Managing Tomorrow: Knowing How to Lead AI Agents

The manager's role has always consisted, in part, of organizing work. Understanding objectives, assigning responsibilities, coordinating people, tracking performance, stepping in when a problem arises, and ensuring that available resources are used effectively are among their fundamental responsibilities.

Until recently, this operational capacity rested primarily on people supported by relatively deterministic computer systems. Software executed defined functions, and humans generally retained responsibility for interpreting information, deciding on the next steps, and acting.

The arrival of artificial intelligence agents is beginning to change this dynamic. Some systems can receive an objective, consult several sources of information, use different tools, carry out a sequence of operations, and pass their result to a person, an application, or another agent. As these capabilities advance, some teams will operate with a combination of people, traditional automation, and AI agents with varying degrees of autonomy.

This evolution does not require managers to become technical specialists in artificial intelligence. It does, however, require them to develop a new skill: understanding this capability well enough to determine which responsibilities can be entrusted to it, with what authority, within what limits, and under what supervision. Managing tomorrow will also mean learning to intelligently organize work performed by people and AI agents.

The first professional uses of generative artificial intelligence mainly took the form of a relationship between a person and a tool. The user makes a request, receives an answer, evaluates it, and then decides what to do with it. In that configuration, control of the process clearly remains in human hands.

Agents introduce a different dynamic. They can pursue an objective, search for information, interact with several systems, use tools, trigger certain actions, and determine some of the intermediate steps needed to accomplish their mission. They thus become active participants in certain operational processes.

A sales agent could, for example, prepare a prospect's file ahead of a meeting, look up relevant information, update certain data in the CRM, and propose the next actions. A procurement agent could monitor inventory levels, analyze consumption, and prepare recommendations. A support agent could classify requests, consult a knowledge base, and automatically resolve certain simple situations. In each of these cases, the system contributes directly to the outcome.

The management question then changes in nature. It is no longer enough to decide which tools to make available to employees. It becomes necessary to determine which responsibilities can be entrusted to agents and under what conditions.

In some respects, this decision resembles a responsibility managers already know well: delegation. A manager determines what an employee can decide alone, which expenses they can authorize, which files they can handle, and which situations require approval. That authority depends on their role, their experience, the risks, and the possible consequences of an error.

A comparable logic can be applied to AI agents. Can an agent only consult information, or can it also modify it? Can it prepare an order or send it? Can it communicate directly with a client? Can it create a user account, grant credit, trigger a payment, or change an IT configuration?

The answer does not depend solely on what the technology is capable of. It depends above all on what the organization decides to allow it to do. Technical capability and operational authority are two different things.

This distinction naturally leads to autonomy proportional to risk. An agent that summarizes documents does not present the same level of risk as an agent authorized to modify an order worth several hundred thousand dollars. An easily reversible operation does not carry the same consequences as a decision that contractually binds the company.

Some activities can therefore be fully automated. Others can be carried out up to a step that requires human validation. Agents may act automatically when certain values remain within defined parameters and escalate exceptions to a manager. The greater the potential consequences, the more robust the control, validation, and escalation mechanisms must be.

This logic connects directly to AI Onboarding. Before integrating an agent into a process, the organization must define its role, the knowledge it can access, its identity, its permissions, its level of autonomy, its escalation rules, and the expected results. The objective is to turn the technical deployment of an agent into a controlled organizational integration.

Speed, however, constitutes an important difference between human delegation and delegation to a machine. A poor human decision can create an incident. A poor automated decision can be repeated hundreds or thousands of times before anyone notices the problem. Supervision must therefore be adapted to that speed and make it possible to detect unusual behaviors quickly.

The manager's role could then evolve toward management oriented more toward outcomes and exceptions. When an agent carries out several steps of a process, manually checking each of its actions cancels out much of the intended gain. The manager must instead define the expected outcome, the limits to respect, the indicators to monitor, and the situations that require their intervention.

An agent tasked with handling certain routine requests could thus operate within precise parameters. The manager would track its resolution rate, its errors, its exceptions, the human rework required, complaints, turnaround times, and costs. Supervision would focus more on the quality of the process and its results than on each individual operation.

This evolution requires well-designed control mechanisms. Systematic human supervision of every action sharply limits the value of automation, while broad autonomy without monitoring mechanisms increases risk. Maturity consists of defining the appropriate level of control according to the nature of the process, the autonomy granted, and the possible consequences.

This transformation will also change the composition of teams. We are used to representing a team through an org chart made up of people. That representation could progressively become insufficient. A sales team may already be supported by agents handling research and administrative preparation. A finance team may use agents to perform certain controls or prepare analyses. An IT team may work with agents capable of analyzing incidents, preparing documentation, or proposing certain configurations.

These agents obviously have no employee status, no legal responsibility, and no career or personal motivation. They nonetheless represent an operational capacity that must be organized. The manager must understand where that capacity comes into play, avoid duplication, define responsibilities, and ensure that people know what the agents do, what information they use, and when human intervention becomes necessary.

Managing a hybrid team therefore also becomes a discipline of process design.

This perspective can transform resource planning. When a team reaches its capacity limit, the traditional response was to increase headcount or improve processes. Managers will progressively have another dimension to consider: determining which part of the increased workload can be absorbed by additional technological capacity and which part requires more human expertise.

Growing volumes of research, classification, file preparation, or certain administrative tasks can sometimes be absorbed by agents. Complex negotiations, human relationships, strategic decisions, the resolution of exceptional situations, and many forms of judgment will continue to require significant human contribution. Workforce planning could thus evolve into genuine hybrid capacity planning.

