Playbook · Policy

An AI agent policy you can build in five minutes

Staff using AI tools is one question. Agents acting in your accounts is another: they refund, deploy and email on your behalf. A short written policy says what they may do alone, and what waits for a person.

Updated November 2026

In short

An AI agent policy lists the agents you run and the systems each may reach; separates what an agent may do alone (read, search, draft, prepare) from what waits for a person’s approval (moving money, deleting, publishing, messaging customers, changing access); says how keys are kept (in a vault or broker, never in files, never a person’s own login); and sets what is logged, who owns each agent, and what happens when one gets something wrong. Keep it to one page and review it twice a year.

Build your AI agent policy

Pick what your agents work in and how strict to be. The policy writes itself; copy it into your handbook and adapt it.

Where your agents work

What waits for a person

Keep the logs for

Review the policy

Your company: AI agent policy

This policy covers the AI agents that act in Your company’s tools with their own access: payments and invoices and email. Staff use of AI tools is covered separately.

1. Owners

Every agent has a named owner who answers for it. Agents that aren’t on the list get no access.

2. What agents may do alone

Read, search, summarise, draft and prepare changes.

3. What waits for a person

  • These always wait for a named person’s approval: moving money, deleting data, publishing, messaging customers and changing access.
  • Refunds, credits and invoices are proposed with the amount, the reason and the evidence.
  • Agents draft replies; a person sends them.

4. Keys

Keys are kept in a vault or broker, never in files, chats or tickets, and never a person’s own login. Each agent gets only the access its job needs.

5. Records

Agent access and every approval are logged and kept for one year. Each owner reads their agent’s log regularly.

6. When something goes wrong

Stop the agent, revoke its keys, read the log, undo what can be undone, tell the people affected, and change the rule that let it happen.

7. Review

This policy is reviewed every six months.

How it differs from a shadow-AI policy

A shadow-AI policy is about people: which AI tools staff may use and what data may go into them. An AI agent policy is about software that acts: agents with their own access to your payment provider, website, code or inbox. Most companies need both; this one covers the agents.

The five decisions it writes down

Every line of the policy above comes from one of these. Decide them once, and new agents fit into the rules instead of making their own.

  • Which agents run, who owns each one, and which systems each may reach.
  • What an agent may do alone: read, search, draft and prepare.
  • What always waits for a named person: money, deletions, publishing, messages to customers, changes to access.
  • How keys are kept: in a vault or broker, never in files or chats, never a person’s own login.
  • What is recorded, for how long, and who reads it.

Roll it out in a week

A policy only helps once it matches what is running. Five short steps get you there.

  • Day 1: list every agent and every key it uses.
  • Day 2: move keys out of files and into a vault or broker.
  • Day 3: switch the riskiest actions to approval.
  • Day 4: name an owner for each agent.
  • Day 5: share the page, and put the review date in the calendar.

Checklist

Before an agent gets a new key

  • The agent is on the list, with an owner.
  • The key gives only the access the job needs.
  • The key lives in a vault or broker, not in a file.
  • Irreversible actions wait for a person.
  • Its access is logged, and someone reads the log.

How MoltenRock puts it into practice

The MoltenRock apps make several of these rules the default on a Mac: keys sealed in a vault and released per agent, business tools read for the agent without handing it a key, and changes in Stripe, Cloudflare and GitHub that wait for your approval.

Questions

Do we need an AI agent policy if we have a shadow-AI policy?

Yes, if any agent acts in your company’s tools with its own access. The shadow-AI policy covers people and the data they share; the agent policy covers what software may do on its own.

How long should it be?

One page. A policy nobody reads protects nobody; the builder above keeps it short.

Is this legal advice?

No. It’s a practical starting point. For regulated work, have it reviewed by someone who knows your obligations.

A practical starting point, not legal advice.

More guides