The message that arrives at nine at night
A gym owner I know gets a WhatsApp most evenings around nine: "Is there a spot in tomorrow's 7am class?" For years she answered these herself, phone propped on the kitchen counter, checking the roster from memory. Lately something drafts the reply for her: it checks the actual booking sheet, writes "Yes, see you at 7, bring a mat," and waits. She reads it, taps send, gets her evening back a minute sooner than she used to.
It could go one step further. It could just send it, no tap required. A few products already work exactly that way. Before you buy anything with the word "agent" on the box, that's the real question hiding under the marketing: does a human see this before it goes out, or not? The two answers describe two genuinely different products, and for most small businesses, the less flashy one is the right one.
Two shapes wearing one label
"AI agent" gets used for both. One shape drafts and waits: it prepares the reply, the invoice, the reorder, and stops there until you say go. The other shape acts: it sends, books, charges, with no pause for a human to look. Same word on the packaging, very different thing happening underneath it.
It's tempting to think of the second shape as the "grown-up" version, the one you graduate to once you trust the tool enough. I don't think that's right. They solve different problems. The drafting shape solves "I don't want to type this myself." The acting shape solves "I don't want to be involved at all." Almost nobody running a business with a real name and real customers actually wants the second thing for anything that could embarrass them if it went wrong.
What's actually getting used
Zapier looked at real adoption across the tools it connects and split AI's role into four buckets: Communicator (writing replies and summaries: 84% adoption), Clerk (pulling data into records: 79%), Analyst (making a yes/no call on its own: 44%), Coordinator (kicking off tracked work with no review step: 25%). The first two involve a human seeing the output before anything happens. The second two don't. Their own advice, not mine: build the reviewed version first, "because if AI gets something wrong, a human has a chance to catch and correct it." That's also where the actual usage is concentrated.
That split is worth sitting with. The most "advanced"-sounding use of AI (the kind that acts entirely on its own) is also the least used, by a wide margin. Not because businesses haven't caught up yet. Because most of them tried it, felt the risk, and pulled the reviewing step back in. If you're wondering how much of your own day that reviewing step should actually cost you, I wrote about exactly that here: Do I Really Have to Check Everything My AI Does?
The trust problem is closer than it looks
This week Anthropic's own CEO, Dario Amodei, called the public backlash against AI "fundamentally a crisis of trust." His point: people don't trust institutions in general, AI companies included, and no amount of careful messaging fixes that; only results do. He's talking about an industry-sized problem. Yours is smaller and far more personal: does the person who gets your reply, your invoice, your booking confirmation, trust that what came from your business is actually right?
That trust doesn't come from a privacy policy or a press release. It comes from the fact that, when it mattered, someone looked before it went out. A drafted-and-approved message can still be wrong. I'm not claiming otherwise. But a wrong draft gets caught at your desk. A wrong autonomous action gets caught by your customer, after the fact, which is a much worse place for either of you to find out.
Where full automation is actually fine
None of this means every AI action needs a human tap. If you're using AI to flag which of five hundred lead-form entries look like spam, nobody's reviewing five hundred entries one by one. That's the whole point of automating it, and getting a handful wrong costs almost nothing. The line isn't "review everything" or "review nothing." It's whether a wrong output, sent unreviewed, would cost real money, damage a real relationship, or embarrass you if a customer saw it. High-volume and low-stakes: let it run. Anything that touches someone else's money, health, or evening: keep the tap.
One question before you buy
Before you adopt anything calling itself an AI agent, ask the seller one thing: what does a human see and approve before this acts, and can I turn that review on for the parts that actually worry me? If the honest answer is "nothing, it just acts," that's not a feature to be impressed by for anything customer-facing. It's the one thing to ask them to add before you trust it with your name.