Even the Companies Betting Hardest on AI Agents Still Keep a Human in the Loop
Even the Companies Betting Hardest on AI Agents Still Keep a Human in the Loop
The 2026 pitch is that your business runs itself. An agent reads the enquiry, writes the reply, books the job, chases the invoice, and you hear about it afterwards.
Zapier went and asked the companies actually doing this. The answer is quieter than the marketing.
In its State of Agentic AI Adoption survey, Zapier found 72% of enterprises are using or testing AI agents. That's real adoption, not curiosity. But when it asked how those agents are managed, the most common answer was human-in-the-loop, at 38%. Approval gates built into the workflow. The agent drafts, a person checks, and only then does anything reach a customer. Just 20% said their AI systems run autonomously with minimal oversight.
Worth being straight about who was asked. Zapier surveyed 525 US executives at companies with 1,000 or more staff, in October 2025. That's not a small service business. But it's the group with the deepest budgets, the most engineers and the most pressure to look like they're ahead. If anyone were going to hand over the keys entirely, it would be them. They haven't.
Oversight isn't cold feet
There's a habit of reading a finding like this as hesitation. Big companies being slow. Risk committees getting in the way.
That's not what it is. Approval gates are a design decision, and usually the right one. Anything that touches a customer carries consequences an agent can't see: a pricing exception, a client who's already annoyed, a job that can't actually be done next Tuesday. Judgement is the expensive part of the work. Keeping a person there isn't a compromise on the way to full autonomy. For most of the work worth automating, it's the finished state.
Which is a relief, because it means the useful version of this is already available to you and costs very little.
Automate the follow-through, not the judgement
The parts of a service business that go wrong are rarely the clever parts. They're the boring ones. The enquiry that arrived on Saturday and got buried. The quote nobody chased. The review that never got asked for. The client who went quiet and nobody noticed for six weeks.
None of that needs an agent with independent decision-making authority. It needs something that remembers, reliably, without being asked.
That's the split worth drawing.
Runs unattended: logging the enquiry, acknowledging it, scheduling the reminder, queuing the follow-up, sending the review request once the job is marked complete, flagging a client who's gone quiet.
Waits for a person: the price, the exception, the apology, the scope change, anything where being wrong costs more than being slow.
The first list is where most of your dropped balls live. It's also the list that can run at 2am with nobody watching, because nothing in it requires an opinion.
A test that saves a lot of arguing
Before automating anything, ask what happens if it fires at the worst possible moment with nobody there.
If the honest answer is "nothing much" - a reminder goes out, a record gets created, a task lands on someone's list - automate it fully and stop thinking about it.
If the answer is "a customer gets the wrong number", or "someone gets a cheerful email in the worst week of their year", it needs a human gate. Not a smarter agent. A gate.
Most owners get this backwards. They automate the sensitive things because those feel impressive, then keep doing the boring admin by hand because it feels like their job. The enterprises in Zapier's survey are doing the opposite, and they have considerably more money to waste on the impressive version.
What to do with this
You don't need full autonomy to get the benefit. You need the follow-through covered, consistently, and a person still making the calls that matter.
Good infrastructure is boring until it's missing. The businesses that feel calm aren't the ones running the newest thing. They're the ones where the small stuff happens whether anyone remembers or not.
If you want a look at what that setup actually involves, here's how Cadence Engine handles it.
Source: Zapier, State of Agentic AI Adoption survey, published December 2025, n=525.











