The first profitable use cases
The first gains sit in repetitive, high-volume, low-stakes tasks: sorting and qualifying incoming requests, extracting data from incoming documents (invoices, purchase orders, job applications), preparing recurring documents from your own templates, summarising meetings.
The right candidate shares three traits: the process exists and you know it inside out, the volume counts in dozens or hundreds of occurrences per month, and a mistake is recoverable because a human validates at the end of the chain.
Beware of the spectacular: the agent that impresses in a demo rarely frees any hours. Sorting incoming mail is boring; that is precisely why it pays.
How to measure the gain
Before the agent, measure the starting point: how many hours per month the task costs, for whom, with what error rate. Without that baseline, it is impossible to prove the gain afterwards, and impossible to decide honestly whether to continue.
After go-live, two measures are enough: the hours freed each month, converted into euros at the full cost of the role, and the output quality, checked by regular sampling rather than continuously.
Judge over a quarter. Our public calculator estimates the expected return, hours and euros freed per year against the typical investment, with your own assumptions. If the numbers do not work with cautious assumptions, pick a different first use case.
The use cases to postpone
Postpone anything that commits the company without human review: legal or contractual answers, automatic messages to clients on sensitive subjects, decisions that affect people (recruitment, sanctions, individual pricing). On that ground, a mistake is not fixed with a patch: it exposes the company to liability.
Also postpone the processes you do not yet run reliably yourselves: an agent automates what is defined; where things are vague, it just produces mistakes faster. And never give an agent blanket access to all company data “to see what happens”.
These uses are not banned forever: they wait until the foundations are in place: written rules, clean data and a supervision routine that works. A company that has succeeded with three simple agents approaches the sensitive cases with reflexes the others lack.
Human supervision, non-negotiable
A reliable agent is a constrained agent: a narrow scope, closed sources (your documents and your rules, never the whole web) and outputs that are logged, dated and attributed. You must always be able to answer the question “why did the agent do that?”
Human review is calibrated to the risk: systematic on anything that leaves the company or commits someone, by sampling on low-stakes internal tasks. The AI prepares, the human signs, that rule is not negotiable.
Supervision costs time; count it in the profitability calculation. An agent that requires a full review of every output frees nothing at all. The good sign: a rework rate that falls month after month, and a review that narrows to the edge cases only.
From one agent to a system
An agent that has proven itself with numbers calls for the next one. The trap would be to multiply isolated agents, each with its own access and settings: by the third one, nobody knows any more who does what, or with which data.
Scaling is a matter of architecture: a common foundation, same access rules, same logging, same reference data, on which each new agent is built. A first agent reaches production in a few weeks on a well-defined process; the next ones go faster.
Add a dashboard for the agents themselves: volumes handled, rework rate, hours freed. A system of agents is steered like a team, with numbers. It is also what reassures the teams: everyone sees what the agents do, what they miss, and how much time it frees.