AI-driven attacks added USD 1 million to the average breach in 2026, and 92% of organizations that suffered an AI-related breach had no proper access controls on their AI systems, according to IBM's 2026 Cost of a Data Breach Report. The same year, a Gartner survey of more than 2,000 CIOs found that 71% of organizations expect to deploy agentic AI within 12 months, while only 21% currently have agentic AI governance capabilities in place, a 50-point gap Gartner analysts flagged as a near-term risk.
That gap is exactly what SOC 2, ISO 27001, ISO 42001 and the EU AI Act are now asking organizations to close for AI agents specifically, not just for AI models. The good news: once you look closely, all three frameworks converge on the same three demands. A risk-scoped identity for every agent action. A human checkpoint in front of anything irreversible. An immutable, queryable log of what happened and who approved it. Build those three things once, at the runtime layer where your agents actually act, and you satisfy every auditor who asks, instead of building a separate story for each framework.
This guide covers what AI agent compliance means, how SOC 2 and ISO auditors actually test agents in 2026, the corrected EU AI Act timeline, and one technical mapping table that spans all three. For the broader discipline this sits inside, see our guide to AI agent governance.
What AI agent compliance actually means
Model compliance asks whether an output is accurate, fair or safe to show someone. Agent compliance asks a different question: was this action allowed to happen, and can you prove it? A model that writes a biased sentence is a quality problem. An agent that approves a refund, changes a production config, or emails a customer is an action with real-world consequences, and every framework in this guide cares about actions, not just outputs.
The three questions every framework is really asking
Strip away the acronyms and SOC 2, ISO 27001/42001 and the EU AI Act are each asking a version of the same three questions:
- What can this agent do? (scope, permissions, tools)
- What did it actually do? (logs, traces, evidence)
- Who said it was allowed to? (approvals, human oversight, accountability)
Everything in the rest of this article maps back to one of those three.
Why your existing SOC 2 or ISO 27001 posture doesn't cover your agents
Most compliance programs were designed for humans logging into static software. An employee has one identity, a fixed set of permissions, and a predictable set of actions. An agent is different on all three counts.
Built for humans and static software
A traditional access review checks whether Sarah in finance still needs access to the payments dashboard. It was never built to ask whether an agent's tool access matches the task it is actually running today, especially when that task can change based on what the agent decided mid-run.
The identity problem
Agents create non-human, often delegated and chained identities: an agent acting with a service account, calling an MCP server, which calls a third-party API, under credentials nobody individually reviews. IBM's 2026 report found that only 40% of organizations use access controls on their AI models and data at all, and fewer than half actively secure the non-human identities (API keys, service accounts, machine credentials) those agents run under. Most audit frameworks simply do not have a mature category for this yet, which is why auditors are improvising with the tools they already have.
SOC 2 for AI agents: what auditors actually test in 2026
There is no separate "AI Trust Services Criteria" and there will not be one soon. The AICPA has not added a new SOC 2 category for AI. What it published instead, in TQA Section 9561 (September 2026), is nonauthoritative guidance on how the existing criteria apply when AI sits in the examined environment. In practice, auditors are reading the same rulebook with new eyes, and the bar has moved even though the text has not.
| Common Criteria | What it already covers | What it now asks of an agent |
|---|---|---|
| CC1, CC2: control environment | Staff competence, communication | AI literacy, an acceptable use policy staff have actually acknowledged |
| CC3, CC4: risk assessment | Named technology risks | AI-specific risks named: prompt injection, model drift, agent over-permissioning |
| CC6: logical access | Who can reach what | Agent service accounts and API keys under least privilege, reviewed on a cadence, deprovisioned promptly |
| CC7: system operations | Monitoring, incident response | Anomalous agent output flagged, AI scenarios written into the incident response playbook |
| CC8: change management | Code review before release | Agent config, tool permissions and model version changes under the same approval gate as code |
| CC9: vendor management | Subservice organization review | Model and MCP providers assessed, zero-retention terms documented |
A SOC 2-ready agent log entry needs to carry more than a timestamp. At minimum: the agent's identity and version, a trace ID linking the full run, the action type and tool called, redacted inputs and outputs, whether a human approved it and who, the outcome, and which policy was evaluated before the action ran.
