Before you give an AI agent access
This is the companion guide to the video where I gave an AI agent my real email, my real calendar, and let it send messages as me. It did not go perfectly. This is the checklist I used to decide what to test first, and what I now tell other managers before they hand one of these agents anything real. If you have not watched yet, start there.
Episode 4Watch on YouTubeThe one idea
The question is not whether an AI agent will ever make a mistake. It will. The question is whether it catches its own mistakes, tells you the truth about them, and demonstrably fixes the process, not just the one email. Everything below follows from that.
Step 1: start with something reversible
Give it something low stakes first, so you can see how it handles judgment before you trust it with anything that can’t be undone. Use this as a rough order.
WHAT TO HAND OVER FIRST LOW RISK, START HERE - Sorting or labeling your inbox (nothing is sent or deleted) - Researching a topic and building you a summary or spreadsheet - Drafting a message for you to review before it goes anywhere - Pulling together a briefing from documents you already have HIGHER RISK, WATCH CLOSELY THE FIRST TIME - Sending a message as you, to someone you have not personally approved - Scheduling something on your calendar that other people will see - Editing a shared document other people are already using DO NOT HAND OVER YET - Anything involving money: payments, purchases, refunds, discount codes - Anything that deletes data with no undo - Anything you have not personally watched it do at least once
Step 2: watch the first irreversible action closely
The first time it does something that can’t be quietly undone, sending a message as you, scheduling something other people will see, ask it to show you exactly what it’s about to do before it does it. Not a summary after the fact. The actual content, before it goes.
Step 3: ask the right question after a mistake
When it breaks, and it will, don’t evaluate it on whether it apologizes. Evaluate it on whether it can name the actual cause and show you the fix.
THE QUESTION TO ASK AFTER ANY MISTAKE Don't judge it on whether it apologizes. Ask directly: "What exactly broke? Why didn't your own check catch it before it happened? What have you changed so the same mistake can't happen again? Show me the fix before you continue." A vague answer is the real red flag, not the mistake itself.
Step 4: decide in advance what never gets full autonomy
Pick this before you need it, not in the moment. For me, that’s anything involving money, and anything I haven’t personally reviewed at least once. Write down your own list now.
The full checklist
All four, in one place, exactly what I used.
BEFORE YOU GIVE AN AI AGENT ACCESS 1. Start with something reversible. Inbox sorting, research, drafts. Nothing that can't be undone. 2. The first time it touches anything irreversible, like sending a message as you, watch it closely. Ask it to show you exactly what it's about to do before it does it. 3. When it makes a mistake, and it will, don't judge it on the apology. Ask what broke and what it changed. A vague answer is the red flag, not the mistake. 4. Decide in advance what never gets full autonomy. For me: anything involving money, and anything I haven't personally reviewed at least once.
Ready to deploy an agent your team can actually trust?
We design and deploy AI agent systems with the guardrails built in from day one, not bolted on after the first mistake.
Start a Conversation