Contact Us
JumpCloud Logo

The EU AI Act’s Hidden Identity Requirement

Why “Who Did This?” Is About to Become a Compliance Question

High-risk AI systems operating in the European Union will face a record-keeping mandate.1 High-risk systems must automatically record their actions across their entire operating lifetime.1 And the organizations that provide or deploy high-risk systems then have to keep these records for a minimum period.2

Read quickly, this sounds like a simple data retention policy. It’s more than that.

The EU AI Act isn’t simply asking organizations to keep a record of what an AI system did. It’s asking them to answer a harder question. Who is accountable for it? Not which system. Not which department. Which person.

There’s a gap here.

According to our Agentic IAM Pulse Report, 72% of organizations already have AI agents in production. Fewer than four in ten have fully integrated these agents into their identity and access management systems.³ 

That distinction is the subject of this guide. It holds regardless of exactly when or how any single enforcement deadline lands. The requirement itself, and the gap most organizations have in meeting it, exists today.

This guide walks through three things:

  1. What the Act’s logging and retention requirements for high-risk systems establish.
  2. Why that amounts to an identity requirement, not just a records-management one.
  3. And what it takes to bridge the divide: a verified link between every AI agent action and the person accountable for it.

What the EU AI Act Actually Asks For

Before we can get into what the EU AI Act means for how you govern and manage AI systems, we need to start by going through what the law says. Extensive obligations attach to systems that fall into the Act’s high-risk category under Annex III.

  • Automatic Logging, Built In

    Article 12 of the Act requires high-risk AI systems to be technically designed so they automatically record events, or logs, across their entire operating lifetime.¹ This is a design requirement that exists for one stated purpose: traceability.

  • Who Keeps the Record

    Article 19 states that providers of high-risk AI systems must retain these automatically generated logs for a minimum period. That’s six months at a floor, longer if a sector-specific rule requires it.² The duty doesn’t stop with the provider. Article 26(6) requires the deployer, the organization using the system under its own authority (most likely yours), to keep the logs under its control for the same minimum six-month period.⁵

  • Traceability Over Time

    Put these two requirements together and they don’t look like a simple file-keeping exercise. They’re a demand for reconstruction. It’s the need to show, after the fact, exactly what a system did, for as long as the law requires. A log that nobody can act on isn’t traceability. It’s a blinking light on a dashboard you can’t do anything about.

That distinction, between keeping a record and being able to use it, is what reframes what the Act is asking you to do.

Why a Logging Requirement Is Also an Identity Requirement

Reconstruct what a system did, to what end? Reconstruct it for whom? Those questions are where the Act’s underlying ask comes into focus.

What “The Natural Persons Involved” Actually Means

Read Article 12 closely and the target of all that logging isn’t only an event. It’s a person. For one category of high-risk system, remote biometric identification, the Act is explicit: Article 12 states that the logs must capture the identification of the natural persons involved in verifying the system’s results.1 In practice, considering the Act’s traceability purpose and its human-oversight requirements, that means an AI agent’s actions should be mapped back to a specific human, not just record that an action took place.

Two Different Questions: What Happened, and Who’s Accountable

Most enterprise logging, the kind your existing audit trail likely already produces, is built to answer one question: what happened. It wasn’t built to answer a second, harder one: who is accountable for it.

According to the Agentic IAM Pulse Report, 83% of organizations lack clear security ownership over their AI agents’ actions.³ Having logs is one thing. Being able to say who’s responsible for what’s in them is another. Most organizations have only built the first.

  • A Gap Regulators Didn’t Name, But Exposed Anyway

    The people who drafted the Act weren’t writing to a specific enterprise identity failure. They didn’t have to. The traceability standard they wrote is exacting enough to surface one on its own.

    We call that failure the accountability gap. It’s the space between an AI agent taking an action and any organization’s ability to name who’s responsible for it.

    Could your organization answer that question today?

Why Legacy IAM Can’t Answer “Who Did This?”

For most organizations, the answer is no. Not for lack of trying. It’s because the credentials AI agents run on were never built to answer an accountability question in the first place.

  • The Default State: Shared Accounts and Static Keys

    AI agents typically operate under shared service accounts, long-lived API keys, and static tokens issued to applications, not people. That’s the pattern that breaks the link between an action and a person. 

