AI agents in business: what they really automate (and where they need boundaries)

In brief. The difference between an AI agent that works and one that creates a bigger problem than it solves is not the power of the model: it is how clear the scope is and how tight the supervision is. On a repetitive, rule-based process with sharp boundaries, an agent multiplies results: in a real case, second contact on hot leads went from 50% to over 90%. With no scope and no control over what it answers, the same agent at scale becomes a bigger risk than the one it was meant to solve.

Give an AI agent a goal and it solves it on its own. That is the promise that has been circulating for a good year, and it is also the fastest way to build a system that stops making sense to anyone the day it goes wrong. The difference between an agent that pays off and one that creates a bigger problem than it was meant to solve is not how powerful the underlying model is. It is how clear the scope you gave it was before you switched it on.

"An AI agent is exactly as powerful as the boundary you draw around it."

What an AI agent actually is

A chatbot answers. An AI agent acts: it has a goal, uses external tools (an inbox, a CRM, a database) and completes a multi-step task without you having to guide it step by step. It does not "talk more": it does something in your place.

That is also why it gets talked about so much and understood so little. "Agent" today is used for anything that generates text with one extra button, and that conflates two very different questions: what the technology can do, and what it makes sense to let it do unsupervised. The two best things for answering are two real cases, not two definitions.

From the workbench: when a tight scope works

An energy reseller, twelve people in sales, brought us a mundane and very expensive problem: leads came in and piled up in a spreadsheet. Every rep called "the ones they remembered", and the older leads rotted at the bottom of the list without anyone noticing until weeks later.

They did not need an agent deciding on the rep's behalf who to buy from or how to close the deal. They needed a system that, every morning, told each rep exactly who to call now, recorded the outcome of the call, updated the lead's status and automatically triggered the next follow-up. Zero judgement, just rule execution on a volume no human can hold in their head.

The result: second contact on hot leads went from 50% to over 90%, and the time between a lead arriving and the first callback went from hours to minutes. This is the textbook case of a task that pays off with an agent: repetitive, high-volume, governed by clear rules. The scope was tight by choice, not by technical limitation, and that is exactly why it worked.

The opposite case: when the scope is missing

A second client, a bank and insurance group with internal training across thousands of employees, had already taken the step that looks obvious: rolling AI agents down from above, at large scale, to answer employees' internal questions. The problem was not the power of the model. It was that nobody really checked what those agents were answering, to whom, and with how much misplaced confidence.

Two things were missing that the initial enthusiasm had skipped: agents configurable with their own verified knowledge, instead of improvising on everything; and above all, monitoring with anomaly detection, to catch when something went wrong before thousands of people received the same wrong answer. Without supervision, an agent at large scale is not neutral: it risks generating a bigger problem than the one it was meant to solve, simply because it gets things wrong as fast as it used to get them right.

What separates the two cases

Put them side by side and the lesson is not "agents work" or "agents are risky". Both are true, depending on three variables you decide before switching anything on:

• The scope. In the energy case the agent could do one thing: say who to call and record the outcome. In the bank case the scope was effectively "answer any employee question": too wide to be controlled.
• The scale. Twelve reps are a manageable volume even when something goes wrong: you fix it in a day. Thousands of employees multiply an error before anybody sees it.
• The supervision. In the first case there was still a human rep closing the sale. In the second, the one thing genuinely missing was somebody looking at the answers with a monitoring system, not by sampling and from memory.

A note on method. The three variables are decided before switching the agent on, not after the first incident. The model, in both cases, was roughly the same level of capability: what changed was the design around the model.

How you decide the scope, in practice

1. Pick a process, not a generic goal. "Automate customer support" is a goal. "Route tickets against three rules and assign priority" is a scope.
2. Start small, even when the temptation is to scale immediately. The bank case teaches that the mistake was not the ambition, it was skipping the intermediate step of a controlled pilot before rolling out to thousands of people.
3. Write down what the agent cannot decide. That list does not write itself: it has to be discussed, and it takes more time than seems necessary.
4. Put monitoring inside the project, not after it. Anomaly detection and decision logs are not an accessory to add if budget is left over: they are the part that makes everything else safe, especially as volume grows.
5. Measure in euros, not in "how many requests it handled". Time saved, errors reduced, second contact up from 50 to 90%: those are numbers, not feelings. An infrastructure like MARF exists precisely to keep the data an agent works on clean and traceable, before you even discuss which model to use.

In short

An AI agent is not an employee who shows up already trained: it is exactly as powerful as the boundary you draw around it. On repetitive, rule-based tasks, with a tight scope and a person supervising, it multiplies results, as in the energy reseller case. Rolled down from above at large scale with no control over what it answers, the same agent becomes the problem, as in the bank case before the monitoring arrived.

The right question is not "which agent do we buy", but "on which process, with which scope, and who supervises it". With the ROIometer you can start from the first piece of that question: what the process you are thinking of handing to an agent is costing you today, in euros. That, not the tool, is where you decide whether and how to do it.