. For you as an individual, the practical rule is simple: personal data goes only into tools your organisation has approved for it, and only as much as the task needs.

Special category data

Some personal data is more sensitive, and the law gives it extra protection. 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 to identify someone
• health data
• data about a person's sex life
• data about a person's sexual orientation

Criminal offence data (convictions, offences, related security measures) is a separate category with similar safeguards. Treat it the same way.

These are red-band information. They don't go into a consumer AI tool, even with the name removed, because in a small workplace the details often identify the person on their own. They go into an AI tool only if your organisation has approved that specific tool for that kind of information, in writing, and you've read what it says.

At Fernway, Maya Roberts, the office manager, receives a colleague's fit note and a letter from their GP about workplace adjustments. She'd like help drafting a supportive reply. This is health data about an identifiable colleague. It stays out of AI tools unless Fernway has a written policy approving a specific tool for it, which it doesn't. Maya writes the reply herself, and if she wants help with tone, she asks a general question with no details: "How do I write a warm, professional reply to an employee who has asked for workplace adjustments?"

Confidential business information

Not everything sensitive is about people. Some information matters because of the harm to the business if it got out: financial results before they're announced, margins, pricing strategy, bids, acquisitions, legal disputes, product plans, security arrangements.

A quick test: would you hesitate to email it to someone outside the company? If so, it belongs in an approved work tool at most, and only if your policy permits that kind of material there.

At Fernway, Dan Okafor in Finance is preparing the half-year figures for the board. He'd like help turning a table of numbers into a short narrative. Before the results are shared, they're confidential. He uses Fernway's approved Copilot, signed in with his work account, which Fernway's policy allows for internal finance documents. He would not use a personal account for this, whatever its settings.

Client and customer information

Customer information combines both risks: it's usually personal data, and it's often commercially confidential too. Contracts, account histories, pricing agreements, complaints and correspondence all fall here.

There's a third layer. Many contracts include confidentiality clauses that limit how you can use or share the customer's information. Sharing it with an AI provider your organisation hasn't approved could breach the contract as well as data protection rules.

So use your organisation's approved work tool for customer material. If you only need help with tone or structure, describe the situation without the customer's details. Don't upload whole contracts or account exports to a personal account, however useful a summary would be. And if a customer has told you something in confidence, treat it as red.

Employee information

Information about colleagues and job applicants is personal data, and much of it is sensitive: pay, performance, absence, disciplinary matters, grievances, health, family circumstances. Employees reasonably expect it to be handled carefully.

At Fernway, Nina Carter, a part-time HR adviser, is shortlisting applicants for a new account manager role. She's tempted to paste the CVs into a chatbot and ask it to rank them. There are three problems. The CVs are personal data. A ranking is a decision about people (see below). And an unapproved tool gives Fernway no agreement about what happens to those CVs.

A better approach: Nina uses AI with no candidate data at all, to turn the job description into a clear scoring grid. Then she and Tom Elliott, the sales lead, score each application themselves against the grid.

Here is a job description for an account manager role at an office equipment company: [paste the published job description]. Turn the essential and desirable criteria into a scoring grid I can use to assess applications consistently. For each criterion, give a short description of what weak, adequate and strong evidence would look like in an application. Don't include anything about age, health, family circumstances or other personal characteristics.

Schools, health and safeguarding

Some sectors carry extra responsibilities. Even if you don't work in one, you may handle their information as a supplier.

Schools. The Department for Education's guidance on generative AI in education says: "It is recommended that personal data is not used in generative AI tools." Pupil information involves children, so treat it as red. Report writing, behaviour notes and special educational needs information stay out of unapproved tools. A teacher can use AI for lesson planning or adapting a worksheet with no pupil data in it.

Health. Health data is special category data wherever it appears: a patient letter, an occupational health report, a note that a colleague is off with a particular condition. None of it goes into an unapproved tool, with or without a name.

Safeguarding. Safeguarding information, about a child or an adult at risk, never goes into an unapproved AI tool. Not to summarise, not to help write it up, not with the names removed. It follows your organisation's safeguarding procedure and goes to the designated safeguarding lead.

At Fernway, an engineer servicing the photocopier at Larkstone Primary School (a fictional customer) overhears something that worries him about a pupil's welfare. He tells the school's designated safeguarding lead before he leaves, as Fernway's procedure requires, and tells Priya Shah, Fernway's head of operations. He writes his note of what he heard himself, in his own words, the same day. It doesn't go near an AI tool, not even to tidy up the wording.

Other people's documents and intellectual property

Documents you receive at work often belong to someone else: a supplier's proposal, a partner's report, a paid-for industry study, a training manual, a book chapter. Two things to consider before you paste them into an AI tool.

Confidentiality. If the document was shared under a confidentiality agreement or marked confidential, uploading it to an AI tool may break that agreement. Check the terms, or ask the person who manages the relationship.

Copyright. Uploading someone else's material and asking an AI to rewrite it or reproduce large parts of it can raise copyright questions. This is a complex area and this guide can't give legal advice. The cautious approach: use AI to help you understand or summarise material for your own work, credit the original, and don't publish AI rewrites of other people's content as your own. If you're unsure, ask your manager or legal team.

The same applies in reverse. If you publish something AI helped you write, check it doesn't reproduce someone else's words closely.

At Fernway, Tom Elliott receives a competitor comparison from a paid research service, marked "for subscriber use only". He'd like AI to pull out the three points most relevant to Fernway. He checks the subscription terms first. They restrict sharing the report with third parties, so he reads the relevant section himself and writes his own notes.

Is AI right for this task?

