An AI texted me last week: "I saw you missed your 8:30 meeting with James. Did you send him a follow-up?"

I had missed the meeting. I hadn't sent anything. I typed "no, can you?" and it did.

I didn't start that conversation. In three years of using AI every day, that was the first time it came to me.

The same week, I asked the same agent, Instinct (I have 5 invites so email me if you need one), to help with a trip over the worst travel week of the year, right after Christmas. It did the research, found places with no visible availability online, drafted an email to each reservation desk, and queued them in my Gmail. All I did was hit send. Then it watched my inbox, noticed one place never replied, and nudged me to follow up.

Around the same time I was shopping for a pull-up bar. I started where I always start, with Google and Google Shopping. Then I asked Muse, Meta's new personal agent, the same question. It came back with three types of pull-up bar that matched what I wanted better than anything my Google searches had found.

Some people have gone a step further and given their agent a credit card so it can finish the job and buy. I haven't done that yet, but I'm close to turning it on.

If you run a company, two things in those stories matter more than the hotels. The agent acted first. And I'm now weighing what to hand it. Both are about to show up at work, with other people's data attached.

At most of the companies I work with, nearly everyone uses AI now. Most of them use it a few times a week.

The weekly users aren't against it. Today's AI makes them do the work of getting started. They have to notice that a task could use help, open a separate tool, figure out what to ask, paste in the context, and carry the answer back to wherever the work happens. For a busy person, each of those steps is a reason to do it the old way.

The daily users turned all of that into a habit. Most people never will use AI proactively. Three years in, that's where adoption has stalled at most companies I see: everyone has access, and a minority gets most of the value.

I used to think the fix was more training or give it more time. The text about James changed my mind.

Sonia Baschez put it better than I could last week:

The AI most employees have today is the husband in her comparison. Go find flight options. Draft a reply. Useful, but you still have to think of the chore, explain it, and check that it got done. That's exactly the work the few-times-a-week crowd skips.

The wife carries the mental load (at least in my household): the birthday gift nobody has bought, the diapers running low, the dentist appointment someone forgot. Nobody hands her the task. She notices it.

The text about James was the first time I saw an agent move from the first category toward the second. Nobody told it to check whether I'd followed up.

I've been trying four of these personal agents: Instinct, Muse, Dots, and GrokBot. Instinct and Muse have surprised me the most. Dots just launched. They don't sit in a chat box waiting for me to come back. They keep working after I've moved on to something else.

Now put that agent inside your company.

It's always on. It has the business context: the inbox, the calendar, the CRM, the shared drive, the notes from last week's meeting. And it can use a computer the way your people do. Fill out the vendor portal. Pull the report from the system nobody ever built an integration for. Post to your project management system even though IT denied the MCP integration.

For the employee who opens AI twice a week, the job shifts from starting the work to approving it. The agent sees the email that needs a response and drafts the reply. It notices the customer who hasn't heard back in a week. It fills in the expense report from receipts already sitting in the inbox. It pings you about the meeting you missed before the other person has to. The employee says yes, no, or fix this.

That's a much smaller ask than "remember to open the chatbot." It's the first version of AI I've seen that could reach the people who haven't changed how they work yet, and at most companies that's the majority of the payroll.

I don't know exactly when it will be reliable enough to trust. Given how quickly the personal versions have improved, I'd be surprised if some form of it isn't showing up at work within three-to-six months.

Which brings me back to the credit card.

At home, the question is whether I trust an agent with my card. At work, the question will be what you trust it with: the inbox, the CRM, the ERP login, the ability to send a customer an email nobody read first.

Those decisions will arrive one system at a time, probably because an employee asks for access rather than because anyone planned it. I'll make my credit card call on my own. Your version involves customers, vendors, and systems full of other people's information, and it's closer than most roadmaps assume.

For three years, getting value from AI meant remembering to ask. I think that's about to change. The few-times-a-week employee may end up gaining the most, because they won't have to go looking for it. The part that falls to you is deciding what the agent gets to see.

Two things I'd genuinely like back from you. First, hit Reply and tell me which part of this matches what you're seeing. Second, if there's someone in your world who should be on this list, forward it to them.

- Robbie