A Workplace Generative-AI Use Policy: A Build Checklist

Before you draft: what this policy is for

A generative-AI acceptable-use policy is not a ban and not a permission slip. It is the document that lets employees capture the productivity benefit without creating confidentiality, IP, accuracy, or discrimination exposure that lands on the company. The organizing idea throughout: AI changes the workflow, not the legal duties. Competence, verification, candor, confidentiality, and non-discrimination remain the company’s obligations, and “the model produced it” is not a defense. The notes below are general information, not legal advice; AI law is moving quickly and varies sharply by jurisdiction, so scope every regulatory reference to your jurisdictions and confirm its current status, scope, and applicability before relying on it.

1. Approved tools — give people a safe option

The single most effective anti-shadow-AI measure is a sanctioned tool that is actually good enough to use. A policy that only prohibits drives usage onto personal accounts where the company has no visibility and no contractual protection.

Name the approved tools. Maintain a short, current list of sanctioned, enterprise-grade tools — ideally ones covered by a contract with confidentiality, no-training, and security commitments. The list, not the prohibition, is what people will follow.
Define “shadow AI” and why it’s off-limits. Consumer accounts and unvetted browser extensions or plugins carry none of the enterprise protections. Make clear that work tasks go through approved tools only.
Give a path to request new tools. People adopt tools faster than policy can list them. A lightweight intake — who approves, what gets reviewed — keeps the sanctioned list a living document instead of an instant relic.

2. Data rules — the one bright line

If employees remember nothing else, they should remember this section. Most real-world AI incidents are data incidents: someone pasted something sensitive into a tool that retains or trains on inputs.

State the bright line in one sentence. Never input confidential, client, trade-secret, personal or otherwise regulated, or privileged information into any tool that is not approved for that data. Make it short enough to repeat from memory.
Explain where prompts go. Employees should understand that a prompt leaves the building — it travels to a vendor, may be logged, and in consumer tiers may be used to train the model. “Where does this input go, and could it train the model?” is the question to ask before pasting.
Map data categories to tools. Some data may be fine in an enterprise tool and never acceptable in a consumer one. Tie the rule to the company’s existing data-classification scheme rather than inventing a parallel one.

3. IP and ownership — manage two risks at once

Generative AI raises a question on both ends of the work: whether you own what comes out, and whether what comes out infringes someone else.

Flag that AI output may not be protectable. Current U.S. Copyright Office guidance treats purely AI-generated material as generally not copyrightable, with human authorship driving what is — an area still developing, so confirm the current standard for your use. Material the company needs to own and enforce deserves meaningful human authorship.
Address third-party infringement risk. Outputs can echo copyrighted, trademarked, or trade-secret material from training data. Treat AI output like any other unvetted third-party content before it ships externally.
Preserve the company’s own IP. Don’t feed proprietary source code, designs, or unpublished work into tools whose terms could compromise ownership or secrecy. Mark or log AI-assisted work where a client, contract, or court requires it.

4. Accuracy and verification — AI drafts, humans decide

The non-negotiable rule of the whole series: a human must verify every AI output before it is used, sent, or relied on. The tool produces a draft, never an authority. Fabricated citations, invented facts, and confident-sounding errors are features of the technology, not anomalies.

Require human verification, by role. Whoever uses the output owns its accuracy. Verification means checking against an authoritative source, not re-reading the output for plausibility.
Match scrutiny to stakes. A brainstorm needs less checking than a filing, a financial figure, or a customer commitment. The policy should make the gradient explicit so people calibrate.

5. Prohibited and high-risk uses

Some tasks carry enough legal or safety exposure that AI may only assist under human control — or not at all. Many in-house teams find it useful to list these explicitly rather than leave them to judgment.

Use	Posture
Legal advice or legal conclusions	Human counsel decides; AI output is never the answer of record
HR and employment decisions (hiring, discipline, pay, termination)	High-risk; automated decision tools may implicate frameworks such as NYC Local Law 144, Illinois’ AI Video Interview Act and BIPA, or the Colorado AI Act — confirm current status, scope, and applicability before relying on any of them
Regulated, safety-critical, or medical outputs	Human review required; sector rules govern regardless of the tool
Security-relevant or production code	No deployment without human review and testing

6. Disclosure — when to say AI was used

Know the audiences with disclosure expectations. Clients and counterparties may require it by contract; some courts have entered standing orders on AI use; some regulators and customer-facing contexts expect notice. The duty of candor doesn’t change because a machine helped draft — confirm what the relevant forum or contract currently requires.
Set a default and an escalation path. When in doubt, ask before representing AI-assisted work as fully human, and route novel disclosure questions to legal rather than guessing.

7. Confidentiality and privilege

Treat the input as potential disclosure. Pasting third-party confidential information into an unapproved tool can breach an NDA. The obligation runs to the data owner, not the vendor.
Protect privilege deliberately. Sharing privileged material with a tool that retains or processes it through third parties can risk waiver. Privileged work belongs in approved, protected channels, and AI questions touching legal strategy should run through counsel.

8. Security

Enterprise accounts, not consumer. Require single sign-on, administrative controls, and contractual terms that consumer tiers don’t offer. The account type is a security control, not a convenience choice.
Turn off training on your data — and verify it. Confirm the no-training setting or contract term for each approved tool, and re-confirm it as terms change.
Bring AI into existing controls. Data-loss-prevention, logging, and access reviews should cover AI tools the way they cover any other system that touches company data.

9. Training, enforcement, and fit

Train on the rule people must remember. A short, concrete session beats a long policy nobody reads. Lead with the bright-line data rule and the approved-tool list.
State consequences and apply them evenly. Tie violations to existing discipline, and enforce consistently — selective enforcement undercuts the policy and creates its own exposure.
Connect it to policies you already have. This document should reference, not contradict, your confidentiality, IP-assignment, acceptable-use, and data-classification policies. Reconcile conflicts before publishing.

The pattern that works

A policy that just says “don’t use AI” gets ignored and breeds shadow AI on personal accounts. The ones that hold up give employees a sanctioned, enterprise-grade tool plus one bright-line rule everyone can recite: never paste confidential or client data into a tool that is not approved. Give people the safe option first; the prohibition only works once the alternative exists.