{
"timestamp": "2026-11-22T09:14:03Z",
"agent_id": "margo@rerun.build",
"agent_version": "2.3.1",
"trace_id": "run_8f21c9",
"action_type": "send_payment",
"tool": "stripe.refund.create",
"target_ref": "cus_8KpQ_redacted",
"requires_approval": true,
"human_approver": "finance-lead@company.com",
"approval_timestamp": "2026-11-22T09:14:41Z",
"outcome": "executed",
"policy_evaluated": "irreversible_action_policy_v4"
}That single record, repeated for every action an agent takes, is most of what CC6 through CC9 ask to see.
ISO 27001 and ISO 42001: two certifications, one agent
Teams that already hold ISO 27001 often assume it covers their agents. It covers the credentials those agents use. It does not cover whether the agent should exist in the first place.
ISO 27001 governs the agent's credentials and access
Annex A controls 5.17 and 8.5 (authentication and secure authentication) apply directly: API keys, OAuth grants and service account rotation for every agent, exactly as they would for a human user.
ISO 42001 governs the agent's purpose and impact
ISO 42001 is the newer AI management system standard, and it asks a question ISO 27001 never does: what is this AI system for, and what happens if it goes wrong? Its impact assessment, under Annex A clause 5, has no 27001 equivalent at all. Speakeasy's comparison puts the split cleanly: ISO 27001 asks whether an agent is "inventoried, identified, restricted, logged and revocable," while ISO 42001 asks whether the organization governs it "through policy, impact assessment, life-cycle control and operating evidence."
The two standards overlap enough that document control, internal audits and corrective action from an existing ISO 27001 program carry over. They do not replace each other.
"Does ISO 42001 certification satisfy the EU AI Act? No. ISO 42001 is a voluntary management standard, while the EU AI Act is binding law. An AI management system may help an organization manage its compliance work, but certification does not prove that every legal obligation has been met." , Speakeasy, What is ISO 42001 for agentic AI?
That line matters more than it looks. A management-system certificate, under either ISO standard, attests to a documented process. It never certifies that one specific agent, running in production today, is safe.
The EU AI Act for agentic systems: the corrected 2026 timeline
A lot of what is still written online about the EU AI Act is already out of date. The Digital Omnibus on AI, Regulation (EU) 2026/1744, was published in the Official Journal on 24 July 2026 and entered into force three days later, and it moved four dates that compliance teams need to replan around.
| Obligation | Original date | Current date |
|---|---|---|
| High-risk AI, Annex III (standalone systems: recruitment, credit scoring, education) | 2 August 2026 | 2 December 2027 |
| High-risk AI, Annex I (embedded in regulated products: medical devices, machinery, vehicles) | 2 August 2027 | 2 August 2028 |
| National AI regulatory sandboxes | 2 August 2026 | 2 August 2027 |
| Content marking for systems already on the market | 2 August 2026 | 2 December 2026 |
| Article 50 transparency (deployers must disclose AI interaction) | 2 August 2026 | unchanged, already in force |
Agents are not a separate regulatory category. Classification runs on what the agent does, not its architecture: an agent scoring job candidates falls under Annex III whether it is one LLM call or a twelve-step autonomous workflow. Four articles matter most for agent design specifically:
- Article 9, continuous risk management, including foreseeable misuse such as prompt injection.
- Article 12, automatic logging over the system's full lifetime.
- Article 13, transparency: deployers must be able to inspect why an action was allowed or denied, not just that it happened.
- Article 14, human oversight that can understand, detect automation bias, and override, with a working kill switch.
