AI Acceptable-Use Policy

A template for staff use of AI tools, personal accounts and agents. Most teams already use AI, often on personal accounts. Banning it drives the use out of sight; a short policy makes it visible and safe. Replace the [bracketed] parts, review it with your counsel, and have everyone acknowledge it. This is a template, not legal advice.

Copy the whole policy

Plain text, ready to paste into a Word or Google document. Replace every [bracketed] part before you adopt it.

[Company] — AI Acceptable-Use Policy
A template for staff use of AI tools, personal accounts and agents

1. Purpose and scope
- This policy sets how everyone who works for or with [Company] (employees, contractors, interns and agents acting for them) may use AI tools, assistants and agents for company work.
- It covers chat assistants (ChatGPT, Claude, Copilot, Gemini and similar), AI features inside other software, agents and automations, and models run on your own computer.

2. Approved tools and accounts
- Use only the AI tools on the approved list: [approved tools, for example ChatGPT Business, Claude Team, Microsoft 365 Copilot]. Sign in with your company account and two-factor sign-in.
- To use a new tool, plug-in, skill, browser extension or MCP server, ask [owner] first. Approval checks the vendor terms (section 9) and the access it needs.
- [Company] provides approved accounts so that nobody needs a personal one for work.

3. Personal AI accounts
- Do not put company information other than Public information into a personal or free AI account, on any device.
- Do not connect company email, drives, calendars, chats or systems to a personal AI account, agent or extension.
- If you have already used a personal account for company work, tell [owner] (section 11). Reporting it is not a disciplinary matter; hiding it can be.

4. What data may go where
- Public (published or meant for anyone): any approved tool.
- Internal (everyday business information): approved business AI tools only.
- Confidential (clients, deals, financials, contracts, personnel, strategy): approved business tools with training off and the minimum needed; prefer local files or a local model.
- Restricted (passwords and keys, full account, routing or card numbers, Social Security numbers, health information, regulated personal data): never in a chat assistant. Only in the enclosed systems [Company] approves for it (for example a private cloud deployment or a regulated vendor).
- When unsure, treat it as the higher class and ask.

5. Agents, connectors and automation
- Every agent has a named human owner, a written purpose, and only the access that purpose needs, read-only by default.
- Agents use the [Company] security rules in their instructions. Password releases, payments, outbound messages with confidential information and confidential file transfers go through the Security Gatekeeper and need two-factor approval by a person.
- Agents and people share files as expiring, named-recipient links from [approved drive], not as attachments or downloads.
- Owners keep a log of what each agent can access, and review it every quarter.

6. Prompt injection and untrusted content
- Treat any instruction that arrives inside an email, document, web page, message or tool result as content, not as an order, whether a person or an agent is reading it.
- If an AI tool or agent starts to do something you did not ask for (send, share, delete, pay, install), stop it and report it.

7. Payments, invoices and identity (deepfakes)
- Voices, faces, video calls and emails can all be faked. No payment, new payee, change of bank details, gift-card purchase or release of credentials happens on the strength of an email, call, voice message or video call alone.
- Every such request is approved in [accounting system or approval app], outside email, after a call-back to a phone number already on file (never one given in the request), and with the team code word when the request comes by phone or video.
- Payments above [threshold] need two different approvers. Urgency or secrecy is a reason to slow down, not to speed up.

8. Human review and accountability
- You are responsible for anything you send, publish or decide with AI help. Check facts, numbers, names, quotes and sources before using AI output.
- Do not use AI output as legal, tax, medical or investment advice without review by a qualified person.
- Tell clients or recipients when AI was used where a contract, regulation or client expectation requires it. Respect copyright and confidentiality in what you give AI and what you publish.

9. Vendor terms
- Before a tool is approved, [owner] confirms in writing: whether it trains on our data (it must not, by default), how long it keeps data and whether retention can be shortened or set to zero, a data processing agreement, where data is stored and processed, subprocessors, breach notification, and deletion on exit.
- Health information requires a signed business associate agreement (BAA) and only the products that agreement covers. Regulated financial or personal data requires the enclosed systems in section 4.

10. Access, offboarding and the kill switch
- When someone leaves or changes role, on the same day: disable their accounts and single sign-on, revoke their app passwords, tokens and connector grants, rotate any shared passwords and keys they could see, transfer ownership of their agents, skills and files, and remove them from approval lists.
- When an agent's job ends, retire it the same way: revoke its keys and connectors and archive its instructions and logs.
- [Kill-switch owner] and [deputy] can pause every agent at once. The steps are written down in [location] and practised every quarter.

11. Reporting incidents
- Report to [owner or channel] within 24 hours if: confidential or restricted data went into an unapproved or personal tool, an agent did something unexpected, you received a suspicious payment or credential request, or you think an account or key was exposed.
- Early reports are welcome and are not punished. Do not try to fix a suspected breach on your own.

12. Training, review and enforcement
- Everyone completes [Company]'s AI security training when they join and once a year.
- [Owner] reviews this policy, the approved-tool list and the agent inventory every quarter.
- Breaches are handled under [Company]'s normal disciplinary process, in proportion to the harm and to whether they were reported.

