Orchestrating humans and AI: how to distribute work between teams and agents

As long as AI remained a desktop assistant, the organization didn't need to change: everyone used it, or not, in their own way. With agents, which take over entire tasks without being monitored at each step, the question becomes a collective one. Who does what, between people and agents? Who decides, who controls, who is accountable for the result? How does a team of eight people and three agents meet, manage themselves, and improve? This is what we call orchestrating humans and AI, and it's a management skill before it's a technical issue. This article proposes a way to approach it, based on what we see in the organizations we support.

Article published in October 2025 and completely rewritten in September 2026. The first version summarized, without citing it, McKinsey's report "The agentic organization" (September 2025); this version is based on its five-pillar framework, explicitly cites it and compares it with what we observe in the field.

What the agentic organization promises, and what you need to know about it

In its September 2025 report, McKinsey describes a shift comparable to that of mechanization or digitalization: after the industrial organization, in which 89% of companies would still operate, and the agile organization of the digital age (9%), would come the agentic organization, a decentralized network of mixed teams, which only 1% of organizations practice today. The report articulates this transformation around five pillars: the business model, the operating model, governance, human capital and culture, technology, and data. It cites teams of two to five people supervising fifty to one hundred specialized agents, a European energy company that entrusted the support of three million customers to an AI assistant, and a bank where ten squads of agents process customer knowledge.

The framework is useful, the examples are real, but they describe the frontier, not the average. In the French organizations we encounter, the reality of 2026 looks more like this: widely deployed assistants, a few agents in production on well-defined processes, and teams that don't yet know how to work with them. The right question, therefore, isn't "how to become an agent-based organization?" but "how to properly orchestrate the first three agents?" The five pillars remain a relevant framework, provided they are scaled down to the level of a team, and then to that of a department.

Divide the work: three modes, not two

The initial mistake is thinking in two columns: what the human does and what the agent does. In practice, each activity of a team falls under one of three modes, and it is the choice of mode that structures everything else. The mode human being The person does the work, possibly with an AI assistant to speed things up, and retains complete control. The mode increases The agent prepares, proposes, or pre-fills the form; the person reviews and validates it before it takes effect in the systems. The method delegate : the agent does the work from start to finish, under written rules, and the person checks by sample and deals with exceptions.

Orchestrating begins with mapping the team's activities and assigning a mode to each one, working with the team, not for it. Three criteria guide this choice: the cost of an error (the higher the cost, the longer the team remains in augmented mode), the clarity of the rules (an activity that cannot be described cannot be delegated), and the ability to verify the result in a few minutes. Sensitive situations, new cases, and anything involving relationships—a dissatisfied client, a struggling colleague, or a mediation between two departments—remain in human mode by default. This isn't a technical limitation; it's an organizational choice, and it needs to be documented.

The manager of a mixed team

The McKinsey report describes three profiles: generalist supervisors capable of managing a portfolio of agents, highly specialized experts who ensure quality and handle exceptions, and empowered frontline employees, freed from data entry to focus on building relationships. These profiles exist, but for most teams, the first transformation is that of the manager themselves. They no longer simply assign tasks to individuals; they design the work, decide on methods, interpret indicators that combine human and agent outputs, and arbitrate when the two contradict each other.

Four new responsibilities fall to him. Appointing an owner for each agent on the team, usually the person who was doing the work manually, with the authority to correct their instructions and suspend them. Maintaining the mode map, because an agent who has proven themselves in augmented mode can move to delegated mode, and an agent who is drifting must revert to the previous mode. Clearly defining what happens to the freed-up time, otherwise the team assumes the worst and redoes the agent's work "just to check." And protecting what remains human, by resisting the temptation to delegate everything simply because it's possible.

