How do you control employee use of generative AI?
You control employee use of generative AI by combining four things: a short acceptable use policy, a list of sanctioned tools with a fast route to request new ones, data controls on what can be sent in a prompt, and proportionate monitoring that starts with metadata. A ban alone rarely works, and allowing everything leaves no record of what data left the organisation.
The goal is not to stop people using AI. It is to make the approved route the easiest one, so that use happens where you can see it and apply controls.
The four options compared
Most organisations pass through more than one of these. The question is which one you settle on, and for which data.
| Option | What it means | Advantages | Disadvantages |
|---|---|---|---|
| Ban | Block known AI domains and prohibit use | Simple to state. Buys time while an alternative is prepared | Use moves to personal devices and unlisted services. No route to approval. Staff lose a tool they find useful |
| Allow everything | No restriction, perhaps a reminder to be careful | No friction. Fast adoption | No record of which services hold company data. No control over identifiers in prompts. Nothing to show an auditor |
| Sanctioned list with approval route | Named tools are approved, anything else goes through a request | Clear to staff. Each approved tool has been assessed. Creates an inventory | Depends on the approval route being quick. Does not control what data is entered into approved tools |
| Gateway with data controls | Routed traffic passes a control point that checks identity, destination and prompt content | Rules apply per request. Identifiers can be blocked or redacted. Every decision is recorded | Covers only traffic routed through it. Needs a privacy assessment for content inspection. More to operate |
The last two work best together. The sanctioned list decides which services are acceptable. The gateway decides what may be sent to them, request by request.
What an acceptable use policy should contain
A policy staff actually read is short and concrete. These are the sections that matter in practice. The wording and its legal effect are for your organisation and its counsel to settle.
- Scope. Who it applies to, including contractors, and which kinds of AI it covers: chat services, AI features in SaaS products, coding assistants, APIs and agents.
- Approved tools. A link to the current list, rather than names in the policy text, because the list changes more often.
- Data rules. What may never be entered, such as customer personal data, credentials, payment data or confidential client material, and what is allowed in approved tools.
- Output responsibility. The person using the output is responsible for checking it before it is relied on, published or sent to a customer.
- Requesting a new tool. Where to ask, what information is needed and how long a decision should take.
- Agents and automation. Connecting an AI tool to company systems, or letting it act, needs its own approval.
- Monitoring. What is logged, whether any content is inspected, why, and who can see it.
- Reporting mistakes. How to report data entered by mistake, with a clear statement that early reporting is expected and welcome.
- Ownership and review. Who owns the policy and when it is next reviewed.
How to monitor use proportionately and lawfully
Monitoring is where AI governance meets employee privacy. Whether a given form of monitoring is lawful depends on your jurisdiction and your agreements with staff, so that determination belongs to your organisation and its counsel. A proportionate design usually follows three layers.
- Metadata first. Proxy and DNS logs show which AI services are contacted, from which hosts and how often, without any prompt content. This answers most governance questions on its own.
- Content inspection only where routed and documented. Inspecting what is in a prompt is possible only for traffic sent through a gateway. Limit it to defined categories of data, record the purpose and retention in your privacy assessment, and tell staff it happens.
- Involve the right people early. Bring in your data protection officer before any content control is switched on. Where a works council or employee representatives exist, consult them in line with your local arrangements.
Monitoring that staff know about, with a stated purpose and a narrow scope, is easier to defend and less likely to push use onto personal devices than monitoring discovered after the fact.
Data controls for prompts
Much leakage through AI tools is ordinary: a pasted spreadsheet, a customer email, a configuration file with a key in it. Controls that target structured data catch much of this without reading free text for meaning.
Structured identifiers
Define the categories you care about, such as email addresses, national identification numbers, payment card numbers, bank account numbers and API keys. These have recognisable formats, so rules can find them consistently.
Block or redact
For each category, choose an action. Blocking stops the request and tells the user why. Redacting replaces the identifier before the request leaves, so the user still gets an answer. Allowing with logging records that the category passed, which is useful while you calibrate.
Residency
Some data must stay in a region. A residency rule checks where the provider will process the request and refuses it if the region is not permitted, or if it cannot be determined.
A 30-day rollout plan
- Days 1 to 7: see what is in use. Export a few weeks of proxy or DNS logs and match them against known AI services. Ask teams to declare the tools they rely on, with an owner and a purpose.
- Days 8 to 14: decide the sanctioned list. Pick the services that meet your security and privacy requirements, and agree who approves new requests and how quickly.
- Days 15 to 21: write and agree the policy. Draft the sections above, review them with your DPO, legal and, where relevant, employee representatives, and publish with the approved list.
- Days 22 to 30: route and control. Point applications and approved tools at a gateway, switch on data controls in a logging mode first, then move the categories you are confident about to block or redact.
Employee AI governance checklist
- A first scan of proxy or DNS logs against known AI services
- A published list of sanctioned AI tools, with an owner
- An approval route for new tools with a target response time
- A short acceptable use policy covering the sections above
- A privacy assessment covering any content inspection, reviewed by the DPO
- Consultation with employee representatives where they exist
- Data controls for structured identifiers, each set to block, redact or log
- A residency rule for data that must stay in a region
- A way for staff to report a mistake without penalty for reporting
- A review date for the policy and the sanctioned list
How MERIDVAR helps
MERIDVAR is self-hosted software you install and operate on your own network. Nothing is sent to MERIDVAR.
- Shadow AI discovery. Traffic is compared with your sanctioned provider list in off, observe or enforce mode, and proxy or DNS logs can be uploaded. Endpoint agents, with installers for Linux, macOS and Windows, report local AI tools.
- Admission gateway. Requests arrive through an OpenAI-compatible API, a console playground, a PAC file or an HTTP CONNECT proxy, with OIDC sign-in at the gate. A refused request never reaches the model provider.
- Prompt DLP. Ten categories of structured identifiers, each set to block, redact or allow with logging, using rules rather than machine learning.
- Residency and approved models. Data residency rules fail closed on an unknown region, and approved model lists limit what can be called.
- Approval and evidence. A new tool is approved by an administrator who did not request it, with conditions and an end date, and decisions are kept in an append-only, hash-chained evidence log.
Controls apply only to traffic routed through the gateway, and the product is in beta. See AI security for the admission controls and shadow AI for discovery.