With AI agents, the nature of the risk changes. An error is no longer limited to a false or imprecise answer: it can become a decision, then an action within the information system. According to the Capgemini Research Institute, these agents could generate up to $450 billion in value by 2028, while 15% of business processes could reach a level of partial or full autonomy within the next twelve months.
The more agents enter the value creation chain, the more mastering their power to act will fall under governance as much as cybersecurity. The text thus emphasizes that the debate, long focused on models, their data, their biases, or their hallucinations, must now focus on the architecture of power within the information system.
Do not confuse delegation with the transfer of privileges
The reflex is often to allow the agent to act “on behalf of” its user and to transfer its rights to it. This mode of operation, already observed, is deemed erroneous here. A CFO legitimately accesses treasury, compensation, reporting, or certain strategic operations, whereas an agent tasked with reconciling invoices only needs a fraction of these privileges. Transferring the CFO’s identity to it amounts to confusing the power of the one who assigns the mission with that of the one who executes it.
Software has permissions, while an agent must receive a mandate. According to the text, this mandate must be based on a distinct identity, a defined mission, accessible resources, a duration, authorized actions, and a revocation capability. The agent’s effective power must be limited to the intersection between its own rights, those of the user, and what the mission requires, “never to their addition.” The rule becomes decisive when multiple agents collaborate: delegating a task must never allow for the delegation of more power.
The issue is already concrete in some companies, particularly in the banking sector, where secure and governed generative AI environments are being deployed to employees. When these environments evolve into agents capable of intervening in business processes, protecting data is no longer enough: the authority entrusted to the machine must also be governed.
Giving agents an identity to grant them autonomy
The Cloud Security Alliance estimates that non-human identities—service accounts, API keys, certificates, tokens, or credentials used by agents—are on average 45 times more numerous than human identities in companies. Therefore, it is essential to know precisely who is acting. The text distinguishes in this regard the Agent Owner, responsible for the agent in production and its lifecycle; the agent itself, endowed with its own rights; and the user on whose behalf it acts.
This governance requires end-to-end traceability. The company must be able to reconstruct the agent’s actions, its interactions with the information system, and the data sources it mobilizes. This visibility must enable the detection of non-compliant behaviors, the triggering of alerts or safeguards, and, for critical or sensitive functions, the maintenance of human validation as long as the agent’s operation has not been sufficiently proven.
According to this analysis, the goal is not to limit the autonomy of agents, but to create the conditions for trust. The better a company can identify its agents, define their mandate, track their actions, and retain a capacity for human intervention, the more it can entrust them with significant responsibilities. The question, the text concludes, will no longer be only which AI the company uses, but to whom it has given the keys to its information system, under which identity, and under what control.
