Accountability in agentic AI: who answers for the mistake?

For a long time, responsibility for a corporate decision was tied, in some way, to a person. Even when systems and algorithms supported the process, it was usually possible to identify who analyzed a piece of information, approved an operation, or authorized a given action.

Agentic AI is starting to change that dynamic. An agent can receive a goal, query information, use tools, interact with other systems, and execute actions without a person approving each step individually. As that autonomy grows, a question arises that companies will have to answer in very concrete terms: if an agent’s decision produces an undesired consequence, who answers for it?

This is where accountability gains importance. In this article, the term refers to clear responsibility for decisions, actions, and consequences, including the ability to identify who held authority, who delegated that authority, and who should oversee how it was used.

The question is becoming more urgent because autonomy is advancing faster than controls. The Applied AI Index 2026 from Boston Consulting Group (BCG), based on a survey of 1,330 executives and decision makers, shows that companies expect to significantly expand the decision-making power granted to agents in the coming years [1].

By 2030, 42% of companies expect to give AI agents real autonomy to make decisions without human approval, but only 5% currently have the full set of controls needed to do so safely [1].

That gap is not just a technical problem. When an organization delegates authority to an agent, it also needs to preserve the ability to explain where that authority came from and who answers for the conditions under which it was exercised.

When the decision is delegated, responsibility cannot disappear

The first temptation when an agent makes a mistake may be to look for whoever executed the action. In traditional systems, that association usually works because the executor is also a person or an application whose behavior was defined in advance.

With agents, that relationship becomes less direct. The agent may have chosen a tool, defined a sequence of steps, or made an intermediate decision that nobody explicitly determined. Even so, someone defined the goal it was supposed to pursue, set its limits, granted its access, and decided that the activity could be carried out with a certain level of autonomy.

Deloitte summarizes this shift by arguing that when agents start doing the work, responsibility does not pass to the agent. It stays with the people who created the operating conditions, set the limits, and decided how much authority the system would receive [2].

That does not mean any failure can automatically be attributed to a single person. A corporate agent may involve a business area that defined its purpose, a technical team that implemented it, owners of the data used, security and risk professionals who set controls, and, eventually, external vendors that provided models or tools.

Responsibility can therefore be distributed. The problem appears when that distribution turns into vagueness.

Delegating a decision to an agent does not mean delegating responsibility for the consequences. The company still needs to know who defined its purpose, who set its limits, and who authorized its operation.

Deloitte Legal recommends precisely that organizations define responsibilities, review processes, approval limits, and intervention points before deploying agents, coordinating legal, technology, and business areas [3].

There is an important consequence in this shift. If the organization only starts discussing responsibility after an agent causes an incident, it will probably be trying to reconstruct decisions that should have been defined earlier.

Accountability depends on being able to reconstruct the decision chain

Formally knowing who is responsible for an agent is necessary, but not sufficient. When something goes wrong, the company needs to be able to reconstruct what happened.

Imagine an agent authorized to analyze an operational situation and execute certain actions. During its work, it queries data, uses a tool, calls another agent, and makes a sequence of intermediate decisions before reaching the final result. If the consequence is inadequate, identifying only the last action executed gives an incomplete view of the problem.

You will need to understand which goal had been assigned to the agent, which permissions were available, which information was used, which tools were triggered, which decisions occurred along the way, and where the authority to execute them came from.

BCG argues that enterprise systems should be able to identify not only the agent that performed a given action, but also the person, process, or governance body on whose behalf it was acting. Delegation needs to remain traceable along the whole chain, including when an agent calls other tools or agents [4].

This traceability is especially important because traditional logs were not always designed for this kind of operation. A system may record the credential used to make a call and still not reveal that the action corresponded to a step in a workflow started earlier, under a specific delegation.

Accountability requires more than knowing which agent executed the last action. You need to be able to reconstruct the chain of authority, decisions, data, tools, and delegations that led to it.

PwC moves in a similar direction by recommending that each agent have a verifiable identity, a defined role, permissions specific to the task, and auditable records of its activities [5]. These elements help preserve a relationship between autonomy and responsibility as agents start operating in increasingly complex processes.

The topic is also starting to change the role of managers themselves. According to Gartner, they stop supervising only people and tasks and progressively become answerable also for the behavior, results, and risks of the autonomous agents working with their teams [6].

In that same study, 75% of Chief Human Resources Officers (CHROs) say managers are more overloaded than ever, indicating that introducing agents may increase, rather than simply eliminate, part of the supervision complexity [6].

