Connect
This deadline debate is consuming boardrooms right now. Legal teams have spent months tracking the EU AI Act’s amendment packages and parsing whether high-risk system obligations will land in August 2026 or slide toward late 2027. It’s a reasonable thing to watch, but it’s also the wrong thing to fixate on.
The questions the regulation raises don’t change whether enforcement arrives in eight weeks or 18 months. And most organizations deploying or starting to experiment with AI agents today can’t answer them.
You May Not Be a Provider. That Changes Everything.
Much of the compliance guidance circulating right now treats AI governance as a model-building problem where data governance is a necessity for training sets, technical documentation, conformity assessments, and lifecycle risk management systems. That depth matters if you’re an AI provider building systems from scratch. But the vast majority of mid-market organizations aren’t building models. They’re buying them, connecting them to internal data, and turning them loose on real workflows.
In the language of the EU AI Act, that may make your organization a deployer, and deployer obligations under Article 26 look fundamentally different from provider obligations. Broadly, the Act asks deployers of high-risk systems to assign meaningful human oversight, operate systems in line with the provider’s instructions for use, monitor how they behave in practice, and retain the logs those systems generate.
Whether or not a deployment qualifies as high-risk is a different question, but it’s worth noting that Annex III explicitly lists AI used in the recruitment and evaluation of job candidates among its high-risk categories. If AI agents are screening résumés or scoring applicants in your organization, high-risk deployment is in scope.
Notice what’s missing from the deployer’s list: nobody is asking you to rewrite model weights or audit training data. The obligations are operational. They’re about knowing what’s running, who’s accountable for it, and being able to show your work.
The Core Question: Who Governs Non-Human Access?
An audit conversation won’t open with an ethical discussion of your AI strategy. It will sound like this:
- What systems did this agent access?
- Who authorized that access, and when?
- Was the access scoped to what the task required, or did the agent hold standing permissions?
- Which human was overseeing it?
- Where is your record?
Notice what’s missing. None of these questions are about models. They’re all about identity: the same questions identity and access programs have been answering about human users for two decades.
What’s changed is the subject. An autonomous agent doesn’t log in at 9 and log out at 5. It acts at machine speed, invokes APIs, calls external tools, and because of its non-deterministic nature, it chooses its own path through your systems at runtime. The frameworks built around predictable human sessions weren’t designed for a workforce that behaves this way.
This is the gap forming inside organizations right now. Agents created by enthusiastic teams and wired into corporate data, while IT has no inventory of what exists, no named owner, and no record of who approved what these agents are permitted to touch. You were used to tackling shadow IT. Now you’re facing shadow AI, which is less a future risk than a current visibility and control problem.
Visibility Only Gets You So Far. You Need Governance.
A common first response is visibility: put monitoring on the network and API layer, watch agent traffic, flag anomalies. That observability has real value for threat detection, and most security programs will want it.
But traffic analysis is a detective control. By the time an anomalous data transfer appears on a log, the boundary has already been crossed and the data has already moved. Detection tells you what happened. Oversight, in the sense regulators and boards mean it, is about what’s permitted to happen. You need controls in place that provide guardrails before an agent acts, not just forensics assembled after.
A governance posture built entirely on watching traffic leaves a vacuum at the front door: anything can request access to anything, and you see the impact later. That’s a difficult story to tell an auditor. It’s an even worse one to tell your board after an incident or a potential penalty.
What Governance Can Look Like
The teams getting ahead of the governance question are moving before the deadline. They’re building a set of foundational behaviors that hold up regardless of when enforcement lands. The same architecture that produces a defensible audit trail is the architecture that makes autonomous systems safe to scale.
Here’s how it takes shape:
- An inventory with owners. Every agent and integration is known, uniquely identifiable, and mapped to a named person who is accountable for its behavior. No anonymous automation.
- Access that expires. Agents receive the permissions a specific task requires, for the time the task requires them. Not permanent, open-ended credentials that outlive their purpose and accumulate risk.
- Authorization that precedes action. A recorded human decision sits in front of what an agent is allowed to do. It’s not something you reconstruct from logs after the action already happened.
- Records by default. Evidence of who approved what, when, and why is a byproduct of how access works. Not a quarterly archaeological dig across spreadsheets and traffic captures.
This isn’t a compliance checklist, and that’s the point. It’s what a disciplined identity practice looks like when the workforce includes agents that act on their own. The upcoming regulation is converging on principles security leaders need in place anyway. The deadline just sets a tempo.
Build for the Questions, Not the Date
Maybe the high-risk obligations bite in August 2026. Maybe Brussels grants more runway. Either way, the questions are already here and they’re showing up in vendor assessments, in board packets, and in deals. The organizations that can answer them won’t be the ones that guessed the enforcement date correctly. They’ll be the ones that started keeping honest accounts of their autonomous systems before anyone forced them to do it.
We recently surveyed IT and security leaders about exactly where these non-human control gaps are forming and how their teams are adapting. Read the full Q3 IT Trends Report for a closer look at how your peers are closing the visibility gap.
