Log in
Subscribe
Jun 30, 2026
June 2026: The meter, the badge, and the kill switch
June 2026: The meter, the badge, and the kill switch
00:00
08:16
Transcript
0:00
This is AI at Work, the monthly. I'm Matthew Kessler, and this is June. Here's the thing about June. It wasn't a feature month.
0:08
It was the month the agents got a price tag, a badge, and a kill switch all in about three weeks, and it was the month the limits showed up too.
0:17
A government reached in and switched off the most powerful model on the market, and a federal court drew the first hard line around what an AI is actually allowed to decide.
0:27
If that sounds like a lot, it is, but it's all one story. Every headline this month is the same headline told four ways.
0:36
Artificial intelligence took another step from being a thing you open toward being a thing that runs on its own, and every single time it took that step, somebody quietly moved the risk onto you. Four stories.
0:47
I'll tell you what's genuinely good about each one, and then I'll tell you the part nobody put in the press release. Start with the meter.
0:55
In a single month, Microsoft and OpenAI both moved their agents onto consumption billing.
1:00
Microsoft's Work IQ, the context layer underneath Copilot, went generally available on June sixteenth, and the agents that run on it now bill in something called Copilot credits.
1:11
OpenAI's Workspace agents start charging on July sixth at roughly five to twenty-five credits a run.
1:18
And Copilot Cowork, the feature that does the long multi-step work for you, went generally available the same week, metered at a cent per credit. Here's the good part, and it's real. This is honest pricing.
1:31
You pay for the work you actually use, not a flat seat fee for software half your team never opens. Microsoft even shipped a cost dashboard in the admin center so you can see the spend and set limits. Genuinely good.
1:45
Now, here's the part nobody says out loud. Your AI spend just stopped being a number you can predict. It used to be seats times thirty dollars. Boring, flat, forecastable.
1:57
Now it's a variable bill that moves with how much your people lean on it. The meter is off by default, which sounds safe until you realize it means the burn is invisible right up until the invoice lands.
2:07
And pay for what you use is a very polite way of saying the budget risk just moved from the vendor's side of the table to your finance team's. Good products, all three of them. Watch the meter. Story two, the badge.
2:21
At its build conference this month, Microsoft showed off Scout, an always-on agent that acts without being asked. Scheduled work, flagged decisions, no human in the loop by default.
2:32
The detail that matters isn't what Scout does. It's that every Scout runs under its own identity in your directory. Microsoft gave the agent a badge, a directory entry, the same kind your employees have.
2:46
And here's the good part, and give Microsoft credit because they saw it coming. Scout's access is scoped per task. Sensitive actions require a human to sign off.
2:56
The company that built an agent to act on its own also built it to stop and ask a person. That's not caution slowing down progress. That's the org chart asserting itself on purpose.
3:07
But put June together and look at what you actually have now. Agents have identities. They have salaries, those co-pilot credits, five to twenty-five a run. They have standing access to your systems.
3:19
And as we're about to see, they have an off switch someone above your pay grade can reach. That combination has a name, and the name is not feature, it's staff. You don't enable an employee, you onboard one.
3:33
And right now, most organizations are hiring dozens of these things with nobody named as the manager. So here's a test, and it takes ten minutes.
3:41
For every agent running in your tenant, write down four things: the identity it runs under, the data it can reach, what it spends, and the human who reviews its work. Every blank cell on that list is your exposure.
3:56
Story three is the short one, and it's the one that should make you a little uncomfortable. On June ninth, Anthropic launched Fable-five, the most capable model it had ever shipped.
4:07
Three days later on the twelfth, the US Department of Commerce sent a letter ordering Anthropic to suspend all access to that model for any foreign national anywhere in the world.
4:17
Anthropic couldn't tell foreign nationals from everyone else in real time, so it shut the model off for everybody worldwide. Whatever your team built on Fable-five that week stopped working that Friday.
4:29
The good, if you can call it that, is the lesson delivered for free. Vendor continuity is not a hypothetical you put in a slide and forget. It's real, and it can arrive on a government's timeline, not yours.
4:42
The catch is everything else. The most stable-looking AI vendor on the market can have its best product switched off by a directive you'll never see, with no warning and no restoration date.
4:53
And the same week, Anthropic filed to go public, so your AI supplier is now a regulated, publicly traded company with all the constraints that brings.
5:03
If something load-bearing in your stack runs on one model from one vendor, June just told you what that costs. Check your continuity clauses. Keep a fallback.
5:13
Which brings me to the one story this month that pushed the other way. While the tools were moving deeper into your work, a federal court was quietly drawing a line around them.
5:23
In a ruling that surfaced this month, a judge in Manhattan held that an executive's conversations with an AI where he worked out his own legal strategy were not protected by attorney-client privilege.
5:35
The reasoning was blunt. The AI is not a lawyer, and the platform's own terms say it can keep and disclose what you type. Sit with that for a second 'cause it's the counterweight to everything else in this episode.
5:48
All month, the message from the vendors was, let the agent do more. Let it run. Let it reach into your contracts and your calendar and your inbox. And then a court said, "Fine But understand what it is.
6:03
It is not your counsel. It is not your judgment. And the things you type into it are not protected just because they felt private.
6:11
Here's the line that ruling draws, and it's the most useful sentence I can give you this month. There are two kinds of work: retrieval and judgment.
6:19
Retrieval is find me the clause, summarize the term, list the obligation. Judgment is, is this enforceable? Should we sign? What's our position? AI is genuinely remarkably good at the first one.
6:34
It has no standing in the second. Ask it to retrieve, and you get thirty minutes of your life back. Ask it to decide, and you've created a record of a judgment a court can read back to you later.
6:46
So put the whole month on one table. The agent's got a meter. They got a badge. They got a kill switch. And a court reminded everyone they don't get the final word. Four stories, one shape.
6:58
AI took another step toward deciding for you, and at every step, the risk, the cost, the exposure, the call got handed back to a human, usually you. That's not a coincidence, and it's not a problem to be solved.
7:11
It's the actual job now. The vendors are very good at building the prepare. They have quietly, repeatedly left you the decide.
7:19
The scout that escalates to a person, the dashboard that makes you set the limit, the court that says you're still the lawyer. Every one of them is the same instruction. The machine prepares, and you decide.
7:31
You have two tests this month. For your agents, the staffing list. Identity, access, spend, manager. For anything high stakes, retrieve, don't opine. That's it. That's the work. That's June.
7:45
If you want the detail, the sources, the dates, the prompts I left out of this, it's all in the two AI at Work issues from this month in your inbox or on the site. This show is the monthly companion.
7:58
One month, one through line, no filler. If it was useful, follow it wherever you're listening and forward it to the one person in your org who's about to approve an agent budget without asking who the manager is.
8:10
I'm Mathieu. The machine prepares. You decide. See you next month.
AI at Work (the monthly audio companion to the AI at Work newsletter)
Listen on
Apple Podcasts
Apple Podcasts
Spotify
Spotify
YouTube
YouTube
Pocket Casts
Pocket Casts
Overcast
Overcast
Amazon Music
Amazon Music
Recent episodes
No results found
