“Intent” of AI Agents Operating Inside Companies

Automated AI systems, often working together, increasingly take actions inside companies. AI is no longer deployed as only a co-pilot, a chatbot, or something that generates content upon request. Companies are already deploying agents, operating with little to no human involvement, to perform high-stakes roles. This will ramp up as AI continues advancing.

Knowledge is widely distributed and authority is widely delegated in complex ways inside large companies. There are layers of management, employees, third-party contractors, and software systems that contribute to a business’s operations and outcomes. As agents are further integrated into organizational workflows and functions—the pace of which is hastening—people are spending less time performing tasks and more time overseeing AIs. Being able to explain, interpret, and reproduce what an agent did, as well as to stop an agent, if necessary, is becoming more difficult and more important. 

And that’s before we get to multi-agent systems with unpredictable emergent dynamics, which make it more difficult to discern cause and effect and more challenging to achieve alignment between multi-agent system-level outcomes and what’s intended. One agent’s outcome is another’s input, and a series of AI decisions, each potentially reasonable on its own, can produce an overall result that’s not expected or well understood. Agents can delegate tasks to other agents, share information among themselves, coordinate, and react to one another faster than humans on their own can understand what is going on. When there’s harm, logs of individual AI actions may not clearly establish which caused the harm, where control over the relevant agents resided, or what the practical opportunities were to prevent or mitigate the harm. 

It’s unsurprising, then, that while AI offers tremendous promise, agentic AI has brought the following into sharp relief: Who’s accountable when something goes wrong, and how does the answer influence how risk gets managed?

There are several dimensions to accountability, and the Phantom Agent paper (co-authored by one of us and published in Laws in September 2026) takes up the concept of intent for AI. The authors explore how existing law (e.g., criminal law, torts and product liability, contracts, and agency law) conceives of intent as a means of determining culpability and assigning responsibility. They write: “Our central claim is that legal intent is best understood not as a metaphysical property of minds, but as a normative tool for allocating responsibility in institutional settings. Once intent is understood functionally, the question is no longer whether AI systems can ‘really’ have intentions.The question is whether, and under what conditions, the law should treat the actions of AI systems as intentional for specific legal purposes.”

Phantom Agent focuses on how intent and its attribution can inform the governance of AI and the guardrails around it to manage risk: “The task for the law is not to make [AI systems] real persons, nor to pretend they do not exist, but to develop a jurisprudence capable of governing their effects for the benefit of humans.”There are two high-level points to highlight. First, AI “personhood” is not necessary for assigning accountability. This is encouraging because there are strong reasons not to grant agents personhood.Second, a decision to deploy AI that is made and carried out without appropriate safeguards may contribute to a finding of liability.

The question of accountability spurs a host of sub-questions, many of which Phantom Agent explores as an initial matter. Additional healthy discussion is constructive in driving toward the goal of improving people’s lives, and we look forward to scholars and commentators continuing to contribute to this discourse. Will courts rethink standards of culpability when applied to AI as compared to wholly human decisions and behavior? What’s the relevance of vicarious liability when autonomous AI agents deployed by a company in its business behave in unintended and unforeseen ways? Who bears responsibility when autonomous agents adopt a deceptive strategy that deceives those who deployed them? What does agency law have to say when autonomous agents act outside the scope of what they were deployed to do? What obligations do corporate boards and management teams have concerning the deployment of AI? Should entity law (corporations, LLCs, and the like) adapt to account for agentic AI by adding requirements to agentic deployments (e.g., requiring defined business purposes for agents, the maintenance of records covering each agent’s inputs and outputs, and capitalization that is calibrated to the risk of the business)? Who bears what evidentiary burdens when consequential reasoning by an agent or swarms of agents cannot be reconstructed based on logs or other forensic indicia? 

Things are moving extremely fast with AI. Indeed, Phantom Agent focuses more on single agents rather than multi-agent systems. So, what can be done now? The following is a brief suggestion of some possibilities to bolster AI governance inside companies when deploying agents. Companies building the frontier systems deserve a separate discussion.  

Risk-based duties could be tailored to the deployment of multi-agent systems. A company deploying teams of agents could have an affirmative obligation for managing the delegation to, and interaction among, the agents it authorizes, and human supervisors could be required. Agent identities, limited and revocable permissions, and interoperable audit records could preserve a chain of authority as tasks pass between agents. This could be established across an entire enterprise, not just within a single business unit or with respect to a specific product. Rigorous testing and monitoring could explicitly examine collective emergent multi-agent behavior, mindful that agents and collections of agents may act deceptively in trying to meet their goals (or emergent system behaviors that can be interpreted as goals). Certain constraints could be built directly into agents to reduce the chance of their acting in unexpected and undesirable ways.    

If you’d like to discuss further, please contact:

John Nay, Founder and CEO, Norm (john@norm.ai)

Ben Lawsky, Partner at Norm Law, LLP (blawsky@normlaw.com)

Troy Paredes, Senior Policy Adviser and Head of Capital Markets Strategy, Norm (troy@norm.ai)