The point on Article 14 worth stating plainly: a system prompt that says "ask before acting" is an instruction inside the model's own reasoning. It is not an external, enforceable constraint, and regulators do not treat it as human oversight. Oversight has to sit outside the agent's own judgment, in a layer the agent cannot talk itself out of.
Penalties scale with severity: up to €35 million or 7% of global turnover for prohibited practices, €15 million or 3% for high-risk violations, €7.5 million or 1% for supplying false information to authorities. The dates above moved. The obligations and the fines did not get lighter, they got later.
One control set, three audits
This is the table worth bookmarking. Each row is one requirement that shows up under a different name in SOC 2, ISO 27001/42001 and the EU AI Act, and the one technical control that satisfies all three at once.
| What the auditor wants | SOC 2 | ISO 27001/42001 | EU AI Act | The control that proves it |
|---|---|---|---|---|
| Agent access is scoped and reviewed | CC6 | 27001 A.5.17, A.8.5 | Art. 9 risk management | Least-privilege identity per agent and tool |
| Risky actions need a human | CC7 | 42001 A.9 | Art. 14 human oversight | Approval gate before irreversible actions |
| Every action is recorded | CC6, CC7 | 42001 A.6.2.8 event logs | Art. 12 record-keeping | Immutable, exportable action log |
| Decisions can be inspected | CC7 | 42001 A.5 impact assessment | Art. 13 transparency | Logged policy and reasoning per decision |
| Config changes are controlled | CC8 | 27001 change management | Art. 9 ongoing risk review | Agent config and tool changes under approval |
| Credentials rotate and expire | CC6 | 27001 A.8.5 | Art. 9 | Partial: still needs an identity provider |
| A runaway agent can be stopped | CC7 incident response | 42001 A.6.2.6 | Art. 14 override | Kill switch outside the agent's own loop |
| Vendors and models are reviewed | CC9 | 27001/42001 A.10 supplier | Art. 9, Art. 25 value chain | Not a runtime control, needs a vendor review process |
Why Zapier, Make and n8n cannot produce this evidence
None of this is a knock on flowchart automation tools, it is simply a different job. Zapier, Make and n8n run fixed paths someone wired by hand: if A happens, do B. A flowchart has no judgment to log, only whether the static path executed, so there is nothing for an auditor to review except "did the Zap run." When the input does not match the template, the automation stops rather than deciding. That is a feature for predictable work and a dead end for Article 13's demand to inspect a decision, because there was never a decision in the first place, only a wire.
Chatbots fail the same test from a different angle. They never take an action on their own, so there is nothing for CC7 or Article 12 to log until a human reads the answer and acts on it manually, and the record of that handoff usually lives nowhere. Our guide on AI agents vs chatbots covers the distinction in more depth, and the Rerun vs n8n and Rerun vs Make comparisons go through the trade-offs for teams weighing a flowchart tool against an agent that can actually decide.
Where Rerun fits, and where it doesn't
Rerun is a framework-agnostic operations layer for running AI agents on real business work. On Rerun, these agents are called coworkers, and every coworker runs with a least-privilege identity by default, an approval gate before anything irreversible, and a live, exportable record of every action it took and who signed off on it.
To be clear about the limits: Rerun does not grant SOC 2 or ISO certification, and it does not make a system EU AI Act compliant on its own. Certification and conformity assessment are performed by accredited third-party auditors against your organization's full program, its policies, its risk assessments, its documentation. What Rerun produces is the runtime evidence an auditor or a conformity assessment will actually ask for: the access scopes, the approval records, the action logs.
If you already have someone accountable for this inside your org, a coworker can run next to them rather than replace the judgment call:

Compliance Lead
Run Compliance Gap Assessments
A compliance coworker that reads your agent logs, flags gaps against your chosen framework, and drafts the evidence package before the audit starts is a few minutes of setup, not a new hire.
