AI Onboarding: integrating an AI agent with the same rigor as a new organizational capability

When a company hires a new employee, it does not generally hand them a computer, passwords to several systems and a list of objectives and wish them good luck. It explains the organization, spells out their role and responsibilities, introduces its working methods, passes on the necessary knowledge and gradually determines the access they will need. A period of support then makes it possible to check their understanding, adjust certain responsibilities and increase their autonomy as they master their environment.

This approach seems so natural to us that we have given it a name: onboarding. Yet when an organization introduces an artificial intelligence agent, the logic often remains very different. It selects a technology, connects it to a few data sources, gives it access and quickly starts measuring the time savings it can produce.

This approach is becoming less and less suited to agents as they gain autonomy. An AI agent can now consult multiple sources of information, use tools, analyze a situation, propose an action, carry out certain tasks, communicate with other systems and sometimes pursue an objective over an extended period. Mila describes precisely this evolution as the shift from simple software tools to semi-autonomous collaborators, and emphasizes that integrating them requires governance, human oversight and risk management mechanisms built directly into operations.

The evolution of their capabilities should therefore bring about a comparable evolution in how they are integrated. If an AI agent is to take part in a team’s work, the organization must determine how it enters that team, what it needs to know, what it can do, with whom and with what it can interact, how far its autonomy can extend and under what circumstances a person must take back the decision. Quantum Beyond calls this approach AI Onboarding.

The analogy with the arrival of an employee is useful because it forces us to consider artificial intelligence as a new operational capability rather than as a mere technological feature. It obviously has its limits: an agent is neither a person nor an employee in the legal, social or human sense. The comparison nonetheless remains highly relevant when it comes to thinking about its role, the knowledge it needs in order to function, its access, its level of autonomy, its supervision and its life cycle.

A person’s general competence is never enough to make them instantly familiar with the company they are joining. An excellent accountant initially knows nothing of the organization’s particular chart of accounts, its approval procedures, its suppliers, its historical exceptions and its internal deadlines. An experienced salesperson must learn the products, the customers, the acceptable margins, the brand positioning and the commitments they are allowed to make. The same reality applies to artificial intelligence.

A model can have remarkable general capabilities while being deeply ignorant of the organization in which we want it to work. It does not naturally know its processes, its responsibilities, its business rules, its exceptions, its vocabulary, its customers, its policies, its reference data, or the boundary between what can be executed automatically and what must be submitted to human judgment.

AI Onboarding therefore begins before deployment with a basic question: what is this agent’s role? The organization must determine why it exists, the outcome it is meant to contribute to, the tasks that belong to it, the people it will interact with and the decisions it can make. It must also distinguish the situations in which it can act from those in which its role consists of recommending an action or preparing a decision that remains a human responsibility.

These questions have as much to do with how work is organized as with artificial intelligence. A poorly defined position creates confusion when it is filled by a person. A poorly defined agentic role can produce the same phenomenon, with one considerable difference: an agent can work at very high speed, repeat an error hundreds of times and act simultaneously in several systems. The precision of the role therefore itself becomes a control mechanism.

Mila’s work on agent integration aligns with this logic by insisting on the need to move beyond isolated experiments and to embed governance directly into workflows. Human oversight must in particular come into play when certain risk thresholds are reached. The agent’s role should therefore contain its limits from the design stage rather than having them added only after a problem appears.

Once that role is defined, the organization must provide the agent with the knowledge it needs to perform it. This step is probably one of the most underestimated dimensions of artificial intelligence adoption. The industry talks a great deal about the power of models, when their value also depends on the quality of the informational environment in which they are placed. An agent cannot sustainably produce reliable work when the knowledge it depends on is contradictory, obsolete, scattered or difficult to interpret.

Humans continuously compensate for these imperfections. An employee discovers that a written procedure is no longer up to date and asks a colleague how things actually work. They learn that a decision made six months earlier changed a practice, that two seemingly similar terms have different meanings, or that a historical exception explains a step that seems pointless. A considerable share of organizational knowledge thus resides in the experience, conversations and habits of teams.

An AI agent does not naturally possess this informal memory. Its integration therefore requires organizational knowledge to be prepared. Policies, procedures, responsibilities, business rules, definitions, reference frameworks and official sources must be structured well enough for the agent to determine which information to use and how much authority to grant it. Contradictions must be identifiable, and obsolete references must be distinguished from those that actually govern current operations.

This is precisely where AI Onboarding meets the Qb Knowledge Standard – AI Readiness & Knowledge Governance. An organization that is hard for its own employees to understand will be just as hard for its agents. Introducing AI then acts as a revealer of the quality of organizational knowledge and creates an opportunity to structure it better. Preparing the agent thus amounts in part to making the organization more intelligible, which benefits both the humans and the systems that will work with this knowledge.

