How to control employee ChatGPT use

Last reviewed: 28 September 2026

This guide explains regulation and good practice. It is not legal advice; have your own legal counsel review your policy and any monitoring before you roll them out.

Why isn't blocking ChatGPT enough?

A blanket block removes the approved route but not the demand. People who find ChatGPT useful keep using it on personal accounts or phones, where the company has no account controls, no data terms and no records.

Blocking has a place: it is the right decision for tools you have assessed and rejected. For ChatGPT itself, most companies get better control by approving one company workspace, setting rules for what may go into it, and blocking the routes around those rules. VAHTOR's own guide on pasting data into ChatGPT makes the same point: a full block can push use to personal phones and accounts.

Which controls exist, and what does each one stop?

No single control covers every risk. The table shows what each layer actually stops and where it stops working, so you can combine them deliberately.

Sources for each row are listed at the end of the page.

How do you set it up, step by step?

Start with what people already use, decide per tool, then add the account, the policy, the technical controls and the training. Most companies with 25 to 100 employees can do steps 1 to 5 within a few weeks.

1. 1Find out which AI tools are already in use. Check SSO logs, expense claims, browser extensions and web proxy logs, or use a discovery tool. Record each tool, how many people use it and for what. Our guide on seeing which AI tools employees use covers the options.
2. 2Decide per tool: approve, restrict or block. Approve tools with a company account and acceptable data terms. Restrict tools that are fine for some tasks, for example "public information only" or "no file uploads". Block tools you have assessed and rejected. Write the decision and the reason into an AI tool register.
3. 3Offer an approved company account. ChatGPT Business starts at two users; see OpenAI's pricing page for current prices, which vary by country. OpenAI states it does not train on Business or Enterprise workspace data by default, supports SAML SSO and offers a data processing addendum. Enterprise adds SCIM provisioning, role-based access control and user-level analytics.
4. 4Publish the policy and collect acceptance. Name the data that never goes into AI tools: personal data about customers or staff, confidential client material, credentials and unreleased financials. Record who accepted which version. In ChatGPT Enterprise, admins can also show a workspace AI policy notice, which OpenAI says appears every 30 days and whenever its content changes.
5. 5Enforce data rules in the browser. Add a control that stops sensitive files and secrets at the point of use, before they reach the AI tool. The options range from Microsoft Purview DLP in Edge and Chrome Enterprise Premium DLP rules to a lighter browser extension; our guide to blocking file uploads to ChatGPT compares them.
6. 6Close the personal-account route where it matters. On ChatGPT Enterprise, a secure web gateway or proxy can insert a ChatGPT-Allowed-Workspace-Id header so only your workspace works on managed traffic, and block the logged-out endpoint. OpenAI's documentation says this needs SSL inspection, and the desktop and mobile apps need your certificates added through device management because they use certificate pinning.
7. 7Train people and keep the record. Since 2 February 2025, providers and deployers of AI systems must take measures for their staff's AI literacy (EU AI Act Article 4). The Commission's example for a company using ChatGPT to write advertising text is to inform staff about risks such as hallucination. An internal record of training is enough; no certificate is required.
8. 8Connect identity and review monthly. Enforce SSO for the company workspace, remove leavers on their last day (automatically via SCIM on Enterprise), and review the tool register, alerts and new tool requests once a month.

What can an employer see about employees' ChatGPT use?

It depends on the account and the tools in place. An employer sees nothing through OpenAI about a personal ChatGPT account, while ChatGPT Enterprise lets workspace owners export conversations through OpenAI's Compliance Platform.

Whatever you choose, tell employees in the policy what is monitored and what is not. Hidden monitoring erodes trust and is hard to justify under data protection rules.

What do EU rules say about controlling and monitoring AI use at work?

The EU AI Act asks deployers to support AI literacy and bans some workplace uses, such as emotion recognition. GDPR limits how much employee data your monitoring may collect.

• AI literacy (Article 4). Applies since 2 February 2025. The Digital Omnibus on AI, in force since 27 July 2026, keeps the obligation but no longer requires a specific level; national market surveillance authorities supervise it from 2 August 2026.
• Emotion recognition at work is prohibited (Article 5(1)(f)), except for medical or safety reasons. Do not use monitoring tools that infer employees' emotions from their typing, voice or faces.
• Data minimisation (GDPR Article 5(1)(c)). Collect only what the purpose needs. Recording tool use and detection types is easier to justify than storing every prompt.
• Employee data (GDPR Article 88) allows national rules on processing in employment. In Finland, for example, the Act on the Protection of Privacy in Working Life (759/2004, section 21) puts technical monitoring and the use of email and networks into the co-operation procedure, after which the employer must define the purpose and methods and inform employees.
• High-risk AI at work (Article 26(7)). Deployers must inform workers' representatives and affected workers before putting a high-risk AI system into use in the workplace. ChatGPT used for drafting is not high-risk by default, but AI used for recruitment or evaluating staff can be (Annex III, point 4).

Involve your data protection officer or employee representatives before you switch on monitoring.

What are the most common mistakes?

Most failures come from relying on one layer. The frequent ones:

• Blocking everything. Use moves to phones and personal accounts, where you have no controls.
• A policy with no technical control. People forget rules under deadline pressure; a browser rule catches the file they did not mean to upload.
• A technical control with no policy. Employees see a blocked upload but do not know what is allowed instead.
• Buying company seats but leaving personal accounts open. The company workspace only protects the data that goes into it.
• Collecting prompts "just in case". Storing full conversations creates a sensitive dataset you must justify, secure and delete.
• Forgetting leavers. Without SSO and offboarding, former staff can keep access to the workspace and its shared projects.
• Training once. Tools change monthly; repeat short training when you approve a new tool or change the policy.

Checklist

• AI tool register with approve, restrict or block decision and reason for each tool
• Approved company AI account in place, with SSO enforced
• AI policy published; acceptance recorded per version
• Never-share data classes named in the policy
• Browser control for files and secrets on managed devices
• Decision taken on personal accounts on the corporate network
• AI literacy training delivered; records kept per person
• Monitoring described to employees; DPO or employee representatives consulted
• Leavers removed from AI tools on their last day
• Monthly review of register, alerts and requests

How does VAHTOR help?

VAHTOR covers the policy, per-tool and browser layers in Chrome and Edge without collecting prompts or conversations. It does not replace a company ChatGPT account, and it does not make a company compliant on its own.

• Discovery and per-tool decisions. VAHTOR shows which AI tools people use and lets you approve, restrict or block them; devices pick up the decision on the next sync.
• Policy and training records. Publish a company AI policy from a template, track who accepted the current version, and keep literacy quizzes and pass marks for the team.
• File and secret rules. Rules can block PDFs, Excel and office files attached to AI tools, and credentials and secrets are flagged on the device. Chat text is not stored.
• Limits. VAHTOR works in Chrome and Edge on devices with the extension installed. Desktop AI apps, mobile devices and other browsers are outside its reach.

Starter (EUR 199 per month + VAT, up to 25 employees) includes discovery, policy setup, acceptance tracking and approve, restrict or block. Browser DLP rules and AI literacy training tracking are in Business (EUR 499 per month + VAT, up to 100 employees). See pricing.