You Didn't Hire These Agents. They Came With the Subscriptions.

So in the first nine days of July, four different companies put an agent inside software you're probably already paying for.

July 1, Google made Gemini Spark available on the Mac, where it can act on local files in whatever folders you point it at. Same day, Notion shipped External Agents in 3.6. Same day, Microsoft's new pricing took effect and Copilot stopped being a promotional add-on, becoming a permanent part of the Microsoft 365 lineup, with Copilot Chat picking up inbox and calendar awareness down in Business Basic and Business Standard. Then on July 9, OpenAI launched ChatGPT Work, which plugs into Slack, Gmail, Drive, SharePoint, Salesforce, GitHub and a couple dozen other things, takes an outcome instead of a prompt, and runs tasks on a schedule.

Nobody on your team asked for any of that. It arrived in tools you'd already bought.

This isn't app sprawl again

We've written about app sprawl plenty. 42 subscriptions, half of them barely used, a few nobody remembers signing up for. Real problem. But it's a passive one. A subscription you forgot about doesn't do anything except charge you.

An agent is not passive. It reads. It writes. It sends. It runs Tuesday at 6am because somebody set that up in March and then left the company.

Gartner called this back in August 2025: 40% of enterprise applications would have task-specific AI agents by the end of 2026, up from under 5% the year before. That's landing more or less on schedule. What the prediction didn't say, and what actually matters if you run a 25-person company, is that this doesn't arrive as a purchase decision. It arrives as a product update.

The count gets away from you fast

Gartner's April numbers on this are Fortune 500 numbers, and they are not your world: they project the average global Fortune 500 will be running over 150,000 agents by 2028, up from fewer than 15 in 2025.

Ignore the size of that. Look at the direction. Fewer than 15 to a five-figure count in three years, and not because anybody went shopping. Because every application grew its own.

Now scale it down honestly. You've got 30 tools. Some meaningful share of them ship an agent this year. A few come on by default. A few more come on for whoever clicks the button in the corner, which is not a decision anyone wrote down. So by the end of the year a 20-person company is running more agents than it has people, and it interviewed exactly none of them.

That'd be fine if they coordinated. They don't. We wrote about that in March, back when the interesting part was that none of these things could see past their own app. Nine months later the interesting part is different. Now they act.

Nobody's watching, and that part is measurable

The number I'd put on the wall is in that same Gartner release: only 13% of organizations think they have the right AI agent governance in place.

The shadow version is worse. Forbes reported last week on a Black Fog study finding 49% of employees using unapproved AI tools, and 58% of those on the free versions. Free versions being the ones with the weakest data handling.

So the honest picture at a normal small company right now is some number of agents you approved, a larger number that showed up in updates, and an unknown number your team signed up for personally and pointed at customer data.

That Forbes piece put the cause well: small businesses adopted AI one subscription at a time, and nobody wrote a rulebook, because as long as the thing lived in a browser tab there wasn't much to write rules about.

It doesn't live in a browser tab anymore. It has your Drive.

Shutting it down is the expensive answer

The instinct is to lock everything, write a policy, wait for this to settle.

That costs more than it looks like. The businesses that got AI working are pulling ahead on actual hours, not vibes. In Bluevine's survey of 942 owners, 48% of the ones using AI are getting back at least half a workday every week. You don't give that up because the inventory is messy.

So the move isn't fewer agents. It's knowing what each one can see.

That's the layer we're building. Liaison connects on top of the tools you already run, so there's one place that knows what's connected, what data each connection can actually reach, who authorized it, and what got done with it. Your team keeps their tools. The agents keep working. You stop guessing. And because it's a layer rather than a destination, nothing has to move for that to be true.

The question was never whether to allow agents. That got decided for you in a product update on July 1.

The question is whether you can name what's running in your business right now. Most people can't. That's the gap worth closing.

If you want to find out what's actually connected to what, let's talk.