Articles / Industry Use Cases / AI Governance
If an AI agent makes a harmful decision, can a company say the agent acted on its own?
No. An AI agent can choose steps and use tools within its permissions, but the organization still decides why it is deployed, which systems it can reach, where approval is required and what evidence is kept. The useful management question is not whether the software seems independent. It is whether a reviewer can trace an action back to a human owner, an operating boundary and a recorded outcome.
What changed in the policy conversation
On September 25, 2026, U.S. Federal Trade Commission Chair Andrew Ferguson told Reuters Momentum AI that he resisted treating AI agents as independent actors with their own wills. He said responsibility could instead attach to the people or companies that develop or instruct them, depending on the facts. This is an enforcement view from the FTC chair—not a new statute, a court judgment or a universal rule for every country. (Reuters, September 25, 2026)
The statement matters because “the agent decided” can sound like an explanation while hiding the real operating choices. A business selected the model, connected the tools, wrote or approved the instructions, set the limits and decided whether a person must intervene. The FTC has previously warned that existing consumer-protection rules do not disappear simply because a product uses AI. (FTC, September 25, 2024)
Trace the decision chain, not the personality
Teams often describe an agent as if it were a new employee. That analogy is useful only up to a point. Software does not become a separate accountable organization because it plans several steps. For operations, a better view is a chain with four parts:
- Instruction: What goal or request started the action?
- Tool use: Which system, record or external service did the agent call?
- Boundary: Which permission, threshold or approval rule applied?
- Outcome: What happened, who approved an exception and what was recorded?
NIST’s AI Risk Management Framework Playbook recommends connecting AI risk work to existing organizational governance, defining intended uses and documenting processes. That does not prescribe one liability answer. It does support a practical control: make the chain reviewable inside the same governance system used for other business risks. (NIST AI RMF Playbook, Govern)
A maintenance-parts example
Consider a fictional maintenance agent asked to avoid a packaging-line stoppage. It checks approved suppliers and drafts an order for replacement seals. The quote is US$1,800. The agent’s purchasing limit is US$500, so the system blocks the order and requests approval from the named maintenance supervisor.
The useful record is not “the agent wanted to buy seals.” It is the supervisor’s instruction, the supplier and price returned by the tool, the US$500 policy version, the blocked result and the identity of any later approver. If the limit were missing, the responsible fix would be to change the deployment boundary—not to blame the software for appearing autonomous.
Run a 20-action evidence audit
Choose one agent-enabled workflow and review the 20 most recent consequential actions. For each action, look for the instruction, tool call, applicable boundary or policy version, required approval or exception, and recorded outcome. Report the result as a count—such as “complete chain for 14 of 20 actions”—rather than a vague maturity score.
Do not treat a complete record as proof that the decision was lawful, fair or correct. It only shows that the organization can reconstruct the event. Legal duties vary by jurisdiction and use case, and a record can reveal a bad policy just as easily as a good one. Specialist legal review may still be necessary for high-impact decisions.
Three takeaways
- An agent’s multi-step behavior does not erase the organization’s choices about purpose, access and approval.
- A defensible operating record connects instruction, tool use, boundary and outcome.
- Traceability is evidence for review; it is not automatic proof of compliance or a correct decision.
Next action: Audit 20 recent agent actions in one workflow and count how many contain the complete four-part chain. Assign an owner and due date for the largest missing evidence field.
Answer to the opening question: A company should not rely on “the agent acted on its own.” It should be able to show who authorized the workflow, what limits applied and what evidence supports the action.
Sources
- Reuters: “FTC chair pushes back on treating AI agents as independent actors”, published September 25, 2026; event held the same day.
- U.S. FTC: “Operation AI Comply”, published September 25, 2024.
- NIST AI RMF Playbook: Govern, accessed September 27, 2026.
Leave a Reply