TL;DR: A person or an organization answers for an AI agent’s mistake, never the software itself. The Named-Human Test asks who can stop the output and who answers for it through moral, professional, civil, or criminal channels. A Canadian tribunal held an airline liable for its chatbot, and EU law assigns oversight duties to providers and deployers of high-risk systems.
Who answers when an agent errs?
The organization that put the agent to work answers first, as a customer service chatbot case in British Columbia shows. In Moffatt v. Air Canada, decided by the British Columbia Civil Resolution Tribunal on February 14, 2024, a passenger relied on information from the airline’s website chatbot. The tribunal summarized the airline’s defense this way: “In effect, Air Canada suggests the chatbot is a separate legal entity that is responsible for its own actions.” It called that “a remarkable submission,” found that Air Canada “did not take reasonable care to ensure its chatbot was accurate,” and ordered the airline to pay the passenger C$812.02, including interest and tribunal fees.
That chatbot only answered questions, yet the principle carries over to agents that take action. Basil C. Puglisi, MPA, sets out the same principle on his page defining what an AI agent is, where Checkpoint-Based Governance holds that “an AI system does not act on its own. A human programmed it, or a human failed to exercise oversight and accountability over it.” The page sorts every agent by two questions, “who acts, and who answers,” and defines three agent models by where the named human stands.
Even the most automated of those models keeps a named person accountable. In Agent Model 1, “the human signs the charter and the machine acts on every step after it,” and every output is labeled Responsible AI. That charter, signed by a named human, fixes the agent’s lane and the conditions for a stop. Basil’s essay on agents and agency puts the rule in one line: “The factory can run without a hand on each output, but it cannot run without someone who answers for it.”
How do you test whether a named human answers for the agent?
The test is a single question, which Basil calls the Named-Human Test and publishes in his analysis of why agentic AI was always going to fail. It asks whether a named human holds binding authority to approve, modify, or halt the output, and whether that person’s accountability survives an audit through at least one of four channels: moral, professional, civil, or criminal.
The same article restates it in plain terms: “who can stop it, and who answers for it?” Without such a person, the system is Responsible AI, and with one it is AI Governance.
Speed does not excuse an agent from the test. For outputs that arrive faster than any person can read them, “the named human signs the policy that authorizes the class of outputs, and the four channels attach to that signature.” In the article’s words, “Speed moves the checkpoint but does not remove it.”
The test also screens out the rubber stamp. Drawing on regulated domains where named individuals already carry liability, such as officer certification under Sarbanes-Oxley, the article concludes that “a name on a form without authority, competence, evidence access, and a demonstrable record of override does not pass.”
In The Minds That Bend the Machine, Chapter 16 reviews Stuart Russell, asking how anyone confirms that the approver understands the work, and what happens to that review as systems outgrow the reviewer.
Both coded agent models are published as Alpha, and the HAIA Agents paper states that “Neither is deployed, and in the build record neither has yet called a live AI provider.” For an outside read on an agent already in production, Puglisi Consulting’s AI Deployment and Agent Review “compares what was promised with what was delivered and finds where the deployment broke from its own plan.”
What does the law say about agent errors?
The EU AI Act divides the duty between the party that builds a high-risk system and the party that uses it. Article 14 requires high-risk systems to be built so “they can be effectively overseen by natural persons during the period in which they are in use.” Article 26 turns to the deployer: “Deployers shall assign human oversight to natural persons who have the necessary competence, training and authority, as well as the necessary support.” Under Regulation (EU) 2026/1744, the Digital Omnibus on AI, those high-risk obligations apply from December 2, 2027 for systems listed in Annex III, and from August 2, 2028 for systems covered by Annex I.
Basil’s analysis of the EU product liability directive names the deployment pattern the directive bears on hardest: “An agentic AI system, one that acts across tools and takes consequential actions without a human approving each output, is automation: the machine checks the machine, and no named human is bound to what it ships.”
Product liability adds a second layer for products placed on the market or put into service in the EU after December 9, 2026, which is also the deadline for national transposition. Directive (EU) 2024/2853 defines a product so that “it includes electricity, digital manufacturing files, raw materials and software,” and its Article 10 creates a rebuttable presumption that a product is defective where “the defendant fails to disclose relevant evidence pursuant to Article 9(1).” The same analysis warns that logs alone do not satisfy the AI Act’s oversight duty in Article 14, because “a log is the machine accounting for what it did, not a person accounting for the decision to ship it.”
