Two enterprise AI agent use cases, placed on the same process, can raise very different questions for the committee that must authorise them. In the first, the agent reads invoices, matches them against purchase orders and hands a list of discrepancies to an accountant. In the second, on the same invoices, it triggers payment. The business function served is identical, and the consequences of an error no longer have the same reach.
These use cases are therefore better classified by what the agent is allowed to do than by business department. Four families stand out: collecting and preparing files, read-only; sorting and routing requests; proposing an action that a person validates; executing actions within set limits. Before authorising an agent, a committee benefits from examining, in this order, the task, the actions, the permissions, human oversight and measurement. This classification and this grid are a proposed method. The notion itself is defined in our AI agent entry.
Can the task be described and checked?
The file presented to the committee describes the task by what triggers it, the expected end state, the way that state is checked and what happens in the event of an error. A request worded as “an agent for customer service” contains none of these elements.
The end state separates the candidates. “The documents in the claim file are gathered and their dates are consistent” can be checked by a rule or by a person. “The supplier has been chased tactfully” cannot be checked in that way: the task can be entrusted to an agent, but its result will have to be reviewed every time.
Two findings send the file back before any technical discussion. If the steps are always the same, a sequence written in advance delivers the service with predictable behaviour; our AI agents page compares these degrees of autonomy. If nobody in the business can write down the expected end state, what is missing is the description of the process, and it is written before a technique is chosen.
Prioritising among the shortlisted candidates and estimating what each one brings in fall under AI strategy.
Which actions will the agent carry out itself?
This is the question on which the families of use cases diverge. The committee asks for the complete list of actions open to the agent, each placed in one of three classes: reading, reversible writing in an internal application, an irreversible action or one that commits the organisation towards a third party.
- In collection and preparation, the agent reads several applications and hands over a file: the documents for a request, a customer’s history before a meeting, the elements of an IT incident. All its actions are reads.
- In sorting and routing, it qualifies an incoming request and assigns it to a queue or a team. Writing exists, it remains internal and can be corrected by reassignment.
- In proposal, it prepares the action without carrying it out: an invoice reconciliation to be validated, a reply to a supplier to be reviewed, a code fix submitted for a developer’s review.
- In bounded execution, it acts itself within limits: creating a ticket, updating a field, restarting a failed job, sending an acknowledgement of receipt.
A single project can span several families: the heaviest action on the list then sets the level of review. A support agent that reads, sorts and drafts falls under proposal as long as a person sends the reply. It changes family the day it sends the reply itself: the message commits the company towards a customer.
The committee also asks for the reverse list, that of the actions the agent must never be able to carry out (a payment, a deletion), so that their absence can be checked in the configuration.
Under which identity and with which rights does it act?
An agent’s rights are best granted like those of a new joiner: by the owner of each application concerned, for a task, with a review date. The team building the agent does not grant them to itself. The file states, for each application, the account used, the rights opened and the person who granted them. The design of tools and technical limits is described in our service page; the committee checks who decided on them.
Reading deserves as much attention as writing. An agent can reproduce in its outputs what it has read: if it accesses staff records to answer a question about leave, what it writes afterwards may carry traces of them. The read scope must therefore be justified in the same way as write rights.
One combination calls for particular scrutiny. The agent reads content from outside (emails, attachments, web pages), it has access to sensitive data and it has an output channel, such as sending a message or calling an external service. A booby-trapped piece of content can then make it pass on what it has read. When these conditions are all present, we propose removing one of them, or placing a human validation on the output channel.
Who supervises, and with what power to stop?
Supervision has a cost, which the file expresses in working time: how many validations per day, requested from whom, on the basis of what information. A validation requested hundreds of times a day from the same person risks becoming a mechanical gesture. Regulation (EU) 2024/1689, read in its consolidated version of 27 July 2026, names this risk: its Article 14(4)(b) refers to a tendency to “automatically or excessively rely” on the outputs of a system, which it calls automation bias.
This article applies only to high-risk AI systems, a category the regulation attaches to the system’s intended purpose. Its vocabulary is nonetheless useful to any committee. Paragraph 3 calls for oversight measures “commensurate with the risks, level of autonomy and context of use”. Paragraph 4 covers the ability of the people responsible for oversight to disregard or reverse an output, in point (d), and to interrupt the system “through a ‘stop’ button or a similar procedure”, in point (e). The application timetable for these provisions is covered in our article on AI Act obligations for businesses.
The file therefore names the supervisors, their deputies, the person authorised to suspend the agent and the person who will decide to lighten oversight. For the proposal family, it specifies what the validator has in front of them: the proposal alone, or also the elements on which the agent relied. The register that keeps these designations falls under AI governance.
How do you measure an enterprise AI agent use case before extending it?
Measurement starts before the agent. Without a record of the current process (processing time, share of files reworked, effort devoted), no comparison will be possible after go-live.
Next comes, where the task lends itself to it, a dry-run period: the agent processes real requests without anything being executed, and its proposals are compared with what the teams did.
Finally, the committee sets down, in writing and in advance, the findings that will lead to extending, maintaining or withdrawing the authorisation, as well as the review date. To the operating indicators detailed in our service page, it adds the full cost per task, supervision time included, and the reason for each human intervention, which shows whether the scope was well drawn.
The authorisation is drafted like an access authorisation: a task, a list of actions, an account and its rights, named supervisors, review criteria and a date. It may cover a narrower family than the one requested: proposal where the project aimed at execution, until the records from the dry run justify going further.