This evolution does not automatically lead to smaller teams. In a growing company, it can make it possible to substantially increase activity volume without having to grow every function at the same pace. It can also allow employees to devote more time to activities where their expertise, creativity, judgment, and interpersonal skills create the most value.

Paradoxically, AI agents could therefore make certain dimensions of management even more human. If part of the information gathering, report preparation, administrative follow-up, or planning can be accelerated, the manager can recover time to support their employees, develop their skills, solve complex problems, meet with clients, and improve processes.

That outcome is not automatic, however. Five hours freed up by artificial intelligence can be used to produce even more reports and communications. They can also be used to better understand the difficulties the team is facing and to work on the real priorities. Technology creates additional capacity; how that capacity is used remains a management decision.

Employees must also understand their relationship with agents. A hybrid team will struggle to function if opaque systems suddenly appear in processes without people understanding their role. Employees must know what the agents do, why certain activities are entrusted to them, what information they use, how their results should be interpreted, and how to report an anomaly.

This transparency becomes particularly important as the distribution of work evolves. It is legitimate for an employee to want to understand why certain tasks are now automated and what that change means for their own role. The manager then becomes one of the central actors of organizational transformation. They must be able to explain how work is evolving, support people, and encourage the development of the skills that are becoming more important.

The adoption of agents therefore cannot rest exclusively with technology teams. It directly transforms the organization of work and must involve those who are responsible for it day to day.

This transformation requires a new form of artificial intelligence literacy among managers. They do not need to know how to train a model or master the mathematics of machine learning. Their competence must be operational enough to understand what an agent can accomplish, where its limits lie, why its results can vary, and how data, knowledge, instructions, tools, and permissions influence its behavior.

This understanding will also enable them to ask better questions of IT specialists, AI experts, and vendors. It will help avoid both excessive confidence in the technology and rejection based mainly on its novelty. The manager does not need to become the mechanic of artificial intelligence. They must learn to manage it as an operational capacity.

That capacity itself rests on the quality of the knowledge agents can access. An agent embedded in an organization whose documentation is fragmented, contradictory, or obsolete will quickly hit limits. Knowledge governance therefore becomes directly tied to the quality of agents' work. This is precisely one of the functions of the Qb Knowledge Standard: structuring and governing information assets so that people and intelligent systems alike can understand the organization well enough to work in it effectively.

Identity is another essential dimension. An agent capable of interacting with applications and data should not act under a generic identity or use an employee's privileges indiscriminately. It must be identifiable, its permissions must match its role, and its actions must be traceable. The principles of least privilege, Zero Trust, and identity management thus become directly operational components of agent management.

Responsibility, for its part, remains human and organizational. An agent does not bear organizational responsibility for a decision. People and the company determine which systems are used, which access is granted, which controls are applied, and which actions are authorized. When an agent has permission to issue a refund, modify a file, or communicate with a client, that authority was granted as part of a governance decision.

The greater the autonomy, the more explicit that governance must become. Responsibilities must be defined, actions traceable, permissions limited to what is necessary, behaviors monitored, and access quickly revocable. As the number of agents grows, this governance can no longer be improvised case by case. A function such as an AI Governance Office can then help maintain common rules, clear responsibilities, and an overall view of the ecosystem.

Agent performance must also be measured. The mere fact that a system works or produces impressive results is not enough to justify keeping it in a process. What is its real cost? What is its error rate? What proportion of cases is resolved automatically? How many require human rework? What types of exceptions arise? Are turnaround times decreasing? Is client satisfaction improving? Do employees genuinely have more time for high-value activities?

These indicators make it possible to evaluate the agent by its contribution to the operational result. They also become indispensable once the organization starts using several specialized agents. Without governance, there is a risk of reproducing the phenomenon observed with the accumulation of SaaS applications: dozens of tools, identities, permissions, dependencies, and costs piling up until no one has a clear view of the whole.

Management must therefore evolve before agents become commonplace. Organizations can already map their processes, identify decisions and levels of authority, structure their knowledge, improve identity and access governance, and develop their managers' AI literacy. This work has immediate value even when few agents are currently in production.

Above all, it prepares the organization for a much deeper transformation: available work capacity will progressively no longer be measured solely by the number of people on the team.

AI agents introduce a new form of operational capacity into organizations. They will progressively be able to search for information, use tools, carry out sequences of work, and take on certain activities with varying levels of autonomy. Their integration will necessarily change how work is organized, delegated, supervised, and measured.

Managers will therefore have an additional responsibility: determining how to distribute activities among people, automation, and agents, defining appropriate levels of authority, supervising results, and recognizing the situations where human judgment remains essential. They will also have to support employees whose roles are evolving and ensure that technological gains genuinely translate into better team performance.

This transformation goes far beyond the choice of an artificial intelligence tool. It touches processes, responsibilities, skills, knowledge, identities, permissions, cybersecurity, and governance. That is why AI Onboarding, organizational transformation, QKS, IAM, Zero Trust, and AI governance must progressively be seen as the different dimensions of one and the same hybrid work environment.

At Quantum Beyond, our approach aims precisely at supporting this evolution alongside executives, managers, operational teams, and technology specialists. The goal is to define roles and levels of autonomy, structure knowledge, secure identities and access, establish governance mechanisms, and measure the real contribution of agents so that the organization retains control of this new capacity.

Tomorrow's managers will continue to lead people, develop their talents, make decisions, and create the conditions for their success. Artificial intelligence will simply add a new category of capacity that they will have to learn to orchestrate.

The real challenge of management in the era of AI agents will be knowing where human intelligence creates the most value, where machine capacity can reinforce it, and how to organize the whole so that the organization, its teams, and its clients obtain better results.