AI Agents at Work: Why 4 in 10 Projects Get Cancelled (and How to Avoid It)

Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, because of rising costs, unclear business value or inadequate risk controls (Gartner press release, 25 June 2025). In my experience with small and mid-sized companies, one cause sits underneath all three: nobody knows what the human work the agent is meant to replace actually costs.

Why do AI agent projects get cancelled?

The three causes Gartner lists (cost, unclear value, missing controls) are not technology problems. They are framing problems. And the second one, unclear business value, drags the other two along with it.

"Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied," says Anushree Verma, senior director analyst at Gartner. "This can blind organizations to the real cost and complexity of deploying AI agents at scale, stalling projects from moving into production."

A project gets cancelled when you cannot defend it. And you cannot defend it when the only figure you hold is what the agent costs. That one is easy, it sits on an invoice. What the agent takes off your plate sits nowhere, and almost no small company has ever measured it.

The Italian gap tells the same story from another angle: 71% of large companies have started at least one AI project, against 15% of mid-sized and 7% of small ones (Artificial Intelligence Observatory, Politecnico di Milano, February 2026). Large companies have management control that can price a process. Small ones almost never do.

How do you work out whether an AI agent pays off?

You need a bill of quantities. It is a construction term: before you price the building, you measure the work, item by item, quantity by quantity. Nobody would build a warehouse starting from the price of cement without knowing how many cubic metres they need. With agents, that is exactly what happens.

A bill of quantities for an agent has three columns:

1. The activity, described the way a person does it today, step by step. Not "order management", but "opens the email, copies the data into the system, checks availability, replies to the customer".
2. The real time it takes, measured or estimated with the person who performs it, not with the person who imagines it. The two figures almost always diverge, and the one who does not do the work is the one who overestimates.
3. The fully loaded hourly cost of whoever performs it, which is not their net pay.

Out of that comes a monthly number, and at last you have two comparable figures. An example from a real project I worked on: an agent at 2,000 euro a month, against roughly 8 hours of employee work saved each month. Done this way, the answer is no, and it needs saying plainly: at that cost, it does not pay back.

That is the value of a bill of quantities even when it tells you to stop. It saves you from finding out six months later, which is the moment the project gets cancelled and the technology takes the blame.

What never makes it into the calculation?

The bill of quantities, as it stands, is still incomplete. Two items move the result and I have never seen either of them in a small company's business case.

Cognitive load

Each of us has a finite number of decisions per day. When a person is relieved of a manual or semi-automatic task, what gets freed is not only time: it is decision capacity. The hours freed by an automated task do not come back as generic time, they come back as clean mental energy that moves onto decisions currently made in a rush at the end of the day, or postponed.

It is not an easy item to put a price on, and I will not pretend otherwise. But it belongs in the business case, because it is the difference between an employee who executes and an employee who oversees.

The capability map

Not every employee can do everything. A supervised agent lets someone who is still learning a skill work today on tasks they could not yet complete alone, with a senior colleague staying on watch.

This is the reading that changes the conversation with the owner: the agent is not a replacement for people, it is a vehicle for transferring experience, gradually and continuously, from the seniors to the rest of the team. The formula I use is "I delegate to the AI through my team, but I stay on the output through my team and through dedicated review sessions". Control does not pass to the juniors. A piece of execution passes, inside a governance framework, with the senior still answering for the result.

Once those two items enter the calculation, a project that said no on saved time alone can become a decision worth weighing. But the order matters: bill of quantities first, these second. The other way round they become a retrospective justification, which is exactly how projects that will be cancelled are born.

How do you tell a real agent from a wrapper?

Gartner estimates that out of the thousands of vendors presenting themselves as agentic, only around 130 actually are, and calls agent washing the rebranding of existing assistants, RPA and chatbots without genuine agentic capability (Gartner press release, 25 June 2025).

On the ground it looks like this: many agents on offer are wrappers around OpenAI or Claude. Built that way, they carry no value. An agent carries value when it enters the process, is codified the way an employee would be, and works alongside the real job. Everything else is automation dressed up as an agent, and automation dressed up as an agent costs like an agent and delivers like a macro.

If a quote for "AI agents" lands on your desk, these are the three questions to put to the vendor before signing:

1. How have you handled compliance? Many startups have not thought about it, and behind it sit the AI Act, Italian law 132 and potential data breaches implicit in terms of service nobody read properly. If the answer is vague, the risk stays with you, not with them. Since August 2026 some obligations are already live.
2. Is the agent tied to the process, or is it a piece of the CRM or ERP? Purely technical agents, hooked to a piece of software rather than to a workflow, help very little.
3. Have you measured whether it is worth it, and how? This is the question every innovator falls down on, and the one a serious vendor answers by asking you for data instead of handing you a percentage.

Why does a pilot stall before production?

In my experience the pilot does not stall because of the model. It stalls because the agent gets wedged into a poorly governed process.

The typical case sits between marketing and sales. If agents contact leads without discipline, the mechanism jams quickly: the salesperson finds the same customer logged twice, the message does not match the stage of the deal, and internal trust in the tool collapses within weeks.

What is missing in those cases is always the same two pieces:

• A named person who feeds the agent feedback continuously, with integration fine tuning that does not end at go-live. A person with a name, not "the team".
• A refresh mechanism for that feedback, which keeps the recent input, discards the old and normalises it. Without it, the agent accumulates contradictory corrections and gets worse instead of better.

Anyone designing an agent without those two pieces is not building a production system, they are building a demo that works on demo day.

What is an agent worth once it reaches production?

A concrete example, from a financial consulting firm. The agent we built reads a document archive and produces analytical playbooks in structured, navigable text.

The analysis work it replaces used to be measured in weeks. Today the same analysis closes in half a day of agent work, with human review on top.

What separates it from the pilots that stall is not the model, which is the same one available to everyone. It is that the agent sits inside a process someone owns, with a named person correcting it and an output that a professional reads and judges before it goes to the client.

Want to know whether an agent pays off for your company?

Start from the bill of quantities, not from the vendor. Measure the work you want to replace, put the fully loaded hourly cost next to it, add the two items nobody adds, and look at the number that comes out. If it says no, you have saved six months. If it says yes, you have a business case that holds up in front of your accountant.

If you want help building it, the free Pre-Assessment is made for exactly this, and it is the first step of our AI consulting path for small and mid-sized companies. In the meantime you can start from the ROI calculator, read why so many AI projects produce no return and why assessment comes before investment. And if you want to see what an agent looks like from the inside, there is the practical guide to building one with n8n.

Frequently asked questions

How much does an AI agent cost for a small company?

In the projects I have worked on, running an agent in production costs from a few hundred to a few thousand euro a month, on top of the initial build. The figure alone says nothing: it only counts against the cost of the work it replaces, worked out with a bill of quantities. An agent at 2,000 euro a month that frees 8 hours of work does not pay off, one at 500 that frees 60 does.

Do AI agents replace employees?

In small and mid-sized companies, almost never. The effect I see most often is different: the agent removes the repetitive task and frees decision capacity, and lets someone learning a skill work on tasks they would not close alone, with a senior staying on watch. It is a transfer of experience inside the team, not a replacement of people.

How do I know whether a vendor is selling a real AI agent?

Ask three questions: how they handled compliance with the AI Act and Italian law 132, whether the agent is tied to the business process or is just a piece of the CRM, and whether they measured the return and by what method. Gartner estimates that out of thousands of vendors calling themselves agentic only around 130 really are, and calls the rest agent washing.