Using ChatGPT and AI tools safely at work: UK SME guide (2026)

Your staff are probably already using AI. They use it to tidy emails, summarise documents, and take meeting notes. An NCSC blog on shadow AI cites a study where 71% of employees said they used AI tools their employer had not approved.

That is not a reason to ban AI. It is a reason to set a few simple rules. This guide is for owners with 5 to 50 staff and no IT team. It reduces risk. It does not make any tool risk-free.

The one rule: think before you paste

What you type into an AI chat tool goes to the company that runs it and is stored there. Depending on the account and settings, its staff may review it and it may be used to improve future models.

The NCSC's advice is simple: keep sensitive information out of public AI tools, and do not ask anything that would cause problems if made public. See its blog ChatGPT and large language models: what's the risk?

What not to paste into AI tools

• Client personal data: names with addresses, dates of birth, National Insurance numbers, health details, or anything about a named person's money or family
• Staff personal data: sickness, disciplinary notes, salaries, or HR complaints
• Passwords, MFA codes, API keys, or bank login details. Never, in any tool
• Contracts, quotes, and anything covered by an NDA or client confidentiality terms
• Unpublished financials: management accounts, bank statements, sale or funding plans

Strip details out first. Replace names with "Client A". Remove account numbers. You usually get just as good an answer.

UK GDPR in plain terms

If staff paste personal data about a client or colleague into an AI tool, your business is processing personal data. UK GDPR applies, and your business stays responsible for it. In practice:

• Use only what you need. If the task works without names, leave them out.
• Be open. Your privacy notice should cover how you use people's data, including in AI tools. The ICO explains this in the right to be informed.
• Use tools with proper terms. Business plans usually come with data protection terms. Free personal accounts usually do not.
• Know what to do when it goes wrong. Client data in the wrong tool can be a personal data breach. Some breaches must be reported to the ICO within 72 hours of you becoming aware.

The ICO has a hub for small organisations and more detailed guidance on AI and data protection. This guide is general information, not legal advice.

Free personal accounts vs business accounts

A personal account belongs to the person, not the firm. You cannot see what was shared, set the privacy settings, or remove access when they leave. It is the same problem as shared and personal logins.

Business plans give you admin control and business data terms. Pick the one that matches what you already use:

• Microsoft 365 firms: Copilot signed in with staff work accounts.
• Google Workspace firms: Gemini signed in with Workspace accounts.
• Either: ChatGPT Business or Enterprise.

Keep client work in accounts the firm controls. Until you have one, keep client data out of AI tools.

Check whether the tool trains on your inputs

We checked these against each vendor's help pages in October 2026. Menus change often, so check again.

ChatGPT

• Personal plans (Free, Plus, Pro): OpenAI says training is on by default. To turn it off, open your profile, then Settings, then Data controls, and switch off Improve the model for everyone. It applies across your devices.
• Temporary Chat: OpenAI says these chats are not used for training and are deleted after 30 days.
• Feedback: OpenAI says rating a reply with a thumbs up or down may let it train on that whole conversation, even if you opted out.
• ChatGPT Business and Enterprise: OpenAI says it does not train on business inputs or outputs by default.

Microsoft Copilot

• Work account: Microsoft says prompts, responses, and data Copilot reaches through your Microsoft 365 account are not used to train its foundation models. Check staff are signed in with their work account, not a personal one.
• Personal Microsoft account: Microsoft released a new Copilot app on 18 August 2026. For the new app, Microsoft says prompts, responses, and file contents are not used to train foundation models. In the older app, look under Privacy for Training on conversation activity and Training on voice conversations, and switch both off.

Google Gemini

• Personal Google account: Google says chats may be read by human reviewers and used to improve its AI while Keep activity is on, and asks users not to enter confidential information. Turn it off at gemini.google.com under Activity. Google says it still keeps chats for 72 hours and may review them for safety.
• Workspace account: Google says your content is not reviewed by humans or used to train models outside your organisation without permission.

