COMMENTARY: Every enterprise has been told to become an agentic company. Not in five years. Now. And most are starting in exactly the wrong place: standing up a committee to figure it out.While leadership debates strategy in a governance working group, a marketing director already has three AI tools open and has shipped work the company has no record of. For many companies, the problem revolves around the gap between where leadership thinks the organization has progressed and the actual work taking place on the ground.[SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Read more Perspectives here.]Agentic AI introduces new challenges that go beyond just the adoption of the past. The question of governance no longer relates to simply what AI can say. It’s now what AI can do. Generative AI created a content risk, but agentic AI has created an execution risk. Before, an employee may have just asked AI to draft a quick email. Now, an employee can work with an agent to write code, approve workflows, or access customer data. It’s critical that governance not stop at policies and prompts, but rather extend through all actions.I have spent 30 years in security and identity, and I have watched this movie before. Cloud. RPA. Data governance. Every wave produced a Center of Excellence that promised to lead the transformation and became the thing that throttled it. We are repeating the experiment with AI and expecting a different result.Here are the mistakes I see most often when it comes to agentic AI and what the organizations getting it right do instead.
- The council that becomes the bottleneck: Our instincts are to route every AI project through a central review board. It feels responsible. Within one economic quarter, it’s the slowest thing in the building. It’s the exact pattern that killed Cloud CoEs by 2020 and RPA CoEs by 2023. Renaming it an "AI Council" does not change the physics. It’s the same shape every time: a gate that the business learns to route around.
- Watching the door instead of the room: Most programs measure who has access to which AI tools. That’s the login boundary, and it tells us almost nothing. The moment an agent gets past login, the question we’re no longer asking about who conducted the attack. It’s what can the agent do, on whose behalf, with what scope, for how long. If we can issue credentials fast, but cannot answer what was done with them, we have built a faster way to lose visibility.
- Manufacturing the company’s own shadow AI: Here’s the uncomfortable math: People do not take the safe path because it’s safe. They’ll only take it because it’s faster. Add a ticket queue, a vendor evaluation, and a review board in front of the approved tool, and instead of governance, we’ve created a reason to skip over it entirely. The unsafe path wins by lunchtime, not because it's reckless, but because they have a job to do, they have 14 other tasks on their plate this morning.
- Believing policies and training change behavior: Nobody reads the wiki. People read defaults. The behavioral research predates the AI wave and it’s overwhelming: auto-enrollment moved 401(k) participation from roughly 23% to over 80% without changing the underlying offer at all. The path of least resistance wins every time, regardless of the rules the company wrote. A policy on a Confluence page does not make up a governance program. The default is the governance program.
- Measuring activity instead of outcomes: Counting agents deployed and assets published produces a vanity-metrics machine. The RPA wave died on this rock. The only question that matters is whether the business shipped faster because of the work, not how much work there was.
- Treating AI as a top-down rollout: Don’t think of AI as a department buying a tool. It touches every employee, every customer interaction, every decision under five thousand dollars. It’s not about a team adopting software. It’s an organization renegotiating how people do their work. We cannot push that down from a strategy deck. It has to happen where the work already happens, with the people already doing it.
- Standing up a team to “figure out AI”: It’s the most expensive mistake of all, because it feels like progress. While we’re designing the new function, our competitor's existing people are shipping new product. There’s no separate team that goes off, figures out agentic, and hands it back to the org later. That’s the slowest possible path. The people already doing the work are the team.
