Staff Using AI Tools at Work: The Real Risks and a Workable Policy
Your team is already pasting work into AI chatbots. A ban will not hold. Here are the risks that matter, the ones that do not, and a one-page policy that works.

Ask a room of fifteen staff who uses AI tools for work and perhaps four hands go up. Ask instead what they used last week to tidy up an email, summarize a long document or draft a job advert, and the number is closer to twelve. Several are using personal accounts on personal phones, because the work laptop has a block on the site.
That gap is the thing to manage. Not whether AI tools are used, which was decided for you some time ago, but whether the use is visible and whether anyone has told people where the lines are.
The risks that are real
Four of them matter to a small business. Keep these separate from the general anxiety, because mixing them produces policies that are both too strict and too vague.
Data leaves your control
Text pasted into a chatbot has been sent to an outside company. Depending on the account type it may be retained, reviewed by staff for safety or quality, and used to improve the service. On consumer plans, conversation content is commonly used for training unless you turn that off. Business and enterprise plans usually exclude it contractually, which is the main practical reason to pay.
The high-risk pastes in a small business are predictable: a customer list "to tidy up the formatting", a draft contract "to check it reads clearly", a spreadsheet of salaries "to summarize", an incident timeline containing client names. Each is a disclosure to a third party, and under data protection law the business remains responsible for it.
Output that is confidently wrong
These tools generate plausible text, which is not the same as correct text. Invented figures, invented citations, invented legal provisions and invented product capabilities all appear, and they appear in fluent, authoritative prose. The risk scales with how far the output travels: a reworded internal email is low stakes, a quotation sent to a customer or a clause pasted into a contract is not.
The tools around the tools
A chatbot in a browser tab is the easy case. The harder cases are AI browser extensions, which can read every page the browser displays including your banking and your email, and meeting notetakers, which join client calls and record entire conversations. Both get installed casually and both have far broader access than the chatbot people worry about. Notetakers also raise a consent question, since everyone in the meeting should know they are being recorded.
Attackers use the same tools
Phishing emails aimed at small businesses have got noticeably better written, and poor grammar is no longer a reliable tell. Voice cloning from a short sample of public audio is now cheap. This changes your training content rather than your technology: teach verification by an independent channel instead of teaching people to spot bad English. See AI-written phishing and voice cloning scams.
The risks that are overstated
Worth saying plainly, because misplaced fear produces policies nobody respects.
Your pasted text is not going to be served up verbatim to a competitor who asks the right question. That is not how these systems work in practice, and it is the wrong thing to be frightened of. The real problem with pasting a customer list is simpler and more boring: you have sent customer data to a company you have no contract with.
Equally, AI use is not inherently a security incident. A member of staff using an approved tool to redraft a sentence has done nothing wrong, and treating it as a violation is how you push the whole thing underground.
A one-page policy that holds
Five sections. No more. Keep it in the same place as your one-page security policy and go through it in a fifteen-minute briefing.
- Approved tools. Name two or three, on business accounts the company pays for and administers. Say that other tools need asking about first, and that the answer is often yes. An approved list with a quick route to add to it is the only version that survives.
- Never enter this. Your own concrete list: customer personal data, staff records, financial and health information, passwords and keys, unpublished pricing and contracts, anything under NDA, client material you do not own. Use examples from your actual business rather than abstract categories.
- Check before it goes out. Any output going to a customer, a regulator, a contract or your website is read and verified by a person. Figures, names, quotes and legal references get checked against a source.
- Extensions and notetakers need approval. Separately and explicitly, because of the access they hold. Add that meeting participants must be told when a notetaker is present.
- Ask if unsure, and nobody gets blamed. One named person, and an explicit line that someone who realizes they pasted something they should not have must say so immediately. That early report is what lets you assess the exposure, and it only happens if reporting is safe. See building a no-blame reporting culture.
Buy the business plan for one tool. It usually costs about the same per user as a lunch, gives you contractual terms that exclude your data from training, and gives you an admin console. More importantly it gives staff a sanctioned place to work, which is the only thing that reliably reduces use of personal accounts.
Finding out what is already in use
Ask, without consequences attached. "What AI tools are you using, and what for? Nobody is in trouble, I want to make the useful ones official." In a small business this one conversation produces a more accurate picture than any technical control, and it usually surfaces two or three tools nobody in management had heard of.
Then check the places people do not think of as tools: browser extensions installed on work laptops, AI features switched on inside software you already pay for, and automation connected to your email or file storage through an integration. Review the third-party apps authorized in your Microsoft 365 or Google Workspace tenancy, because AI integrations granted broad mailbox or drive permissions sit in that list and often outlast the person who added them.
Where AI agents change the picture
Chatbots produce text a person decides what to do with. Agents take actions: send email, move files, make changes in systems. Anything with write access to your mailbox, your files or your accounting deserves a different level of scrutiny, because the failure mode changes from a bad paragraph to a bad action taken on your behalf. Grant the narrowest permission that works, keep a human approval step for anything involving money or external communication, and review what you have connected every few months. See AI agent security risks.
Email that reads too well to be a scam? Fluent writing is no longer evidence of legitimacy. Paste it into CyberWatch AI Scan for a free check.
For the broader threat picture see AI and cyberattacks and spotting AI-generated scams. For the policy and training side, security awareness training that works and managing human risk. The wider program is in our complete guide to small business cybersecurity.
Frequently asked questions
Should we ban AI tools at work?
A ban rarely works. Staff use the tools on their phones and personal accounts instead, which removes your visibility without removing the risk. Approving a small number of tools and setting clear rules about data is more effective.
What is the actual risk of pasting work into a chatbot?
Mainly disclosure. The text leaves your control, may be retained, may be used to improve the service depending on the plan, and may be visible to the provider's staff. Treat it as sending the material to an outside party.
Is a paid business plan safer than a free account?
Generally yes, because business and enterprise plans usually exclude your content from model training and offer administrative controls. Read the specific terms for the plan you buy rather than assuming, and check where data is stored.
What should never go into an AI tool?
Customer personal data, health or financial records, passwords and keys, unpublished contracts and pricing, anything under NDA, and source code or documents you do not own. Write the list down in your own words.
Can AI output be trusted?
Not without checking. These tools produce confident text that can contain invented facts, figures, citations and legal references. Anything going to a customer, a regulator or into a contract needs a human to verify it.
What is shadow AI?
Staff using AI tools the business does not know about, usually through personal accounts or browser extensions. It is the normal state in most small businesses, and asking openly what people use is the fastest way to find out.
Are AI browser extensions and meeting notetakers risky?
They deserve more scrutiny than a chatbot. An extension can read every page in the browser and a notetaker records entire client conversations, so both need approval, and notetakers need consent from everyone in the meeting.
Do we have legal obligations if staff put personal data into AI tools?
Yes. Under data protection law you remain responsible for personal data your staff share with a third party, including the obligation to have a lawful basis and appropriate contracts. Unapproved tools mean neither is in place.
How do we write a policy people will follow?
One page, listing the approved tools, the things that must never be entered, the rule that output is checked before it goes out, and who to ask when unsure. Rules people can remember get followed.
Sources
- ChatGPT and large language models: what's the risk?, UK National Cyber Security Centre
- Machine learning principles, UK National Cyber Security Centre
- AI Risk Management Framework, US National Institute of Standards and Technology
- NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, US National Institute of Standards and Technology
- OWASP Top 10 for Large Language Model Applications, OWASP
- OWASP Top 10 for Large Language Model Applications project, OWASP
- Cloud security guidance, UK National Cyber Security Centre
- 10 Steps to Cyber Security, UK National Cyber Security Centre