Integration also works in both directions. The agent must learn to operate within the organization, while the teams must learn to work with this new capability. Its arrival can change how tasks are distributed, speed up certain activities, turn certain responsibilities into functions more oriented toward validation and judgment, make new information available and lead to the redesign of certain processes.

Employees must therefore understand what the agent does, why it does it, what its limits are and which responsibilities remain human. They must know in which situations its results can be used directly, when verification is necessary, how to report problematic behavior and when to take back control completely. This transparency becomes essential to effective adoption.

Involving teams in the design and integration of the agent also has considerable operational value. The accounts payable specialist knows the exceptions that official procedures do not describe. The salesperson knows which commitments can create difficulties with certain customers. The production manager understands why a seemingly superfluous step still exists, while the IT technician knows certain invisible dependencies among systems. This practical knowledge becomes an essential resource for teaching the agent how the work is actually done.

Employees thus become direct participants in AI Onboarding. Their expertise makes it possible to structure the knowledge the agent needs in order to function, while this work of formalization helps preserve and give value to part of the knowledge accumulated within the organization. Integrating AI can therefore become an exercise in knowledge transmission and governance rather than a mere technology project.

Identity management is another fundamental component. When an employee arrives, the organization normally determines which systems they can access. An agent should be subject to comparable discipline. NIST has notably pointed out that some early agentic deployments reproduce bad historical habits in identity management, particularly when users hand their own credentials over to agents and thereby allow them to act as if they were the user themselves.

A more mature architecture consists in treating each agent as a distinct digital entity with its own identity, its own authentication mechanisms and precisely delimited authorizations. The agent thus receives the equivalent of a digital badge, allowing the organization to know who consulted a piece of information, requested a permission or carried out an action. Its access matches its role and can evolve or be revoked when that role changes.

This distinction also protects people. When an agent uses an employee’s credentials, the logs can suggest that the employee performed the action. An identity of the agent’s own preserves better traceability and allows for more precise governance. An agent may be authorized to read an invoice without being able to modify it, to prepare a payment without authorizing it, to analyze a database without being able to delete its records, or to draft an email without automatically having the right to send it to certain recipients.

AI Onboarding thus connects directly to IAM, the principle of least privilege, Zero Trust and Continuous Trust. The objective is to give the agent the capabilities it needs to accomplish its mission while continuously adapting trust to the context and level of risk. The question, then, is not whether the organization “trusts” its artificial intelligence in general terms, but whether a particular action remains appropriate for this agent, in this situation, with this data and these potential consequences.

The level of autonomy should follow a comparable progression. A new employee is not generally given all the responsibilities of their position during their first hour. They observe, learn, carry out certain tasks under supervision and gradually develop their autonomy. Similar logic can be extremely useful for integrating an AI agent.

The organization can start by having it work in a controlled environment on historical situations and comparing its results with those that were actually obtained. The agent can then run in parallel with the team without intervening directly in operations, and then produce recommendations used by employees. When enough experience demonstrates the quality of its results in the company’s real-world context, certain actions carrying limited risk can progressively be automated.

This approach replaces abstract trust in artificial intelligence with trust grounded in concrete observations. It also makes it possible to determine the situations in which the agent delivers particularly significant value and those in which human judgment remains indispensable. The objective is therefore not necessarily to reach maximum autonomy, but to establish the optimal level of autonomy for each responsibility.

This progression also changes the role of managers. Teams will continue to be made up of people, but some processes will involve more and more agents. Managers will therefore have to learn to supervise a hybrid work capacity and to understand artificial intelligence well enough to determine which responsibilities can be delegated, how to assess the quality of results, when to intervene and how to distribute work intelligently between humans and machines.

They do not need to become artificial intelligence engineers. They must, however, develop an operational literacy that allows them to understand what agents can accomplish, their limits and the consequences of their actions. This skill could gradually become as ordinary as today’s use of management systems, dashboards and collaboration tools. The OECD also emphasizes the importance of workers’ skills and adaptability in the successful integration of AI, which shows that the transformation goes well beyond technical training on a new tool.

AI Onboarding must therefore include managers. They must understand an agent’s role as clearly as they understand the other operational capabilities they are responsible for. They must also know which indicators to use in assessing its performance, which errors are acceptable, which behaviors must trigger an intervention and which decisions remain under human authority.

A more subtle dimension concerns organizational culture. Two companies engaged in exactly the same activity can operate very differently. One will favor a high degree of autonomy, while another will require several levels of validation. Some will accept more commercial risk and others will adopt a far more cautious approach. Speed may be a priority in one environment, while compliance or security takes precedence elsewhere.

