Using ChatGPT, Claude and Copilot Safely in a Small Business

Generative AI tools such as ChatGPT, Claude and Microsoft Copilot can help with routine work: drafting emails, summarising documents, suggesting spreadsheet formulas and producing first versions of policies. They also bring risks, most of which can be managed through clear habits and sensible decisions about which tools staff use and how.

Many of the controls below are manageable in-house. Get advice where the risks warrant it, for example on contracts, permissions reviews or data protection impact assessments.

Where the risks are

A major risk is someone pasting something into a chat box that should never have left the business: a customer's details and complaint history, a supplier's confidential contract, staff sickness records or a password from a shared document. Once it is in a tool you do not control, you may not be able to get it back. If it is personal data, you remain responsible for it under UK data protection law.

Staff therefore need to know what to do if it happens. They should report it straight away. Then change any exposed passwords or API keys, contact the provider and assess whether there has been a personal data breach. Where a breach must be reported, tell the ICO within 72 hours of becoming aware of it. If it is likely to result in a high risk to people, tell them without undue delay.

The answers themselves are the second problem. These tools write fluently whether they are right or wrong, and can invent facts, misstate the law, get figures wrong and cite cases or regulations that do not exist.

Then there is ownership. UK copyright law already has a provision for certain computer-generated works, but whether a particular AI output qualifies, and whether it reproduces someone else's work, can be uncertain. A provider's terms cannot guarantee that an output is protected by copyright or free of third-party rights. Take care with anything you publish or sell, and check the Intellectual Property Office pages on GOV.UK for the current position.

Choosing accounts and tools

The useful distinction is less about free versus paid and more about personal accounts versus accounts your organisation manages. A personal account gives you no admin control or visibility, and its default settings may allow conversations to be used for training. Organisation-managed business or team services generally offer:

• an admin console, so you can add and remove users as people join and leave
• contractual commitments not to train models on your business data
• clearer terms on retention, security, subprocessors and international transfers

Protections depend on the specific service or licence, and some may already come with your existing work accounts, so check exactly what you have. Look beyond where data is stored: check subprocessors, whether anyone overseas can access it and what safeguards cover restricted international transfers. A UK hosting location alone does not settle compliance. Check the provider's role too, and ensure an appropriate processor contract is in place where required. Terms change, so read the current business terms on each provider's website.

Microsoft 365 Copilot needs a specific mention. When accessing your Microsoft 365 files, emails and other organisational content, it respects the user's existing permissions. If those permissions are messy, with salary spreadsheets or HR folders shared with everyone years ago, Copilot can surface them. Before switching it on, review who can access what in SharePoint, OneDrive and Teams.

Checking what comes out

Treat AI output as a first draft from a capable but unreliable assistant. Someone who knows the subject should read it before anyone relies on it, particularly for:

• anything going to a customer, supplier or the public
• facts, dates, figures and calculations
• legal, tax, employment or health and safety points
• quotes and citations, checked against the original source

For legal or regulatory questions, confirm the answer with GOV.UK, the relevant regulator or a qualified adviser.

Your situation may be slightly different. ask a question below ↓ and our editorial team will reply with our advice.

Shadow AI, prompt injection and fraud

Shadow AI means staff using tools the business does not know about, usually on personal accounts. A ban without clear rules and usable alternatives can encourage unapproved use, so provide approved tools and make it easy to ask whether something new is acceptable.

Prompt injection matters as tools gain access to email, files and the web. When an assistant reads an email, document or web page, instructions embedded in that content, whether visible or hidden, can try to steer it into ignoring your request, revealing information or taking actions you did not intend. Give tools only the access they need, switch off unused connectors and actions, and have a person approve any email, file change or payment. Approval helps but is not a complete defence. The National Cyber Security Centre website has current advice.

AI also makes fraud more convincing. A plausible voice, video or email is not proof of identity, so verify unusual payment requests through a contact channel you already know. Check GOV.UK for the current national fraud reporting service; in Scotland, report to Police Scotland.

Setting up sensible controls

A page or two, agreed and explained to staff, covers most small businesses. Work through these steps in order:

1. Choose approved tools. Pick one or two, set them up as organisation-managed accounts and tell staff to use these for work.
2. Check the settings. Check model-training, retention and deletion settings separately, and confirm which your administrator can enforce. Switch on multi-factor authentication.
3. Write a short never-paste list (see below).
4. Reduce data before uploading. Remove identifiers, use dummy data where possible and upload only the extract needed. Replacing names alone may not make information anonymous.
5. Agree the human check. Nothing AI-generated goes to a customer without a person approving it.
6. Keep a simple record of which tools are used, by whom and for what. This helps with accountability.
7. Cover AI in your privacy notice. Saying you "use AI" is not enough: explain the purposes, recipients and any transfers. Establish your lawful basis before processing; health data also needs a special category condition. Before processing likely to result in a high risk to people's rights and freedoms, complete a DPIA.
8. Review every six months or so, as tools and terms change.

The never-paste list might include:

• passwords, PINs, security codes or API keys
• customer or staff personal data, especially health, financial or HR details
• confidential contracts or anything under a non-disclosure agreement
• bank details and card numbers

The ICO publishes guidance on AI and data protection covering lawful basis, transparency, accuracy and risk assessment. It is updated as law and technology develop, so check the ICO website rather than relying on summaries.