Agent-Based Infrastructure Operator: Definition and Role

Réponse courte. Un opérateur d'infrastructure agentique conçoit, déploie et exploite dans la durée les agents IA d'une entreprise, avec des engagements de service, la traçabilité de leurs actions et la réversibilité. Il se distingue du cabinet de conseil, qui recommande, et de l'intégrateur, qui livre puis se retire. Il prend en charge l'infrastructure sur laquelle agissent ces agents : accès aux données et aux applications, identités et droits des agents, orchestration, supervision, traçabilité, maîtrise des coûts et conformité. Il ne vend ni un modèle d'IA ni un logiciel. Il répond du comportement des agents en production : ils font ce qu'on attend d'eux, avec les droits qu'on leur a donnés, et chacune de leurs actions peut être expliquée.

Definition: Agent-based and operator-based infrastructure

An agentic infrastructure is the set of components that enable AI agents to act within a company’s information system in a controlled manner. The terms “agent-based AI infrastructure” or, in English, “agentic infrastructure” are also used.

The term refers to a layer, not a product. It sits between AI models (GPT, Claude, Mistral) and business applications (ERP, CRM, email, ticketing systems). It determines what an agent can access, what actions it is authorized to take, how it is monitored, and how a human can take over. For the basics of agent-based AI itself, see our article “Agent-Based AI: Definition, How It Works, and Business Applications.”

The operator is the person who keeps this layer running on a day-to-day basis. The term says it all: like a network or hosting operator, they are committed to a service in production, not to a completed project.

Why an AI agent needs a dedicated infrastructure

A chatbot generates a response, and then a human decides what to do with it. An AI agent goes a step further: it reads data, selects an action, and executes it within the company’s software. It creates a user account, updates a customer record, or closes a ticket.

As soon as an agent takes action, five questions arise, and none of them depends on the AI model:

• What data can the agent access?
• What actions can it perform, and what permissions does it require?
• Who is alerted when it makes a mistake, and how does a human take over?
• How much does each task he handles cost?
• Three months later, can we still figure out why he made that decision?

These issues fall under the purview of IT operations and security. That is why the challenge of an agent project lies less in the choice of model than in its integration into the information system.

The Eight Components of an Agent-Based Infrastructure

Example: An agent who processes access requests in Microsoft 365

Let's consider an agent responsible for handling requests for access to Microsoft 365 shared spaces. An employee writes: "I need access to the Sales department's Clients folder." The agent reviews the request, identifies the relevant group, and adds the person to it.

Here's what the operator sets up around him:

• Identity. Agents have their own identities in the company directory. Microsoft offers Microsoft Entra Agent ID for this purpose, which enables conditional access rules, centralized logging, and lifecycle management for agents. Copilot Studio agents created before this feature was enabled must be recreated; there is no automatic conversion.
• Permissions. The agent can view the directory and add members to a closed list of groups. The agent cannot create groups or modify administrative permissions.
• Human review. A request involving a sensitive group (finance, human resources, management) is sent to the group manager for approval.
• Traceability. Each addition is logged along with the original request and the reason provided by the agent.
• Security. An unusual number of additions, or a request outside of normal business hours, triggers an alert at the SOC.
• Cost. The operator tracks the average cost of processing a request to compare it with the time required for manual processing.

None of this depends on the AI model used. It is the operational work that enables an agent to act without becoming a risk.

Publisher, integrator, operator: three different professions

A single service provider can perform several of these roles. So the right question isn’t “Who builds the agent?” but “Who is responsible for it on Monday morning?”

What an operator does on a daily basis

Before going live, the operator maps out the access rights the agent needs, creates the agent's identities, sets the thresholds beyond which a human takes over, and records the processing in the GDPR register.

In production, it tracks processed tasks and errors, coordinates manual intervention when an agent goes beyond its scope, monitors costs, and reports security incidents.

Over time, it adjusts entitlements, adapts connectors when an application changes, manages changes to the underlying model, and generates detailed reports, along with their calculation methods.

How to Choose a Service Provider: Eight Questions to Ask

1. Does each agent have a distinct identity and set of rights that can be revoked without affecting human accounts?
2. How does a human take over: who, at what times, and within what timeframe?
3. Are agents' actions logged and accessible to the customer?
4. Are incidents involving staff members monitored by a security team?
5. Is the cost per task tracked and reported?
6. Where is the data stored, and what does the customer receive at the end of the contract?
7. Does the operator already have agents in production, and does it publish measured results along with the methodology used?
8. Is he familiar with the existing IT systems (Microsoft 365, directory, ERP), or is he starting from scratch?