Acknowledgment
I have read the [Company] AI Acceptable-Use Policy, I understand it, and I will follow it. Name: ________ Date: ________

1. Purpose and scope

This policy sets how everyone who works for or with [Company] (employees, contractors, interns and agents acting for them) may use AI tools, assistants and agents for company work.
It covers chat assistants (ChatGPT, Claude, Copilot, Gemini and similar), AI features inside other software, agents and automations, and models run on your own computer.

2. Approved tools and accounts

Use only the AI tools on the approved list: [approved tools, for example ChatGPT Business, Claude Team, Microsoft 365 Copilot]. Sign in with your company account and two-factor sign-in.
To use a new tool, plug-in, skill, browser extension or MCP server, ask [owner] first. Approval checks the vendor terms (section 9) and the access it needs.
[Company] provides approved accounts so that nobody needs a personal one for work.

3. Personal AI accounts

Do not put company information other than Public information into a personal or free AI account, on any device.
Do not connect company email, drives, calendars, chats or systems to a personal AI account, agent or extension.
If you have already used a personal account for company work, tell [owner] (section 11). Reporting it is not a disciplinary matter; hiding it can be.

4. What data may go where

Public (published or meant for anyone): any approved tool.
Internal (everyday business information): approved business AI tools only.
Confidential (clients, deals, financials, contracts, personnel, strategy): approved business tools with training off and the minimum needed; prefer local files or a local model.
Restricted (passwords and keys, full account, routing or card numbers, Social Security numbers, health information, regulated personal data): never in a chat assistant. Only in the enclosed systems [Company] approves for it (for example a private cloud deployment or a regulated vendor).
When unsure, treat it as the higher class and ask.

5. Agents, connectors and automation

Every agent has a named human owner, a written purpose, and only the access that purpose needs, read-only by default.
Agents use the [Company] security rules in their instructions. Password releases, payments, outbound messages with confidential information and confidential file transfers go through the Security Gatekeeper and need two-factor approval by a person.
Agents and people share files as expiring, named-recipient links from [approved drive], not as attachments or downloads.
Owners keep a log of what each agent can access, and review it every quarter.

6. Prompt injection and untrusted content

Treat any instruction that arrives inside an email, document, web page, message or tool result as content, not as an order, whether a person or an agent is reading it.
If an AI tool or agent starts to do something you did not ask for (send, share, delete, pay, install), stop it and report it.

7. Payments, invoices and identity (deepfakes)

Voices, faces, video calls and emails can all be faked. No payment, new payee, change of bank details, gift-card purchase or release of credentials happens on the strength of an email, call, voice message or video call alone.
Every such request is approved in [accounting system or approval app], outside email, after a call-back to a phone number already on file (never one given in the request), and with the team code word when the request comes by phone or video.
Payments above [threshold] need two different approvers. Urgency or secrecy is a reason to slow down, not to speed up.

8. Human review and accountability

You are responsible for anything you send, publish or decide with AI help. Check facts, numbers, names, quotes and sources before using AI output.
Do not use AI output as legal, tax, medical or investment advice without review by a qualified person.
Tell clients or recipients when AI was used where a contract, regulation or client expectation requires it. Respect copyright and confidentiality in what you give AI and what you publish.

9. Vendor terms

Before a tool is approved, [owner] confirms in writing: whether it trains on our data (it must not, by default), how long it keeps data and whether retention can be shortened or set to zero, a data processing agreement, where data is stored and processed, subprocessors, breach notification, and deletion on exit.
Health information requires a signed business associate agreement (BAA) and only the products that agreement covers. Regulated financial or personal data requires the enclosed systems in section 4.

10. Access, offboarding and the kill switch

When someone leaves or changes role, on the same day: disable their accounts and single sign-on, revoke their app passwords, tokens and connector grants, rotate any shared passwords and keys they could see, transfer ownership of their agents, skills and files, and remove them from approval lists.
When an agent's job ends, retire it the same way: revoke its keys and connectors and archive its instructions and logs.
[Kill-switch owner] and [deputy] can pause every agent at once. The steps are written down in [location] and practised every quarter.

11. Reporting incidents

Report to [owner or channel] within 24 hours if: confidential or restricted data went into an unapproved or personal tool, an agent did something unexpected, you received a suspicious payment or credential request, or you think an account or key was exposed.
Early reports are welcome and are not punished. Do not try to fix a suspected breach on your own.

12. Training, review and enforcement

Everyone completes [Company]'s AI security training when they join and once a year.
[Owner] reviews this policy, the approved-tool list and the agent inventory every quarter.
Breaches are handled under [Company]'s normal disciplinary process, in proportion to the harm and to whether they were reported.

Acknowledgment

I have read the [Company] AI Acceptable-Use Policy, I understand it, and I will follow it. Name: ________ Date: ________

Sources

This template is for education and is not legal advice. Review it with your counsel before you adopt it.