    49% of organizations rely on long-lived API keys for their AI agents. 37% use shared service accounts. Among organizations running AI agents in business-critical workflows, long-lived API key use climbs to 71%.³ 

    That means the most sensitive environments are also the most exposed. Only 37% of organizations report their AI agents are fully integrated into formal identity and access management.³

  • When Agents Chain Actions, Oversight Falls Away First

    The problem compounds when an agent chains several autonomous actions together. Each additional step separates the outcome further from any single accountable decision. Traditional identity tools have no native concept of an agent’s lifecycle to combat that drift. The data shows oversight trending the wrong way as autonomy increases.

    Six months ago, 40% of organizations required human review before high-risk AI actions. Today that figure has dropped to 25%. Full autonomy, with no human review at all, more than doubled over the same period. It rose from 11% to 26%. Oversight is declining exactly where the stakes are highest.

  • Zombie Agents: The Gap’s Starkest Symptom

    Some agents operate with no oversight at all, and keep running long after anyone remembers deploying them. We call these zombie agents: active, credentialed, still acting, with no one left who can answer for what they do. The majority of organizations couldn’t quickly shut a zombie agent down even if they found one lingering in their environment. 55% have no centralized kill switch for AI agents. A third say their only option is disabling agents manually, one system at a time.³ 

    Ownership is equally fragmented. Only 17% of organizations have a named security leader accountable for AI agent actions. Nearly half default that responsibility to IT, by omission rather than design.³

  • Checkmark

    49%

    Organizations relying on long-lived API keys to run AI agents

  • Checkmark

    55%

    Organizations with no centralized kill switch for AI agents

  • Checkmark

    17%

    Organizations with a named security leader accountable for AI agent actions

Source: The Agentic IAM Pulse Report, JumpCloud, 2026

What It Actually Takes to Build Agent Accountability

JumpCloud Agentic IAM manages the entire agentic lifecycle with discovery, registration, management, and governance from one platform.

These are the tools and practices that take an organization from an unmanaged fleet of agents to an accountable one.

  • Discover: Find Every Agent, Including the Ones Nobody Registered

    Closing the accountability gap starts with knowing what exists. JumpCloud’s discovery layer surfaces sanctioned and unsanctioned AI agents across devices, browsers, and on-premises environments. It builds the authoritative inventory you must have before you can take any other agent-related action. You can’t govern what you can’t see.

  • Register: Give Every Agent a Named, Accountable Owner

    Once discovered, every agent needs a formal identity record. JumpCloud captures every agent’s purpose, creator, scope of action, and a named human owner accountable for what it does. It anchors an agent’s actions to a human user at the moment of registration, not after something goes wrong.

  • Manage: Scope What Each Agent Can Do, Not Just What It Can Reach

    Registration establishes who owns an agent. Management determines what it’s allowed to do. JumpCloud incorporates real-time risk monitoring and device health checks to make sure agents are only working on healthy, managed devices and accessing the appropriate data. A verified identity running on a compromised or unmanaged device isn’t secure.

  • Govern: Keep the Record, and Keep Watching

    Access granted isn’t access permanently justified. JumpCloud provides the ongoing logs, audit trails, and access reviews that support the Act’s traceability standard. This isn’t a one-time exercise at deployment. It’s continuous.

Organizations that put this framework to work aren’t just setting themselves up for easier tracking and evidence pulls. They’re faster. Our research found that organizations with governance in place are three times more likely to scale AI agent use without limits.³ 

Governance isn’t the brake on AI adoption. For the organizations already building it, it’s the thing that lets them move faster than everyone else.

Where to Start

You now understand that the EU AI Act’s requirements go beyond record-keeping requirements. It’s about traceability, oversight, and accountability. If you want to see the full picture of how other IT leaders are governing, or failing to govern, their AI agents, read the Agentic IAM Pulse Report. It goes deeper into where the gaps are widest and what closing them looks like in practice.

Get the Full Report

Sources

  1. Regulation (EU) 2024/1689 (EU AI Act), Article 12. https://artificialintelligenceact.eu/article/12/ 
  2. Regulation (EU) 2024/1689 (EU AI Act), Article 19. https://artificialintelligenceact.eu/article/19/ 
  3. The Agentic IAM Pulse Report. JumpCloud, 2026. https://jumpcloud.com/resources/agentic-iam-pulse-report 
  4. Q3 2026 IT Trends Report. JumpCloud, 2026. https://jumpcloud.com/resources/q3-2026-it-trends-report 
  5. Regulation (EU) 2024/1689 (EU AI Act), Article 26. https://artificialintelligenceact.eu/article/26/