For any other tool, search its help pages for "training" and "data controls". If you cannot find a clear answer, keep client data out of it.

AI meeting notetakers: tell people first

Notetakers record a call, transcribe it, and write a summary. The recording is personal data about everyone on the call.

• Tell people at the start that you are recording and why. Let them object.
• Do not record HR, health, legal, or other sensitive conversations.
• Switch off auto-join, so a bot does not join meetings you are not in.
• Check who receives the summary. Some tools can send it to every attendee.
• Set a retention period and delete recordings you no longer need.

If a client asks you not to record, do not record. It is their data too.

AI keyboards and voice apps

Keyboard apps can see everything typed through them, including client messages and sometimes passwords. Voice apps may send audio to the cloud to turn it into text.

Before staff install one on a work phone, check who makes it, whether speech is processed on the phone or in the cloud, and what permissions it asks for. If in doubt, stick with the keyboard that came with the phone.

Some newer apps do the processing on your own phone or computer instead. That can reduce how much leaves the device. It does not change your duty to tell people before you record them.

Lock down AI accounts like email

An AI account holds a history of everything staff asked it. Anyone who takes it over can read that. Treat it like email.

• Use a unique password for every AI account, stored in a password manager. See how to stop staff reusing passwords.
• Turn on MFA. Our guides on MFA for Microsoft 365 and authenticator apps for staff walk you through it.
• Connect AI tools to email, files, or calendars only when there is a clear need.
• Remove AI access on the same day someone leaves.

A one-page staff AI policy you can copy

Edit the parts in square brackets. Keep it to one page.

1. Approved tools: [list tools]. Use them only with your work account.
2. Never paste: client or staff personal data, passwords or codes, bank details, contracts, or unpublished financials.
3. Anonymise first: replace names and identifying details before you ask.
4. Settings: [name] checks training settings on every approved tool.
5. Check the output: AI makes mistakes. A person checks facts, figures, and tone before anything goes to a client.
6. People decisions: AI does not decide hiring, pay, or discipline. A person makes the final call.
7. Recording: use only [approved notetaker]. Tell everyone first. Never record sensitive calls.
8. New apps: ask [name] before installing AI apps, browser extensions, or keyboards on work devices.
9. Accounts: unique password from the password manager, and MFA on.
10. Mistakes: if you share something you should not have, tell [name] straight away. No blame. We may have 72 hours to act.

Policy owner: [name]. Review date: [date, at least yearly].

FAQ

Is it safe to put client data into ChatGPT?

Not into a free or personal account. Keep client personal data, contracts, and financials out unless the firm has a business account with suitable terms. Even then, share only what the task needs.

Does ChatGPT use my chats to train its models?

On personal plans, OpenAI says chats can be used for training unless you turn off Improve the model for everyone in Settings, then Data controls. OpenAI says it does not train on ChatGPT Business or Enterprise data by default.

Do we need permission to use an AI notetaker in meetings?

You need to tell people before you record, explain why, and let them object. The right legal basis depends on the situation, so check the ICO guidance. Asking at the start of the call is the simplest habit.

Is Microsoft Copilot safer than ChatGPT for business?

It depends on the account, not the brand. A work account on a business plan comes with business terms and admin control. A personal account does not.

What should we do if someone pasted client data into an AI tool?

Note what was shared, in which tool, and when. Delete the chat. Then decide whether it is a personal data breach. Some breaches must be reported to the ICO within 72 hours.

Sources we checked

• NCSC: The hidden risks of shadow AI
• NCSC: ChatGPT and large language models: what's the risk?
• NCSC: AI and cyber security: what you need to know
• ICO: Guidance on AI and data protection
• OpenAI, Microsoft, and Google help pages on data controls, checked October 2026

Bottom line

Your staff will use AI anyway. Give them an approved business account, a short list of things never to paste, and training settings already checked. Tell people before you record them. Protect AI accounts with unique passwords and MFA. Write it on one page and review it yearly.