Nobody Tells You That an AI Agent Needs a Manager

There’s a sentence in almost every AI pitch deck right now: the agent handles it for you.

It’s not a lie exactly. But it leaves out the part that decides whether the whole thing works — and if you’re a small business about to spend real money, that missing part is the one you most need to hear.

An AI agent is not an appliance you switch on. It’s closer to a new starter who works very fast, never gets tired, and occasionally does something confidently wrong without noticing. New starters need supervision. So do agents.

What the 2026 data actually says

This isn’t a hunch. The numbers that came out this year are unusually blunt about it.

Gartner expects that by the end of 2026, around 40% of business applications will ship with some kind of task-specific AI agent built in. Adoption is real and fast.

But the same firm also predicts that more than 40% of agentic AI projects will be scrapped by the end of 2027 — not because the technology failed, but because of costs that climbed, value that was never clearly defined, and controls that weren’t in place.

McKinsey’s work this year found something even more uncomfortable: a large majority of firms using AI report no measurable effect on their bottom line at all.

So AI agents work. And most projects still don’t. The gap between those two facts is where your money either goes or gets wasted.

The hour nobody budgets for

Here’s the specific thing that gets left out of the pitch.

Analysis of agent deployments this year found that the operators who succeed spend roughly sixty to ninety minutes a day on their agents — reading outputs, catching misfires, adjusting instructions, tightening the rules.

Read that again alongside the sales pitch. You were told the agent frees up your team’s time. The reality is that someone’s day now includes managing the agent.

That’s not an argument against doing it. For the right task, you’re trading three hours of repetitive work for one hour of supervision, and that’s a good trade. But if you budgeted for zero hours of supervision, your maths was wrong before you started — and that’s precisely why so many projects get quietly cancelled a few months in.

Why agents drift

The pattern people describe is remarkably consistent: it works on day one, gets a bit odd by month three, and breaks by month six.

This surprises people because software doesn’t normally behave that way. A spreadsheet formula written in 2019 still does the same thing today.

An agent is different for three reasons.

Your business changes and the agent doesn’t know. You add a service. You change your prices. You stop covering a postcode. The agent carries on answering from what it was told months ago, confidently, to real customers.

Errors compound across steps. A single question answered slightly wrong is a small problem. An agent that performs eight steps, where step two was slightly wrong, produces a result that’s wrong in a way nobody can easily trace. This is why agents do well at short, well-defined jobs and badly at long chains with no checkpoints.

Mistakes don’t look like mistakes. When normal software breaks, it stops and shows an error. When an agent gets it wrong, it produces something that looks completely plausible — same tone, same confidence, same formatting as its correct work. There’s no red light. Someone has to actually read it.

That last one is the real reason supervision can’t be skipped. You’re not watching for crashes. You’re watching for confident nonsense.

What this changes about how you should buy

If supervision is part of the deal, then a few things that sound like details turn out to be the whole decision.

Pick a task you can check in seconds. If verifying the agent’s work takes as long as doing it yourself, you’ve built a very expensive hobby. The best first candidates produce output you can glance at and immediately judge — a draft reply, an extracted invoice total, a categorised enquiry.

Make sure something writes down what it did. If you can’t see what the agent did and why, you can’t fix it when it drifts. Ask this before you buy, not after.

Insist on checkpoints in longer jobs. A five-step process with a human confirmation at step three is dramatically more reliable than the same process running straight through. Slightly less impressive in a demo. Far better in practice.

Decide who owns it. Not who bought it — who checks it. If the answer is “we’ll see”, the agent will drift, because every agent does. One named person spending twenty minutes a day is the difference between a tool that pays for itself and a project that gets cancelled.

Start with one. Research this year suggests individual productivity actually falls once people are juggling more than about four AI tools at once. Three agents each half-managed will beat you. One agent properly managed will not.

The honest version of the pitch

Here’s what an AI agent actually offers, stated plainly:

It will take a repetitive, rule-shaped task off your hands and do it faster than you can, around the clock, without forgetting. In exchange, it needs a clearly defined job, a way for you to see its work, and a person who looks at it regularly and corrects it when your business moves on.

That’s a genuinely good deal for the right task. It is a terrible deal if you were promised that nobody would ever have to think about it again.

The firms getting real value from agents this year aren’t the ones with bigger budgets or better technology. They’re the ones who mapped a process properly before automating it, picked a narrow job rather than a grand one, and gave someone responsibility for watching it.

None of that requires a large company. All of it requires being told the truth before you start.

Before you spend anything

Ask whoever is selling you an agent three questions:

What exactly will this agent do, and what will it refuse to do? A vague answer here predicts a vague result. How will I see what it did? If there’s no straight answer, you’ve found your future problem. How much of someone’s time will this need each week, once it’s live? Anyone who says “none” is either inexperienced or not being straight with you.

If you’d like a second opinion on something you’re considering — or an honest answer about whether a particular task is worth automating at all — tell us what you’re looking at. Sometimes the useful answer is that it isn’t, and we’d rather say so.