This helps explain why accountability cannot be solved just by adding a human approval at every step. The challenge is defining where human judgment needs to remain, which decisions can be delegated, and which evidence must be preserved so the organization stays able to answer for the consequences.

Distributed responsibility cannot mean undefined responsibility

As agents cross different areas and systems, it is unlikely that a single function will be able to control every aspect involved. Business, technology, security, data, risk, legal, and vendors may take part in different ways in the same process.

Deloitte observes that responsibility for AI outcomes is already distributed across different corporate functions and that this characteristic tends to remain. The recommendation is to establish clear decision rights and identify owners for the business, technology, and risk oversight dimensions, while also preserving final responsibility for relevant outcomes [7].

This distinction matters because collaboration across areas does not remove the need for accountability. Several functions may take part in managing an agent, but the organization needs to avoid a situation where, faced with an incident, each one can point to another.

BCG puts the problem in a very practical way: the company should be able to identify who approved the agent’s deployment, who defined its purpose and limits, who granted its access, and who has authority to stop its operation [4].

That last point deserves attention. Mechanisms for interruption, access revocation, or escalation have little value if it is not clear who can use them and under which circumstances. Accountability therefore involves both answering for what happened and having authority to act when the agent’s behavior stops being acceptable.

The more execution is distributed among people, systems, and agents, the more explicit the chain of responsibility must be that connects the autonomy granted to the consequences produced.

This may be one of the least visible changes brought by Agentic AI. The technology makes it possible to distribute execution and decision on an unprecedented scale, but organizations cannot let responsibility fragment at the same speed.

The challenge is not to try to assign personality or blame to the agent. It is to build a model in which the company can demonstrate who granted authority, which limits were defined, how the agent’s activity was monitored, and who was responsible for the decisions that allowed that autonomy.

Accountability, in this sense, needs to be designed before autonomy. Trying to reconstruct it only after an incident is a sign that governance arrived late.

From vision to practice: ARGUS

This need to connect autonomy, authority, and evidence is directly tied to the proposal of ARGUS as a platform for Operational and Agentic Intelligence.

With AgentOps (Agent Operations) and AI Gateway capabilities, ARGUS aims to let agent activity be monitored within an environment where identities, policies, permissions, tool calls, decisions, costs, and results can be observed and related to one another.

That traceability gains value when it stops showing only the agent’s internal behavior and starts incorporating the context in which its activity took place. A decision may have been influenced by the state of an application, a recent change, information received from another agent, a human action, or events in the infrastructure.

That is why agentic intelligence connects to shared operational context. By correlating information from multiple sources, ARGUS aims to help reconstruct not only what an agent did, but also what was happening in the environment when that decision was made and which other elements took part in that chain.

In operations that combine people, agents, automations, applications, and infrastructure, this systemic view can contribute to incident investigation, auditing, and understanding the consequences of a decision. As autonomy increases, preserving that context stops being only an observability matter and becomes part of the very ability to hold someone accountable.

Learn about ARGUS and see how it works in practice.

About the author

Paulo Florêncio de Menezes is an entrepreneur, partner, and co-founder of companies such as Target Solutions, BiPTT, and ARGUS.

He holds a degree in Data Processing from the Federal University of Amazonas (UFAM), an MBA in Business Management from Fundação Dom Cabral (FDC), and a Professional Master’s in Economics from Fundação Getulio Vargas (FGV).

In his Master’s research, he developed the IBEOP-IA (Brazilian Index of Potential Occupational Exposure to Artificial Intelligence), investigating the potential exposure of Brazilian occupations to AI capabilities by combining occupational data, artificial intelligence patents, and natural language processing techniques.

References

[1] Boston Consulting Group. The Formula for Agentic AI Value --- Applied AI Index 2026. Sep. 30, 2026.

[2] Deloitte. How AI Is Making Engineering the New Enterprise Core. 2026.

[3] Deloitte Legal. When AI acts on its own, who is responsible for mistakes? Jun. 25, 2026.

[4] Boston Consulting Group. The Authorization Gap: Why Yesterday’s Controls Won’t Work with Today’s Agents. Sep. 22, 2026.

[5] PwC. AI agents as workforce counterparts---what governance should look like. Jul. 17, 2026.

[6] Gartner. How Agentic AI Is Reshaping Manager Oversight. May 6, 2026.

[7] Deloitte. Rethinking the CISO role for an AI-saturated enterprise. Sep. 3, 2026.