The AI Agent Safety Checklist
What to check before you let an AI tool act on its own - send e-mails, update records, book meetings or spend money: what it can touch, what needs a person's OK, prompt injection, logs, an off switch, and an agent register.
Most AI use so far drafts things a person reads before anything happens. Agents are different: they send the e-mail, update the record, book the meeting, place the order. A mistake - or a trick hidden in an e-mail the agent reads - goes straight into the world.
An AI that writes can be wrong. An AI that acts can be wrong at scale. This checklist is for the moment you let it act.
What's in it
- How much can it do? Four steps from "drafts" to "acts alone", and why most agents should stop at step 3.
- What it can touch: its own identity, only the access the job needs, and money, people and customer data off limits by default.
- What it can do without asking: what always needs a person's OK - anything leaving the company, money and deleting - and the limits to set.
- What it reads, and who it listens to: prompt injection in plain English, and how to stop an agent taking orders from an e-mail.
- Common agent jobs, and where to start them: inbox sorting, booking meetings, customer questions, the CRM, chasing invoices and more.
- Watching it and stopping it: logs someone actually reads, alerts, and an off switch you've tested.
- When it gets something wrong: who's responsible, putting it right, and treating it as an incident.
- The agent register: one row per agent, with its owner, access, approvals, limits and the date the off switch was last tested - plus the red lines and a before-go-live check.
Who it's for
Owners, operations leaders and IT managers at small and growing companies whose people are starting to connect AI tools to e-mail, calendars, the CRM or other systems - or whose software vendors are switching "AI agents" on for them.
Set the rules first with The AI Use Policy Starter, and check the tool with The AI Vendor Questions. Then try it with a small group using The AI Pilot Playbook. If something does go wrong, The Incident Response Plan covers the first hours.
It's a practical checklist, not a security audit or legal advice.
It's a 4-page PDF, made to print and write on.
Want them all? The Complete Collection has every guide in the store in one download, for less than a third of the price.
Reviews
No reviews yet.
Bought this? Log in to review it.
Log in