Even when the information is safe to share, AI isn't always the right tool. It does well at drafting, restructuring, summarising material you give it, suggesting wording and spotting gaps. It does badly at anything that depends on facts it can't see, on recent or local knowledge, or on judgement about people.

Skip AI, or keep its role small, when:

• the task is quicker by hand (a two-line reply, a single figure)
• the task is making a decision about a person
• the work has to be your own, such as an assessment or a professional opinion
• you couldn't tell whether the answer was right

What should the answer be based on?

An AI tool answering from its general knowledge is guessing at what's likely. An AI tool working from a document you've given it is much easier to check. Where you can, give it the source and tell it to use only that: "Using only the attached returns policy, draft a reply that explains the customer's options." Then ask it to say when the answer isn't in the source, instead of filling the gap.

If the task needs facts you haven't provided, such as current prices, legal requirements or someone's contact details, get them from the proper source yourself.

Checking what comes back

AI tools predict likely-sounding text. They don't look things up the way you would. Most of the time that works well; sometimes it produces names, figures, dates, quotes or sources that are wrong or don't exist, written with exactly the same confidence as the correct parts.

Before anything AI produced leaves your hands:

• Check the specifics. Every name, number, date and quotation against the original source.
• Check the sources. If it cites a report, a law or a web page, open it and confirm it exists and says what the AI claims.
• Check the premise. Did your question assume something that isn't true? AI often goes along with a wrong assumption.
• Check what's missing. Summaries can drop the one caveat that mattered.
• Match the effort to the stakes. A reworded internal email needs a read-through. A figure going to a customer needs checking line by line.

At Fernway, Grace Lynch asks Copilot to summarise the payment terms from three supplier contracts. The summary says one supplier offers 60 days. The contract says 30 days, with 60 available only for orders over a set value. Grace catches it because she checks every figure against the contract before sending the summary to Dan.

The guide How to check whether an AI answer is right sets out a five-check routine, and the free Prove It course teaches it in depth. To practise spotting errors, try Spot the Hallucination.

AI and important decisions about people

Some decisions affect people's lives: hiring, promotion, discipline, dismissal, pay, access to a service, credit, grades, eligibility. AI can help organise information for those decisions. It shouldn't make them.

What that means in practice:

• Keep a human decision-maker. A named person makes the decision, having looked at the evidence themselves, not just the AI's summary or score.
• Be able to explain it. If someone asks why a decision went against them, you should be able to explain the reasons in plain language. "The AI ranked you lower" is not an explanation.
• Watch for bias. AI tools can reflect patterns in the material they learned from, including unfair ones.

UK data protection law also gives people rights where decisions about them are made solely by automated means with significant effects. The ICO's guidance on explaining decisions made with AI sets out what organisations need to consider. If your team is thinking about using AI anywhere near decisions like these, involve HR and your data protection contact before you start.

Human accountability

If AI drafted it and you sent it, you are responsible for it. It's the same standard that applies to a template or a colleague's draft.

In practice:

• Read everything before it goes out under your name.
• Don't forward AI output you don't understand.
• Don't present AI-generated analysis as your professional judgement unless you've done the thinking to stand behind it.
• Be open about AI use where people would reasonably expect to know. The guide Should I tell my employer I used AI? covers this.

Keeping records where needed

Most everyday AI use needs no special record. A reworded email is just an email.

Some work does need a trail: anything that feeds a decision about a person, a figure in a report others rely on, regulated work, or anything that might later be questioned. For that work, keep a brief note of:

• which tool and account you used
• what you asked it to do (the prompt, or a summary of it)
• what you checked, and against what
• what you changed

A few lines in the file or the email is usually enough. If someone later asks "did you check that?", you can say yes, and show how. The Prove It lesson on keeping a trail shows a simple format.

Your organisation's policy

Many organisations now have a written AI policy, or AI rules inside a wider IT or data policy. It might name approved tools, ban personal accounts for work, list information that must never go into AI, or require you to label AI-assisted work.

Find it and read it. Look on the intranet or in the staff handbook, or ask your manager or IT team. If it says something different from this guide, the policy wins.

If there isn't one, don't take silence as permission. Treat anything beyond green as off limits in unapproved tools until you've asked. If you're a manager wondering where to start, the AI policy starter produces a first draft to discuss, and the guides for managers cover rolling AI out to a team.

When you're not sure

Stop before you paste. A few minutes' delay costs very little.

Ask the right person. Your manager for "is this OK in our team?"; your data protection officer or data protection contact for personal data; IT or the security team for which tools and accounts are approved. Asking is a sign of good judgement.

Use the decision tool. Can I put this in ChatGPT? takes you through the main questions and gives a lower risk, caution or high risk result.

If something has already gone in

Mistakes happen, usually when people are busy. If something went into an AI tool that shouldn't have:

1. Stop. Don't paste anything more.
2. Tell someone quickly. Your manager and your data protection contact, the same day. Say what went in, which tool and account, and when.
3. Don't try to fix it quietly. Deleting the chat may not remove every copy the provider holds, and your organisation may still need to assess what happened.

Speed matters more than blame. Under UK data protection law, an organisation must report a notifiable personal data breach to the ICO without undue delay and no later than 72 hours after becoming aware of it, as the ICO's guide to personal data breaches explains. Deciding whether a breach is notifiable is your organisation's job, not yours. Your job is to tell them quickly, because the sooner your organisation knows, the more options it has.

Before you use AI: a checklist

Run through this before any work task that involves AI. Most of the time every box ticks in seconds. When one doesn't, that's your signal to stop and ask.

Finish one complete safe task

Choose an internal draft based on approved, non-sensitive material. Check the important claims, record the source and keep a human decision before sharing. Note the point where you would stop if the task became sensitive or consequential.