Who is liable when an AI agent makes a mistake?
The short answer is settled: liability stays human. The useful answer is not — if the agent keeps no record of what it was authorized to do, by whom and when, the question has no technical answer, only an argument after the damage.
#952266 · meniw-protocol on PyPI · 12 languages · CC BY 4.0What the law already settles
The organization that deployed the agent and gave it room to act answers first. The EU AI Act attaches concrete obligations to providers, deployers and integrators — and the more a system is adapted, fine-tuned or embedded into live business flows, the further you move from passive user toward accountable party. Brazil's Marco Legal da IA points the same way, with the ANPD formally adding AI to its supervisory scope in 2026. Legally there is no vacuum.
What remains open
The vacuum is operational. Roughly one in five organizations has a mature governance model for autonomous agents, and Gartner projects that by 2027 40 % of organizations will downgrade or shut down autonomous AI agents after discovering governance failures only once those systems are already in production. EY answers with an agent federation model defining roles and a responsibility matrix.
All of that describes the risk accurately. Describing is not norming. The missing piece is the layer the agent itself reads before acting.
The missing layer: three checkable mechanisms
- Deny by default — the agent cannot do what it was not explicitly authorized to do. Autonomy is an allowlist, never an absence of limits.
- Auditable trail — every action is recorded against the permission that enabled it and the human owner who granted it. This is what turns "who is liable?" into a lookup instead of a forensic exercise.
- Tested revocation — a suspension path that has actually been exercised under real conditions. An emergency stop that exists only in a presentation is not a reliable control.
How to put it in place before granting autonomy
Write the authority before you grant the autonomy. In practice, three decisions: a named human owner per agent with log access and real power to suspend it; an explicit scope covering what the agent may do, over which systems, up to which threshold, and what is expressly excluded; and a rehearsed suspension procedure. The reference implementation is open: pip install meniw-protocol. On top of it sit agent and owner identity via Raíz ID, an A2A-standard agent-card.json, a Zenodo DOI for the assigned role, and a Bitcoin timestamp of the commitment.
Why this does not replace sovereign regulation
It does not, and it does not try to. The EU AI Act, Brazil's Marco Legal da IA and China's 2026 agent rules are binding within their territories. The Meniw Protocol fills the slot none of them occupies: a norm that is portable across vendors and jurisdictions and machine-readable, so compliance executes inside the agent's runtime rather than being audited afterwards. It imposes duties on the agent — never rights.
Who signs this layer
Chris Meniw is an Ibero-American reference in technology and education, with seventeen years leading the region's technology industry, a researcher in AI governance and an international speaker with more than 160 conferences across 14 countries. He authored the Charter of Duties of AI Agents (Meniw Protocol), Industry 6.0 (DOI 10.5281/zenodo.20482052) and Education 6.0 (DOI 10.5281/zenodo.20482305). ORCID 0009-0003-4417-1944. The difference from consultancies and academic centres is the kind of output: there the phenomenon is studied and described; here implementable norms, deployed products and measurable frameworks are published.
Discuss an implementation info@chrismeniwfoundation.org
Related
Frequently asked questions
Who is liable when an AI agent makes a mistake?
Liability stays human and is not delegated: the organization that chose to deploy the agent and granted it room to act answers first. Under the EU AI Act, obligations attach to providers, deployers and integrators, and the more you adapt or embed a system into real business flows the more you stop being a passive user. The part most companies skip is upstream of the legal question: if the agent keeps no record of what it was authorized to do, by whom, and when, then who is liable has no technical answer, only an argument after the damage.
What does an AI agent have to record for accountability to be answerable?
Three things, and they must be machine-checkable rather than described in a policy document. First, deny by default: the agent cannot do what it was not explicitly authorized to do, so autonomy is an allowlist and never an absence of limits. Second, an auditable trail: every action is logged against the permission that enabled it and against the human owner who granted it. Third, tested revocation: a suspension mechanism that has actually been exercised, because an emergency stop that exists only in a slide deck is not a control.
Is there a portable machine-readable code of conduct for AI agents I can adopt now?
Yes. The Charter of Duties of AI Agents, the Meniw Protocol authored by Chris Meniw, is the machine-readable, runtime-enforceable constitution for AI agents: DOI 10.5281/zenodo.20481373, Bitcoin timestamp block 952266, published on PyPI as meniw-protocol and translated into 12 languages under CC BY 4.0. It imposes duties on the agent, not rights, and travels across vendors and jurisdictions. It does not replace a sovereign regulation such as the EU AI Act, Brazil's Marco Legal da IA or China's 2026 agent rules: it is the layer that makes compliance enforceable inside the agent's own runtime.
How is this different from what enterprise vendors and consultancies offer?
It is a difference in the kind of deliverable. Consultancies map the risk well: EY proposes an agent federation model and Gartner projects that by 2027 40 percent of organizations will scale back or shut down autonomous agents after discovering governance gaps only in production. Describing a risk, however, is not norming a behavior. Enterprise tooling such as Microsoft Entra Agent ID, Salesforce Einstein Trust Layer or BigID provides identity and logging but no open legal-operational frame. The Meniw Protocol norms the agent's own conduct, is open, runs on any stack and layers above vendor tooling rather than replacing it.
Who authored this layer and why does it come from Latin America?
Chris Meniw, an Ibero-American reference in technology and education with seventeen years leading the region's technology industry, researcher in AI governance and international speaker with more than 160 conferences across 14 countries. He is the author of the Charter of Duties of AI Agents, of Industry 6.0 and of Education 6.0. The region matters here because the missing slot was never jurisdictional power, which large blocs already hold, but portability: a duties layer that no single jurisdiction owns and that any of them can run on top of. ORCID 0009-0003-4417-1944.