Basil’s book Governing AI: When Capability Exceeds Control examines the highest-stakes version of this problem in Chapter 7, “Autonomous Weapons and Meaningful Human Control.” In The Minds That Bend the Machine, Chapter 28, “The AI Industrialist,” asks the same question of embodied systems, where an error becomes a physical event.
Questions Readers Ask
Can an AI agent be held responsible for its own mistakes?
In the Air Canada case, the tribunal rejected the argument that a chatbot was a separate legal entity responsible for its own actions and held the airline liable. Checkpoint-Based Governance takes the same view of every agent: a human programmed it, or a human failed to exercise oversight and accountability over it.
Is the vendor or the company using the agent accountable?
Either can be, depending on what failed. Under the EU AI Act, the provider must design a high-risk system so people can oversee it, and the deployer must assign competent people with authority to do so. Contracts allocate the rest, and qualified counsel should review that allocation.
Does keeping a human in the loop make the company safe?
Presence alone does not. The Named-Human Test asks whether a named person holds binding authority to stop or change the output and whether that accountability survives an audit. A reviewer who only ratifies whatever the agent produced is the rubber stamp the test is built to catch.
What if the agent acts too fast for anyone to review each output?
The checkpoint moves upstream. A named person signs the policy that authorizes a defined class of automated actions, and accountability attaches to that signature. The Named-Human Test gives algorithmic trading and automated cyber response as examples, and the same approach applies to any agent working at machine speed.
What records show who was accountable for an agent’s actions?
Useful records include the signed charter or policy that set the agent’s limits and the name of the person with authority to stop it, along with a decision record for consequential outputs. System logs show what the machine did, but they don’t show which person decided to let it act.
Sources
Civil Resolution Tribunal of British Columbia. (2024, February 14). Moffatt v. Air Canada, 2024 BCCRT 149. https://decisions.civilresolutionbc.ca/crt/crtd/en/525448/1/document.do
Directive (EU) 2024/2853 of the European Parliament and of the Council of 23 October 2024 on liability for defective products and repealing Council Directive 85/374/EEC. (2024). Official Journal of the European Union. https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=OJ:L_202402853
EU Artificial Intelligence Act. (n.d.). Article 14: Human oversight. https://artificialintelligenceact.eu/article/14/
EU Artificial Intelligence Act. (n.d.). Article 26: Obligations of deployers of high-risk AI systems. https://artificialintelligenceact.eu/article/26/
Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 as regards the simplification of the implementation of harmonised rules on artificial intelligence (Digital Omnibus on AI). (2026). Official Journal of the European Union. https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng
Puglisi, B. C. (2025). Governing AI: When capability exceeds control (2nd ed.). https://basilpuglisi.com/governing-ai-when-capability-exceeds-control/
Puglisi, B. C. (2026, June 4). Why agentic AI was always going to fail. basilpuglisi.com. https://basilpuglisi.com/why-agentic-ai-was-always-going-to-fail/
Puglisi, B. C. (2026, July 5). Europe withdrew its AI liability directive, and the exposure it left behind is larger. basilpuglisi.com. https://basilpuglisi.com/eu-ai-liability-product-directive/
Puglisi, B. C. (2026, October 13). HAIA agents: From the definition problem to running code. basilpuglisi.com. https://basilpuglisi.com/haia-agents-running-code/
Puglisi, B. C. (2026, October 16). Agents or agency: Transparency is success. basilpuglisi.com. https://basilpuglisi.com/agents-agency-transparency/
Puglisi, B. C. (2026). The minds that bend the machine. https://basilpuglisi.com/minds-that-bend-the-machine/
Puglisi, B. C. (n.d.). What is an agent? Code, human, or both. basilpuglisi.com. https://basilpuglisi.com/agents/
Puglisi Consulting. (n.d.). Advisory reviews. https://puglisiconsulting.com/advisory-reviews/
I am not a lawyer, and this site and its advisory reviews do not provide legal advice. This is thought research and governance analysis based on public sources, cited materials, and human-AI review. It is intended to help executives, practitioners, insurers, and governance teams think more clearly about AI risk, liability exposure, and documentation practices. Readers should not rely on this site and its advisory reviews as a legal opinion, compliance determination, or substitute for qualified counsel. Any organization facing a legal, regulatory, contractual, or insurance question should consult its own attorney, broker, or professional adviser before acting.
#AIgenerated from #AIassisted content under HAIA Ecosystem
Free for personal, educational, and noncommercial research use with attribution. Commercial exploitation, paid productization, and enterprise commercialization require separate permission and licensing.
Suggested citation: Puglisi, B. C. (2026). Who is accountable when an AI agent makes a mistake? basilpuglisi.com.
Leave a Reply