These differences influence a multitude of decisions every day. An agent integrated into an organization must therefore operate according to rules compatible with its culture, its values and its risk tolerance. This obviously does not mean that a model understands corporate culture the way a human does. Rather, the organization must translate enough of its implicit expectations into rules, examples, knowledge and governance mechanisms so that the agent’s behavior remains consistent with them.

This work has value that goes well beyond artificial intelligence, since it sometimes forces the organization to make explicit what its employees had learned intuitively. Who can make a given decision? What level of error is acceptable? Which information must never leave the organization? Which operations must always be verified? At what point should a process be interrupted? How do we really define an excellent result? AI Onboarding then becomes an exercise in organizational clarification.

Unlike the traditional onboarding of an employee, however, that of an agent has one important peculiarity: the technology that powers it can change abruptly. A vendor can update its model, new capabilities can appear, an integration can be added, permissions can change and the data the agent accesses can evolve. A business process can also be modified. An agent properly onboarded at the beginning of the year can therefore present a noticeably different capability or risk profile a few months later.

AI Onboarding must consequently be seen as a cycle rather than a one-time event. The role, the permissions, the results, the errors, the human interactions and the level of autonomy must be periodically reassessed. This continuity connects directly to agentic cyber vigilance: the organization must know what it initially authorized the agent to do and verify what it is actually capable of doing today.

This distinction will become increasingly important as models advance. An update can considerably improve reasoning or tool-use capabilities without the company itself having changed the agent’s permissions. Governance must therefore pay as much attention to the evolution of capabilities as to that of access.

AI Onboarding thus finds its place within a broader approach to adoption and organizational transformation. AI Adoption prepares the organization to integrate artificial intelligence into its strategy, its processes and its working methods. AI Organizational Transformation makes it possible to rethink processes, responsibilities and the distribution of work between humans and intelligent systems. AI Onboarding then comes in at a more operational level, to integrate a specific agentic capability into a team and determine its role, its knowledge, its access, its autonomy, its interactions and its life cycle.

The AI Governance Office, for its part, provides the governance needed across the entire agentic ecosystem. As multiple agents appear within the organization, someone must maintain a cross-functional view of roles, risks, responsibilities, levels of autonomy and applicable rules. The Qb Knowledge Standard provides the knowledge governance layer, while IAM, Zero Trust, Continuous Trust and Hypersecurity architecture frame identities, permissions, interactions and the organization’s ability to retain control.

AI Onboarding thus becomes the meeting point of organizational transformation, operational excellence, knowledge, AI governance, cybersecurity and change management. Properly integrating an agent requires understanding, all at once, the work it will carry out, the knowledge it depends on, the people it will collaborate with, the data it will use, the systems it will access, the risks associated with its autonomy and how its performance will be assessed.

Companies will progressively welcome more AI agents into their operations. Some will assist sales teams, others will take part in accounting, human resources, cybersecurity, software development, customer service, procurement, financial analysis or operations management. The decisive question will therefore soon no longer be merely whether an organization uses artificial intelligence, but how it genuinely integrates it into the way it works.

Companies have spent decades developing methods for onboarding employees because they understood that a competent person does not automatically become effective simply by walking through the door. They must understand their role, learn the organization, receive the appropriate tools, know their responsibilities, build working relationships and progressively acquire the autonomy needed to succeed. AI agents warrant comparable onboarding discipline, adapted to their technological nature and to the particular risks introduced by their ability to act.

An agent must therefore have a clearly defined role, reliable organizational knowledge, an identity of its own, appropriate authorizations, escalation rules, a period of learning and validation, supervision proportionate to the risk and regular reassessment of its capabilities. The humans who work with it must also be prepared, so that they understand what it can accomplish, what remains their responsibility and how to collaborate effectively with this new capability.

That is precisely the objective of Quantum Beyond’s AI Onboarding: supporting organizations and their internal teams in the progressive integration of their AI agents with discipline matching the importance those agents will take on in operations. This approach targets performance, security, work quality and the organization’s ability to retain control over the level of autonomy it chooses to grant, all at the same time.

An agent that better understands its role, works from reliable knowledge and has appropriate access can produce more value. A team that understands how to collaborate with it can make better use of its capabilities, while a manager who knows precisely its strengths and limits can delegate more responsibilities to it with a level of trust grounded in experience. Artificial intelligence then progressively becomes a genuinely integrated organizational capability.

Before asking an AI agent to work for the company, then, you have to teach it how to work within the company. That means defining what it must accomplish, what it must know, the people and systems it can interact with, how far it can act, when it must request human intervention and how the organization will assess its work over time. It is through this discipline that a company moves beyond simply using artificial intelligence and genuinely begins to work with it.