There is a particular kind of optimism going round at the moment. Not AI that answers questions, but AI that takes actions. Agents that book the meeting, chase the invoice, update the record, reply to the customer. The demos are genuinely impressive, and for a stretched business the appeal is obvious. Work that runs itself.
Then there is a number from Gartner that is worth sitting with. They expect more than 40% of agentic AI projects to be scrapped by the end of 2027. Not paused. Cancelled.
It would be easy to read that number as a verdict on the technology. It is not. Agents work. The projects that get quietly switched off rarely fail because the model could not do the task. They fail because nobody decided, up front, what the agent was allowed to do, who was answerable when it acted, and how anyone would notice when it got something wrong.
That is a governance problem, not an engineering one. And governance has an image problem of its own, because the word brings to mind a forty-page policy and a committee that meets quarterly. For a business of thirty or fifty people it is far smaller and more practical than that. It comes down to three questions you answer before the agent goes anywhere near a real customer or a real ledger.
What is it actually allowed to do
An agent with loose instructions and broad access is the digital version of handing a new starter the keys to everything on day one and hoping for the best. Scope means writing down, in plain terms, what it may do and what it may not. It can draft the reply, not send it. It can flag the overdue invoice, not change the amount. It can read from the system, not delete from it.
Narrow is not timid. Narrow is what makes the thing safe enough to leave running while you are not watching.
A human where the stakes are real
Not every action needs a person checking it, and putting one in front of everything just rebuilds the manual work you were trying to remove. The skill is picking the few moments where a mistake would actually cost you, and stopping only there.
An agent that sorts incoming emails into folders can run unattended all day. An agent that issues refunds should pause for a human on anything above a figure you set. The point is to decide where that line sits on purpose, rather than finding out where it should have been after something has already gone out the door.
A record of what it did
If you cannot answer “what did the agent do last Tuesday, and why,” you do not have an agent you control. You have one you hope is behaving.
A simple, readable log of every action it took is what lets you catch a problem early, explain yourself to a customer or a regulator if you ever have to, and improve the thing over time. It is dull to set up. You will be grateful for it the first time something looks off and you can actually go and check.
An agent you cannot audit is not one you control. It is one you hope is behaving.
The risk is quiet, not dramatic
What makes an ungoverned agent dangerous is not some cinematic failure. It is a quiet one. When a person makes a mistake, it tends to get noticed, because people work at human speed and other people are around. An agent making the same mistake at two in the morning, across four hundred records, does not get noticed until the month is closed, or the customers have already had the email, or the figures everyone trusted turn out to have been wrong for a fortnight.
Speed is the whole point of an agent. It is also exactly why a small error becomes a large one before anyone looks up.
Start with one you can trust
So we would rather help a business put a single, tightly-scoped agent into production that everyone actually trusts than a dozen clever ones nobody dares leave running. The trusted one earns its keep and quietly takes on more. The clever ones sit switched off, and eventually become part of that 40%.
Before the agent, then: decide what it may and may not touch. Put a person at the one point where a mistake would genuinely hurt. Keep a record you could audit later. None of it is glamorous, and none of it shows up in the demo. All of it is the difference between an agent you come to rely on and one you end up quietly turning off six months in.
It is the part we build in from the first conversation rather than bolt on at the end. If you are weighing up where an agent could safely take work off your hands, that is exactly what the free diagnostic is for.