For the mechanics behind each piece of this: our guides to AI agent governance, AI agent security, AI agent observability and human-in-the-loop AI agents each go deeper on one layer than this article should.
Compliance readiness checklist
Pull the mapping table above into action, and this is what you should be able to check off before your next audit conversation.
AI agent compliance is the evidence, not the certificate
SOC 2, ISO 27001/42001 and the EU AI Act were written by different bodies for different purposes, and they still land on the same three questions: what can the agent do, what did it do, and who said it could. In practice, AI agent compliance comes down to building the identity, the approval gate and the log once, at the layer where your agents actually act, instead of a separate compliance story for every framework that asks.
The deadlines moved. The bar for proving you were in control the whole time did not.
Frequently asked questions
What is AI agent compliance?
AI agent compliance means producing, at runtime, the evidence that SOC 2, ISO 27001, ISO 42001 and the EU AI Act all ask for: a risk-scoped identity for every agent action, a human approval in front of anything irreversible, and an immutable log of what the agent did and who signed off on it. It is not a single certificate, it is the audit trail that lets you pass any of those frameworks' audits.
Does the EU AI Act apply to AI agents?
Yes. The EU AI Act classifies systems by what they do, not their architecture, so an agent that performs a high-risk function such as recruitment screening or credit scoring falls under Annex III like any other system. Under the 2026 Digital Omnibus (Regulation EU 2026/1744), Annex III high-risk obligations now apply from 2 December 2027, and Annex I embedded systems from 2 August 2028. Article 50 transparency obligations are unaffected and already in force since 2 August 2026.
Is there a SOC 2 certification specifically for AI agents?
No. The AICPA has not created a separate AI category. Auditors apply the existing Trust Services Criteria (CC1 through CC9) to agents, guided by nonauthoritative TQA Section 9561 guidance published in September 2026. In practice this means your agent's access controls, logs and change management get evaluated under the same CC6 to CC9 criteria your other systems already follow.
What is the difference between ISO 27001 and ISO 42001 for an agent?
ISO 27001 governs the agent's credentials and access: is it inventoried, least-privilege, logged and revocable. ISO 42001 governs the agent's purpose and impact: does the organization have a policy, an impact assessment and life-cycle oversight for why this agent exists and what it is allowed to do. The two overlap on process controls but ISO 42001's impact assessment has no ISO 27001 equivalent.
Is a system prompt enough to satisfy human oversight requirements?
No. A system prompt is an instruction inside the model's own reasoning, not an external, enforceable constraint. Article 14 of the EU AI Act and most auditors expect human oversight to sit outside the agent's judgment, for example as an approval gate the agent cannot bypass or reason its way around, with a working kill switch.
What EU AI Act deadline applies to high-risk AI agents in 2026?
The only AI Act deadline still landing in 2026 for most organizations is Article 50 transparency, already in force since 2 August 2026, plus a new content-marking grace period ending 2 December 2026 for systems already on the market. The Annex III high-risk deadline was moved to 2 December 2027 by the 2026 Digital Omnibus.
Does Rerun make my agents SOC 2 or EU AI Act compliant?
No, and it is worth being direct about that. Rerun does not grant certification and does not make a system legally compliant by itself. Certification and conformity assessment are performed by accredited third-party auditors against your full program. What Rerun provides is the runtime evidence an auditor or conformity assessment actually asks for: least-privilege agent identity, approval gates on irreversible actions, and an exportable, immutable action log.
What is the fastest way to get audit-ready for an AI agent?
Start with the three questions every framework asks: what can the agent do, what did it do, and who approved it. Scope every agent's tool access to least privilege, put an approval gate in front of irreversible actions, and make sure every action is logged with identity, inputs, decision and outcome. Those three controls cover most of SOC 2's CC6 to CC9, ISO 42001's operation and audit clauses, and EU AI Act Articles 9, 12 and 14 at once.








