Your AI coworker just got its own email address

Here’s what we’re reading and thinking about this week:

Google just gave its new AI agent its own company email address. Its own calendar too, and a spot in the company directory.

A week earlier, OpenAI launched agents called Dots that you message in Slack and Teams that keep working after you log off.

So what happens when everyone at your company can assign your agent work?

Think about what lands in that inbox on day one…

• The office manager forwards a vendor invoice and says "can you handle this?"

• The sales lead asks it to send the usual follow-up.

• The recruiter wants the standard offer letter for someone starting Monday.

A person on their first day would ask "which usual?" and "standard according to who?" Then they'd spend two weeks quietly picking up the answers. The agent won't do either. It only knows what's been written down... and for most teams, that's a lot less than they think.

It will do exactly what it's asked

Here's a lesson we learned the hard way. Someone asked AI to write a project update, and it came back 16 pages long, with a five-page agenda attached.

Nobody could use it. And honestly, the AI did exactly what it was asked. It just had no idea what a good update looks like for us, because nobody had told it.

Now picture that on every request, from everyone with the agent's email address.

👀 

The usual pushback is that these agents are smart enough to figure it out…

But think about the last time someone at work did something dumb. It was almost never a smarts problem, they just didn't know what you wanted. Hand anyone vague instructions and the keys to everything, and they'll try their hardest anyway.

What it looks like when it works

We run one of our inboxes with AI, and last week's numbers are kind of boring, which is the point. About 300 emails came in, and a person needed to see about 60 of them. The AI handled roughly 4 out of 5 on its own, which gave back about 10 hours.

It isn't doing that because it's clever. It's following a playbook (our word for a written process an AI can follow, like a checklist specific enough to run without you). For the inbox, it fits in four lines:

• What starts it: an email lands

• What it needs: the email and our rules

• What it does: reads it, sorts it, then files it or flags it

• What done looks like: filed, drafted, or on someone's list

Rails are what let you walk away

That inbox AI can draft a reply. It can never send one. And the agent that handles email has no access to the calendar at all, so it can't "helpfully" book a meeting nobody asked for.

There are two ways to keep an agent in its lane: tell it not to do something, or just don't give it the tool. The second is a lot harder to get wrong.

That one rule, drafts but never sends, is honestly why we can stop watching it, because on its worst day the damage is a bad draft somebody deletes.

Pick the playbook, not the agent

We've lived through about five AI "revolutions" by now. Whatever Google's or OpenAI's new agents can do this month, every other tool will catch up to soon.

🌶️ Our take: the agent isn't the unlock. The playbook is.

The agent you pick this year will probably get swapped out. The four lines that say how your work gets done come with you. That's Own the Playbook, Rent the Tech in a nutshell.

What to do this week

1. Pick the one request you'd hand a new hire first. The one that shows up every week.

2. Write it in four lines: what starts it, what it needs, the steps, and what done looks like.

3. Add one line of rails: what it can do alone, what it has to ask about, and what it can never do.

What's the first request you'd forward to an AI coworker, and what's the unwritten rule it would get wrong? Hit reply and vent to us 🤣 

LINKS

For your reading list 📚

• 👀 During a test, an AI agent sent a fake murder tip to the Philadelphia police. Anthropic didn't find out for 10 weeks.

• Your customers' AI agents are going to need a front door. Meta, Walmart and Stripe want to agree on how they get in.

• Heads up, Microsoft shops: Copilot agents switch to pay-per-use by default starting Oct 19. Check your spending limit first.

• Your next Webex call might have an AI on the invite list, taking notes and doing the follow-ups. Most of it lands early 2027.

• In case you missed it: Anthropic is spending $100M to train 10,000 people whose whole job is making AI work inside a business. Sound familiar?

That's all!

We'll see you again soon. Thoughts, feedback and questions are much appreciated - respond here or shoot us a note at [email protected]

Cheers,

🪄 The AMP Team (formerly: the AI Exchange Team)