How to handle staff use of personal ChatGPT accounts

If you lead IT or compliance at a regulated firm, you have almost certainly had this thought: *how many of our people are pasting work into a personal ChatGPT account right now?* It is one of the most common forms of shadow AI, and one of the easiest to underestimate. Someone summarises a client email, drafts a report, or cleans up a spreadsheet using a free account they signed up for at home. It feels harmless and saves them an hour. It is also, from a governance point of view, a quiet problem.

This article is a practical playbook for handling it. Not a lecture, and not a call for a blanket ban, because bans rarely work the way people hope.

Why personal accounts are the awkward case

Sanctioned tools sit inside your control. You have a contract, a data processing agreement, retention settings, and an admin console. Personal accounts have none of that. When a member of staff uses their own login for work, three things happen at once:

• There is no audit trail you can reach. The activity lives in someone's private account, outside your tenancy and your retention policy.
• Firm and client data leaves the building. It goes to a consumer service governed by consumer terms, often with different defaults for how inputs are handled.
• You cannot evidence what happened. If a client, auditor or regulator asks what was processed and where, you have nothing to show.

That last point is the one that bites in a regulated context. The issue is not only that data moved; it is that you cannot demonstrate what did or did not occur. Absence of evidence becomes its own finding.

Why a blanket ban usually backfires

The instinct is to block it all. Sometimes that is the right call for specific tools or specific data. But a firm-wide ban on consumer AI tends to push usage underground rather than stop it. People switch to phones, home laptops and copy-paste, and you lose the little visibility you had. You also send a message that the organisation has no answer to a tool staff find genuinely useful, which erodes trust in the policy itself.

A governed route plus a clear record beats a ban that everyone quietly works around.

A better frame is to make the safe path the easy path, and to keep a record of the rest. Four steps.

Step 1 — See it

You cannot govern what you cannot see. Start by surfacing where consumer and personal AI use is actually happening.

Most firms already collect the raw signal. Firewall, secure web gateway (SWG) and SIEM logs record connections to AI destinations, and Evaident surfaces personal and consumer AI use from those logs you already hold. This is detection by destination: it tells you that traffic went to a given AI service and how often, not what was typed. Prompt content sits inside TLS, so log-based detection cannot read it, and it should not pretend to.

Where you need a closer view of web-based AI use, including personal and free accounts that never touch a corporate login, a managed browser extension can capture that activity, recording metadata by default. That gives you a clearer picture of usage without turning into surveillance of every keystroke.

Two points worth keeping straight:

• Detection is not blocking. The browser extension tells you what is happening; actually blocking web access stays with your firewall, SWG or browser controls.
• Default capture is metadata, which is usually the right starting posture for a regulated workforce.

Step 2 — Provide a sanctioned route

Visibility on its own just creates a list of people to be cross with. The point of seeing the problem is to give staff somewhere better to go.

Offer an approved tool that meets your data and retention requirements, so people have a legitimate way to do the thing they were already trying to do. For API-based tools, agents and internal integrations, route them through the Evaident Gateway, where policy is enforced centrally and the activity is logged in one place rather than scattered across personal accounts.

The goal is simple: the sanctioned path should be at least as convenient as the personal account. If it is slower or more locked down than the free tool, you are quietly recreating the incentive to go around it.

Step 3 — Set policy

With a real alternative in place, policy stops being an empty threat and becomes a clear instruction.

• Publish an acceptable-use policy that states which tools are approved, what categories of data may and may not be used with AI, and what personal-account use means in practice.
• Pair it with a staff notice so people understand what is being detected and why. Transparency is not just courtesy; under UK GDPR you need a lawful basis for monitoring, and staff should know it is happening.
• Where you capture content rather than metadata, complete a DPIA before you switch that on, and document the decision.

Written down and communicated, the policy turns "we didn't realise" into a clear expectation, which matters if you ever need to act on a breach.

Step 4 — Evidence it

Finally, bring detection, policy and your sanctioned route together into one tamper-evident record. When a client asks how you govern AI, or an auditor wants to see your controls in action, you can show what is in place, what was detected, and how exceptions were handled, rather than describing intentions.

This is the difference between *saying* you manage shadow AI and being able to *show* it. For a regulated firm, the second is the only version that counts.

Putting it together

Personal ChatGPT accounts are not going away, and pretending otherwise helps no one. See the usage from logs you already collect, give people a sanctioned route that is genuinely easier than the workaround, set policy that staff understand, and keep one record you can stand behind. That is a governance posture you can defend, rather than a ban you cannot enforce.

Want to know where consumer AI use is already happening across your firm? See your exposure.

*This article supports your compliance work and is not legal advice.*