Should I Be Worried About Employees Using ChatGPT at Work?

Some concern is warranted; panic is not. The main risk is rarely the AI tool itself. It is staff pasting confidential or personal information into services your organisation has no contract with and no ability to audit. That is a governance problem with practical fixes.

This page answers the questions owners and IT leads ask in the first week after realising their people are already using AI: the risks worth acting on, the ones commonly overstated and the first steps that help.

Should I be worried about employees using ChatGPT at work?

Be concerned enough to act, not enough to panic. Staff adopt AI assistants to work faster and they move quicker than any approval process, so the exposure lies in what leaves your control: customer records, contract terms, pricing, source code, HR cases. Once that information sits in a consumer account, you cannot recall it or evidence its handling to a client or a regulator. Unsanctioned use of this kind is Shadow AI, and Shadow AI Discovery establishes its real extent.

Is it safe for staff to use ChatGPT?

Safety depends far more on the account than on the model. A consumer account, opened with a work email address and paid on a personal card, carries no processing agreement and no reliable record of what was submitted. A paid business or enterprise tier, configured properly and covered by a contract, gives you named users, administrative control, retention settings and an audit trail. The tool can be safe; an unmanaged account rarely is, which is why AI governance matters more than the choice of supplier.

What are the real risks of employees using AI tools?

Most of the exposure sits in four ordinary places rather than in the model itself. None of them needs a hostile actor and each becomes manageable once you know where AI is in use.

• Confidentiality: commercially sensitive material submitted to a service outside your contractual perimeter.
• Data protection: personal data processed with no lawful basis and no record of the activity, which raises UK GDPR questions.
• Accuracy: plausible output relied on without review, in advice, code or client correspondence.
• Access: assistants connected to mailboxes, files or line-of-business systems with wider permissions than the person using them.

Which risks are overblown?

Two fears are usually overstated: that a single prompt trains a public model and exposes your data to strangers, and that AI will quietly replace professional judgement across the organisation. Paid business and enterprise agreements typically exclude submitted content from model training, so read the agreement you signed rather than assume the worst. The realistic failure is duller: an unreviewed answer from an unmanaged account, with no record of either. Treat this as a control problem and its scale shrinks considerably.

What could go wrong if staff paste customer data into an AI tool?

You lose the ability to answer a client who asks where their data has been processed, and you may breach a confidentiality clause or data processing schedule you have already signed. You may also process personal data without a lawful basis or a documented transfer route, which can become a reportable incident. In regulated sectors the finding often concerns weak oversight rather than the disclosure itself. If it has already happened, our answer on an employee putting confidential data into an AI tool sets out the response.

How do I find out which AI tools our staff are already using?

The evidence sits in systems you already own: DNS and web proxy logs, single sign-on and OAuth grants, browser extension inventories, expense claims and card statements. Combine that technical view with a no-blame amnesty survey, because people describe their workarounds openly when the answer carries no disciplinary consequence. Expect the resulting list to be longer than the one you would predict. Our answers on finding the AI tools already in use and on Shadow AI Discovery set out the method step by step.

Should we ban ChatGPT or allow it?

A ban is easy to publish and hard to enforce. Blocked at work, the same task moves to a personal phone, where you have no logging and no evidence whatsoever. The workable position pairs a sanctioned route with a short list of prohibitions: an approved tool covered by a contract, clear rules on what must never be submitted and a named owner for exceptions. We compare both positions in should we ban ChatGPT.

What are the first three things we should do?

Discover, publish rules, then offer a sanctioned alternative, in that order. Formal frameworks such as ISO 42001 become useful once those three are in place, not before.

• Discover. Identify every AI tool and connected account in use, and who pays for them.
• Publish one page of rules. A short acceptable use policy beats a long standard nobody reads.
• Offer a sanctioned option. Provide a contracted tool that covers the same work, then withdraw tolerance for the rest.

For an independent view of where AI is already in use across your organisation, our AI Security Gap Analysis sets out what to fix first, in priority order. Book a call to discuss scope.