How to Use AI Tools at Work Without Leaking Confidential Data: A Simple Policy for Small Teams

Why does AI use at work create a data-leak risk?

When you type or upload something into an AI tool, that content travels to the provider’s servers. What happens next depends on the tool, the plan and the settings. The UK’s National Cyber Security Centre (NCSC) has advised that queries are visible to the provider, are stored, and are likely to be used at some point to develop the service or model, so you should read the terms of use and privacy policy before asking anything sensitive. It also advises against including anything in a public chatbot query that would cause problems if it became public.

A real example: in 2023, Bloomberg reported that Samsung engineers had accidentally leaked internal source code by uploading it to ChatGPT, after which the company restricted generative AI use on its devices and networks. The employees were not attackers. They were trying to work faster. That is the pattern a policy is designed to prevent.

Do consumer and business AI accounts handle data differently?

Often, yes, and this is the single biggest lever you have. Vendors commonly document different defaults for personal and organisational plans. For example, OpenAI states that business data from its business and enterprise products and API platform is not used for model training by default, while personal plans have separate controls that may allow chats to be used for training unless you opt out. Microsoft likewise distinguishes between work accounts and consumer accounts for Copilot. Other providers follow similar patterns, but terms change, so always check the current documentation for the exact tool and plan.

The practical takeaway: work data should only go into work-managed accounts. “Not used for training” is not the same as “never stored” or “never seen by anyone,” so even on a business plan the Red list below still applies.

What is the Red-Amber-Green (RAG) data rule?

Instead of memorising dozens of rules, classify every piece of information into one of three colours before you use AI.

What should you never paste into an AI tool? (The “Never Paste This” list)

Print this list and pin it where the team works.

• Credentials: passwords, API keys, access tokens, private keys, recovery codes and one-time passcodes.
• Customer and employee personal data: full names with contact details, ID numbers, addresses, dates of birth, payslips, HR or performance records.
• Financial data: bank or card numbers, unreleased revenue figures, investor or fundraising documents, tax records.
• Health information about any identifiable person.
• Legal and contract material: signed contracts, NDAs, settlement terms, privileged legal advice.
• Trade secrets and unreleased plans: proprietary source code, product roadmaps, pricing strategy, M&A discussions.
• Client-owned material you are not permitted to share with third parties under your contract.
• Security details: network diagrams, vulnerability reports, internal system configurations.
• Anything you would not be comfortable seeing in a news story or sending to a competitor.

How do you make Amber data safe to use? (The 3-Step Scrub)

1. Remove identifiers. Replace names, emails, phone numbers, company names and account numbers with labels such as [CLIENT A] or [EMPLOYEE 1].
2. Cut to the minimum. Paste only the paragraph or data you need, not the whole document or thread.
3. Re-read before sending. Scan once for hidden details: email signatures, file metadata, comments, tracked changes, and numbers that could identify a client.

Scrubbing reduces risk but does not eliminate it. If removing the identifiers still leaves a document that clearly points to one client or deal, treat it as Red.

How do you choose which AI tools are approved? (The 5-Question Vendor Check)

Before adding any AI tool to your approved list, answer these five questions. If you cannot answer one, ask the vendor or choose another tool.

1. Training: does the vendor use our inputs to train its models, and is that off by default on our plan?
2. Retention: how long are prompts and files stored, and can we delete them?
3. Access: who at the vendor can read our data, and under what conditions?
4. Controls: does it offer admin controls, single sign-on and the ability to remove users?
5. Contract: are there business terms or a data processing agreement, and where is the data processed?

Also watch for AI features built into tools you already use (note-takers in meeting apps, assistants in email or documents). They count as AI tools too, and the same rules apply.

What does a one-page AI use policy for a small team look like?

Copy the template below into a document, fill in the brackets and ask each team member to confirm they have read it. It is a starting point, not legal advice, so have a qualified professional review it if you handle regulated data.

What should you do if someone already pasted confidential data?

1. Don’t panic or hide it. Speed matters more than blame.
2. Record what happened: which tool, which account, what data, when.
3. Delete the conversation or upload where the tool allows it, and check the retention and deletion options in the tool’s settings.
4. Rotate credentials immediately if any password, key or token was included.
5. Assess notification duties. If personal data or client information was involved, check your contracts and local privacy laws, and get professional advice on whether notification is required.
6. Fix the cause: a missing tool, unclear rule or time pressure. Update the policy accordingly.

How do you roll the policy out in one week?

Avoid blanket bans as your only response. Banning tools outright often pushes people toward personal accounts you cannot see, which is the opposite of what you want. Clear rules plus an approved option work better.

Frequently asked questions

Can employees use ChatGPT or other AI chatbots at work?

Yes, if your policy allows it. Use an approved, work-managed account, follow the Red-Amber-Green rule and never enter Red data.

Does turning off chat history or training make it safe to paste confidential data?

No. It can reduce how your data is used, but the provider may still store or process it, and mistakes or breaches can still occur. Red data should stay out of AI tools regardless of settings.

Are business AI plans safer than free ones?

Generally they offer stronger defaults, such as no training on your data by default, admin controls and business contract terms. Confirm the details in the vendor’s current documentation, because terms change.

Do we need an AI policy if we are only a few people?

Yes. Small teams often have the least oversight and the most informal tool use. A one-page policy takes under an hour and prevents the most common mistakes.

Should we ban AI tools entirely?

Usually not. Bans tend to push use onto personal devices and accounts that you cannot monitor. Approved tools with clear rules are a safer middle path, although some regulated sectors may need stricter limits.

How often should we update the policy?

Review it every six months and whenever you add a new tool, since vendor terms and features change frequently.

Sources and further reading

• UK NCSC: ChatGPT and large language models, what’s the risk?
• Bloomberg: Samsung bans staff’s AI use after spotting ChatGPT data leak (May 2023)
• OpenAI: Enterprise privacy
• NIST: AI Risk Management Framework

Vendor terms, plans and settings change often. Check each provider’s current documentation before relying on any default described here.