What Should My Staff Be Allowed to Put Into ChatGPT, Copilot or Gemini at Work?

10 August 2026

Why This Actually Needs Writing Down

If you think your team is not using AI tools at work, you are almost certainly wrong. Microsoft's UK research found that 71% of employees have used unapproved consumer AI tools at work, and 51% do so every week. A separate Okta and Apprize360 survey found 55% of UK employees report using unsanctioned tools, even though 96% of executives at the same organisations said they were confident they had visibility over AI use. That gap between what leaders think is happening and what is actually happening is the whole problem.

Most of this is not malicious. Someone is behind on a report, pastes a client email thread into ChatGPT to get a faster summary, and moves on with their day. Nobody flags it because nobody told them not to. The tool did exactly what it was designed to do. The business just lost visibility and, potentially, control over data it was legally responsible for.

Formal governance has not kept pace. CIPD research puts the proportion of UK employers with a written AI usage policy at just 31%, and SAP found that 60% of UK businesses admit their staff have not completed comprehensive AI training, even as AI investment is set to rise by 40% over the next two years. You are not behind if you do not have a policy yet. Most businesses do not. But you are exposed until you write one, because 'use common sense' is not something you can point to if something goes wrong.

The Traffic Light Rule for AI Inputs

The simplest working policy is a traffic light system. It does not need to be a 40-page document. It needs to be short enough that staff actually remember it.

Green (fine to use freely, even on free-tier tools): generic writing help, brainstorming, summarising publicly available information, drafting templates, coding help on non-proprietary logic, learning and research questions.

Amber (only on approved business-tier accounts, and only with names/identifiers removed where possible): internal process documents, anonymised or aggregated business data, draft contracts before client names are added, general strategy thinking that does not reference named clients or figures tied to a real person.

Red (never, on any tier, without specific written sign-off): client names combined with any personal detail, financial records, passwords or credentials, health or biometric information, anything covered by an NDA, unpublished financial results, HR or disciplinary information, and anything a client would reasonably expect you to keep confidential.

The genuinely useful trick for the amber category is asking staff to substitute real names and figures with placeholders before pasting anything in. 'Client A, based in the Midlands, spends roughly £X per month' does the same job for getting AI help without creating an identifiable record anywhere outside your systems.

Free Tools vs Paid Business Tiers: The Real Difference

This is where most small businesses get confused. A free ChatGPT, Gemini or Copilot account is not the same product as ChatGPT Team, Microsoft 365 Copilot, or Gemini for Google Workspace, even though the interface looks similar.

Free consumer accounts generally allow the provider to use your conversations to improve their models unless you manually turn that setting off, and even then you are relying on a consumer terms-of-service agreement rather than a business data processing contract. There is no admin oversight, no audit trail, and typically no ability for you as the employer to see what has been typed in or delete it centrally.

Business and enterprise tiers change the contract, not just the interface. ChatGPT Team and Enterprise, Microsoft 365 Copilot (with a proper licence), and Gemini for Workspace include data processing terms that state your inputs are not used to train the underlying model, plus admin controls, retention settings, and in some cases audit logs. That is a meaningfully different risk position under UK GDPR, because you have a documented lawful basis and can show due diligence if the ICO ever asks.

But the licence alone does not do the work. You still need someone to actually configure the admin settings, still need a data processing agreement on file, and still need a DPIA if the use case is high risk. Buying the paid tier and assuming that solves the problem is one of the most common mistakes small businesses make.

What Happens If It Goes Wrong

If an employee pastes personal data (a client's name, address, health information, or financial details) into an AI tool without a lawful basis, that can constitute a personal data breach under UK GDPR. The ICO's rule is clear: you must report a notifiable breach to the ICO without undue delay and, where feasible, within 72 hours of becoming aware of it, if it is likely to result in a risk to people's rights and freedoms. Separately, and with a higher threshold, you may need to tell the individuals affected directly if the risk to them is high.

In practice, if an employee flags that they have pasted something sensitive, the immediate steps are: try to delete the input from the chat if the platform allows it, document exactly what was entered and when, assess whether it meets the ICO's reporting threshold, and check whether your data processing agreement with the AI provider (if you have a business tier) gives you any additional recourse. Do this calmly. Punishing the person who reported it teaches everyone else to stay quiet next time, which is a far bigger long-term risk than the individual incident.

You should also know when a formal Data Protection Impact Assessment (DPIA) is required. The ICO's own guidance states a DPIA is mandatory for processing likely to result in high risk to individuals, and specifically flags 'innovative technology, including AI' as a trigger factor when combined with other risk criteria. A simple writing assistant that never touches personal data may not need one. An AI tool connected to your CRM or customer records almost certainly does.

Is This Right For You?

This matters most if your team already uses AI tools day to day and you have no written rule about it, which is most UK small businesses. Research from Microsoft found 71% of UK employees have used unapproved consumer AI tools at work, and separate research from Okta found 55% continue to use tools their employer has not sanctioned. Meanwhile CIPD research puts the number of employers with a formal AI usage policy at just 31%. That gap between usage and governance is exactly where the risk sits.

It is not right for you if you genuinely handle no personal or commercial data at all, which in practice is almost nobody. Sole traders with no clients, no staff records and no supplier contracts have less to worry about, but the moment you have a customer list, an invoice, or an email inbox, this applies to you.

It is also not a substitute for proper legal advice if you operate in a regulated sector (financial services, healthcare, legal, education). Those sectors have additional obligations on top of what is covered here, and you should get sector-specific guidance rather than relying on a general small business framework.

Frequently Asked Questions

Do I need a DPIA before letting staff use ChatGPT?

Not automatically. The ICO requires a DPIA when processing is likely to result in high risk to individuals, and flags AI as a trigger factor when combined with other risk criteria such as large-scale processing or automated decision-making. A generic writing assistant that never touches personal data usually will not need one. An AI tool connected to your customer database, CRM, or HR records almost certainly will.

What should I do if an employee has already pasted sensitive data into ChatGPT?

Stay calm and act quickly rather than punishing the person who told you. Try to delete the input from the chat history if the platform allows it, write down exactly what was entered and when, and assess against the ICO's threshold: is this likely to risk someone's rights and freedoms? If yes, you have 72 hours from becoming aware of it to report to the ICO. If you are on a business-tier account, check what your data processing agreement with the provider gives you.

Is Microsoft Copilot or Google Gemini safer than ChatGPT for business use?

The safety difference is mostly about which licence tier you are on, not which brand you pick. A free consumer account for any of the three carries similar risk. Microsoft 365 Copilot with a proper business licence, ChatGPT Team or Enterprise, and Gemini for Google Workspace all include broadly comparable data processing terms and admin controls. Choose based on what fits your existing software stack, then make sure the admin settings are actually configured correctly.

Can I just block AI tools instead of writing a policy?

You can try, but the evidence suggests it will not work. Even in organisations that formally block AI tools, research shows a large proportion of staff continue using them on personal devices or personal accounts, simply moving the activity somewhere you cannot see it at all. A written policy with clear green, amber and red categories, backed by an approved business-tier tool staff are actually allowed to use, is far more effective than a blanket ban that pushes usage underground.

Do I need to tell clients I am using AI tools in my business?

It depends on the use case. If AI is used for internal drafting or research that never touches client-identifiable data, disclosure is good practice but rarely a legal requirement. If AI is directly involved in customer-facing decisions, automated responses, or processing a client's personal data, transparency becomes more important, and in some cases expected under UK GDPR's fairness principle. When in doubt, a short line in your engagement terms covering how you use AI tools is a low-cost way to stay ahead of the question.