The rituals that hold everything together

A mixed team is managed with short, frequent rituals rather than quarterly reviews. Three are sufficient to start with. One Weekly review of exceptionsFifteen minutes, where we review what the officers climbed, what they did wrong, and what changes are made to their instructions; this is where the officer learns, and this is where the team takes ownership of its operation. shared dashboard, read weekly, where the team's performance appears on a single line, including both human and agent data (volumes, deadlines, error rate measured on a sample, satisfaction), to avoid the knee-jerk reaction of "it's the agent who made a mistake." And a monthly fashion review, where we decide which activities change columns, in one direction or the other.

These rituals are a better alternative to what the report calls the "agentic budget" and the "guardian agents" who continuously monitor compliance with internal policies. Such systems exist in large organizations with dozens of agents in production; at the team or department level, a human review of a sample and a fifteen-minute meeting accomplish the same task, at a cost that everyone understands.

Governance: light, but before the first agent

At the organizational level, three functions must make decisions together, and they are not used to doing so. The business defines the opportunity and the acceptable level of quality. The IT department guarantees data access, security, and system integration. The HR department anticipates the evolution of roles, organizes skills development, and champions the collective narrative. A monthly committee bringing together these three voices and the owners of the agents in production is sufficient in most cases. The important thing is that this committee exists before the first agent is in production, not after the first incident, and that it maintains an inventory: which agents, on which processes, with which owner, what scope of action, and what review date.

Human responsibility doesn't disappear in this system; it simply shifts to a different level: defining the rules, monitoring exceptions, and calibrating the balance between autonomy and risk. On this point, the McKinsey report is correct, and this must be communicated to teams from the outset: an agent is never solely responsible for anything; a person always is.

Skills to build

When execution is delegated to agents, people shift their focus to goal setting, monitoring, and relationship building. This requires two sets of skills. New skills for everyone: formulating a clear objective for an agent, providing them with the right resources, critically reviewing their work, and recognizing tasks that shouldn't be delegated. And a reinforcement of what the agent doesn't currently do: solving poorly defined problems, negotiating, listening, and making decisions with incomplete information. The report notes that non-technical profiles can build robust agent pipelines in a matter of weeks; this is true, and it means that the rare skill isn't what we think. What's most lacking are managers capable of designing the work of a mixed team, not developers.

Technology: separating what belongs to you from what can be replaced

The fifth pillar is the one that should be given the least attention, on one condition: distinguish from the outset what belongs permanently to the company from what will be replaced. Data, business rules, instructions given to agents, and examples of good deliverables belong to the company. Language models and platforms, whose hierarchy changes every six months, are replaceable. Open standards such as the Model Context Protocol This separation becomes concrete: a connector built to expose a business database can be used interchangeably by Claude, Copilot, or ChatGPT. The rule is simple: internalize only what differentiates them, keep the rest interchangeable, and never let business logic become locked into a single vendor's tool.

Where to start

The McKinsey report recommends putting agentic AI on the executive committee's agenda, defining a target vision, creating a center of excellence, and rewiring one or two key areas. That's just for a large corporation. For most organizations, the starting point is a willing team, a well-mapped process with no one vying for responsibility, and a document that answers four questions: what the agent will do, who will own it, what it won't be allowed to do, and what the team will do with the freed-up time. Then, six weeks in augmented mode, with the three rituals in place from the first week, and an explicit decision on whether or not to switch to delegated mode. The second agent reuses what the first learned; the third is the one who convinces the next level of management.

Orchestrating humans and AI isn't a matter of speed, but of design. An organization that knows, for each activity, how it's managed, who controls it, and when it will be reviewed, can add agents without losing control. An organization that adds them without planning ahead discovers the orchestration process at the first sign of trouble, and that's the worst possible time to learn.

Source. McKinsey & Company, The agentic organization: contours of the next paradigm for the AI ​​eraSeptember 26, 2025, for the figures, examples and five-pillar framework cited in this article.

For further : when AI becomes a colleague and transforms the organizational chart, our article on amplify one's abilities without surrendering one's free will and our study on Generative AI at work: between productivity and motivation.