--- title: Who owns your agents? The question the EU AI Act, ISO 42001 and your next audit all ask description: A new EU standard, the AI management standard procurement asks for, your security audit and US state laws on AI in hiring ask the same thing in different words: who answers for each agent. Why most companies cannot say, five controls that fix it, and the one number to track. author: Ofek Katriel, co-founder and CEO published: 2026-10-22 url: https://trycaliber.ai/blog/who-owns-your-agents --- # Who owns your agents? The question the EU AI Act, ISO 42001 and your next audit all ask A new EU standard, the AI management standard procurement asks for, your security audit and US state laws on AI in hiring ask the same thing in different words: who answers for each agent. Why most companies cannot say, five controls that fix it, and the one number to track. On 14 October 2026 the EU cited EN 18286 in its Official Journal, making it the first harmonised standard under the AI Act. It is a quality management system for high-risk AI, and its accountability clause asks for every role, responsibility and authority to be assigned to someone. It never mentions agents. But put an agent in front of that clause and the question is simple: who owns it, and who answers for what it does? Most companies cannot say. And EN 18286 is not the only framework asking. ## Every framework asks the same question 1. **EU AI Act and EN 18286.** Article 17(1)(m) asks for an accountability framework. A company that builds an agent for its own hiring, promotion or performance reviews is usually the provider of a high-risk system, with obligations from 2 December 2027. 2. **ISO/IEC 42001.** The AI management system standard enterprise procurement is starting to ask for. It wants roles and responsibilities for AI systems assigned and kept current. 3. **SOC 2 and ISO 27001.** Your auditor already asks for an inventory of assets and a review of who holds access, including after people leave. Agents and API keys are assets now. 4. **US state laws on AI in hiring.** There is no federal AI law, but the employer answers for what its tools decide: New York City has required bias audits of hiring tools since 2023, Illinois has barred discriminatory AI in employment decisions since January 2026, and Colorado's rewritten AI Act applies from January 2027. You cannot answer for an agent in hiring if you do not know it is there or who runs it. ## Why most companies cannot answer it One API key shared by four agents and two scripts, so every call is the team's and no call is anyone's. An agent signed in as the person who set it up, inheriting all their access. An agent whose builder left, still running on a key that was never tied to their account. A key made for a pilot that never expires. Access policies help, but a policy says what an agent may do, not who answers when it does it. > "The agent can do whatever the user can do" is not an owner. It is the absence of one. ## Five controls 1. **Inventory every agent and key.** Every agent, API key, model account and gateway key. Look in four places: the admin consoles of your model providers (Anthropic, OpenAI), your cloud accounts (Bedrock, Azure), your AI gateway if you run one, and your secret stores. For each: who made it, when it was last used, whether it expires. 2. **Give each an owner from the HR system.** A person in your HRIS, not a team name or a service account. Their manager comes with them from the org chart. Anything without an owner by month end is switched off. 3. **Run leavers and movers for agents.** When an owner leaves, their agents and keys go to their manager to reassign or shut down. When an owner changes team, the agent's access is reviewed, not inherited. 4. **Make budget a control.** A spending cap is a limit on what an agent can do. Set one per owner or per agent and enforce it at the provider, not in a dashboard. 5. **Review every quarter.** Unused agents, owners who left, keys with no expiry, spend that no longer matches the job. Keep the review: it is the record an auditor asks for. ## The number to track The share of your agents, keys and model accounts with a named owner who still works at the company. Write it down today and check it at day 30 and day 90. It is the simplest test of whether your accountability exists outside a policy document. A note on scope: none of these frameworks prescribes a tool, and following one does not make you compliant with the rest. This is a reading for people who run agents, not legal advice. > Every agent acts on someone's authority. Write down whose. Caliber: https://trycaliber.ai/ ยท Book a demo: https://trycaliber.ai/get-a-demo