Our status at IT Systèmes: an agent already in production

IT Systèmes is an IT services company founded in 2010, with 44 employees in Paris (Malakoff), Lyon, Annecy, and Bordeaux. We manage the information systems of more than 50 clients through IT outsourcing. Our approach to agent-based infrastructure stems from our experience as system operators: an AI agent in production monitors, secures, and corrects itself just like the rest of the information system.

We are already using our own agent. The Hypergérance offering includes Helpy, an AI agent that handles Level 1 support. From June to August 2026, among the 55 customers on Hypergérance contracts, Helpy received 1,889 tickets: 680 in June, 720 in July, and 489 in August, including the summer lull. 60% were closed without human intervention, in an average of 3 minutes. Any ticket taken over by a technician—even if only partially handled by the agent—is excluded from this rate. Source: Helpy monitoring tool.

Ce que nous faisons pour Helpy (suivi au quotidien, reprise humaine, mesure publiée avec sa méthode) correspond exactement à ce qu'un opérateur d'infrastructure agentique doit faire pour les agents d'une entreprise. Deux de nos activités s'inscrivent dans cette logique :

• Construire : nous développons les agents et les logiciels qui les accueillent, selon la méthode d'hyperdéveloppement, où l'IA génère le code sous la supervision de nos experts.
• Deployment: We run our own agent in production on our customers' infrastructure (Hypergérance offering, using the Helpy agent).

Our Microsoft expertise (33 active Microsoft certifications as of September 4, 2026) is key here: for many small and medium-sized businesses, employees rely on Microsoft 365 and its directory, which is where their identities and access rights are managed.

Frequently asked questions

What is an agent-based infrastructure?

It refers to the set of components that enable AI agents to operate within a company’s information system in a controlled manner: connectors, identities and permissions, orchestration, monitoring, traceability, cost control, security, and compliance.

What is the difference between an AI provider and an AI integrator?

The integrator delivers a project and guarantees its compliance with the specifications. The operator guarantees the operation of the agents in production throughout the duration of the contract.

Can an IT services company serve as an agent-based infrastructure operator?

Oui, si elle en a le métier : exploitation dans la durée, suivi au quotidien, sécurité et reprise humaine. C'est ce qui distingue une ESN IA qui opère des agents en production d'une ESN qui se contente d'intégrer des outils d'IA.

Does an SME need an agent-based infrastructure operator?

Not for an assistant who simply answers questions. Yes, as soon as an agent interacts with the information system—whether by entering data into software, modifying permissions, sending messages, or closing requests—it becomes necessary to manage that person’s access, monitor their activity, and be able to take control if needed.

How can you secure an AI agent infrastructure?

The ANSSI guide “Security Recommendations for a Generative AI System” does not specifically address agents, but several of its recommendations apply directly to them: controlling the AI system’s interactions with business applications, limiting automated actions when the AI processes uncontrolled inputs, and logging processing activities. We add to this a dedicated identity for each agent with minimal privileges, log monitoring by a security team, and a regular review of privileges. We detail these practices in our article “How to Secure an AI Agent Project in a Business Setting.”

How much does it cost to implement an AI agent?

At IT Systèmes, an AI agent starts at €5,000 for a single-task agent and goes up to €30,000 for a multi-agent solution.

Does the AI Act apply to AI agents?

The European AI Regulation (Regulation (EU) 2024/1689) classifies systems according to their risk level. Many internal systems, such as support services or document processing, fall under the minimal-risk category. As soon as a system interacts directly with people, the transparency requirements of Article 50 have applied since August 2, 2026: the user must be aware that they are interacting with an AI. For high-risk uses listed in Annex III, such as recruitment or creditworthiness assessment, the obligations have been postponed to December 2, 2027, by Regulation (EU) 2026/1744, known as the Digital Omnibus AI Regulation. Finally, the requirement to train staff who use AI has been in effect since February 2, 2025.

Sources: Regulation (EU) 2024/1689 on Artificial Intelligence and Regulation (EU) 2026/1744, EUR-Lex; ANSSI, Security Recommendations for a Generative AI System; Microsoft Learn, Microsoft Entra Agent ID documentation; Helpy Barometer, June–August 2026, Helpy monitoring tool.