4 Questions CIOs Should Ask Before Deploying AI Agents  

There’s a throughline across just about every conversation I have these days with enterprise leaders. They know they need to move forward on AI. They want to do it. But the news from Silicon Valley is giving them pause. Agents escaping test environments, accessing systems they shouldn’t, taking actions nobody authorized, and some of the biggest names in AI publicly warning about the risks.

What do you do with that if you’re trying to make smart decisions about where and how fast to move?

I unpacked this recently with Ensono’s Chief AI Officer, Jim Piazza, and what came out of our conversation was equal parts reassurance and clarion call. The incidents making headlines look like the machines in The Matrix, but there’s more to learn from them than there is to fear. Those lessons have immediate application for leaders trying to navigate the path forward.

Why AI Agents Take Unintended Actions

AI agents are optimizers. Give one a goal and it will pursue it, hard and creatively, within whatever boundaries you’ve put in place. If those boundaries aren’t tight enough — or aren’t there — the agent will find a way around them.

That’s the core lesson from the recent high-profile incidents: well-intentioned instructions, interpreted too literally, with consequences nobody signed up for.

The broader market data tells a similar story. The Forrester State of Agentic AI, 2026 report finds that while three-quarters of enterprise leaders say they’re adopting agentic AI, only a small minority have reached meaningful production applications. The gap, Forrester finds, isn’t about the technology — it’s about the organizational and architectural discipline required to use it well, what the report calls “orchestration maturity, executable governance, and disciplined nonhuman identity.”

In other words: the agents are ready, but most enterprises aren’t.

Good Intentions Aren’t Enough

What I find myself telling clients right now is that the teams behind these incidents weren’t reckless. They were experts who thought they had the right guardrails in place, acting in good faith, and it still went wrong. I say that not to alarm people, but because it’s important for enterprise leaders to be honest about what it means. If it can happen to AI experts at leading labs, it can happen to any organization that isn’t being extremely intentional about how it builds and governs agents.

Most organizations are rigorous, yet the technology is moving faster than their teams’ ability to govern it. That gap is where the risk lives. If you’re a CIO considering your AI roadmap, you need to look at whether you have the right people, controls and architecture to make sure your agents can only do what you intend them to do, and then look beyond that to the longer-term question of how to manage those agents in production.

How You Build Agents Matters as Much as What You Build Them To Do

Jim’s thinking has shaped a lot of how I talk about this with clients. The short version: most agents today reason probabilistically. They assess a situation and determine the best course of action based on what they know.

That reasoning is valuable, but “best” is defined by the agent, within whatever constraints you’ve given it. If those constraints aren’t precise enough, you get a very different outcome than the one you intended.

The fix Jim lays out in detail in this blog post pairs that reasoning with deterministic execution. That means the agent can think freely, but it can only act within a defined set of options you’ve approved in advance. It’s a practical way to close the control-plane gap Forrester flags as one of the biggest blind spots in enterprise agentic deployments today. If you want the full architecture — including how identity and audit trails fit in — that report is worth a read on its own.

Don’t Let Fear Make Your AI Decisions

AI has genuine, significant benefits for enterprises right now: operational efficiency, cost reduction, accelerated modernization, better and faster decisions. These aren’t hypothetical future benefits. I see enterprises realizing them every day, and the competitive cost of standing still while others move is real.

The fear these stories have generated is understandable, but fear isn’t a strategy. The enterprises that come out ahead won’t be the ones that moved fastest, and they won’t be the ones that waited longest. They’ll be the ones that moved deliberately, with a specific business problem in mind and real capabilities behind them.

4 Practical Questions to Ask Right Now

• What specific business problem are we trying to solve? The enterprises getting the most value from AI started with a concrete, well-defined problem, not a general mandate to “do AI.” Specificity is what separates a successful deployment from a sprawling initiative that’s hard to govern.

• Do we have actual controls, not just a policy document? Do you know what your agents are allowed to do and not do? Do you have kill switches? Do you have humans in the loop until the system has proven itself?

• Do we have the internal capability to build this safely? This is the honest question a lot of organizations need to sit with. If the answer is no or not yet, that’s not a reason to stop. It’s a reason to find the right partner.

• Are we moving? Because your competitors are, and the gap compounds over time.

AI agents are already in the real world, taking real action. That’s not going to slow down. The organizations that figure out how to engage with this technology — capturing the upside while taking the risks seriously — are the ones that will set the pace. That’s a solvable problem, but it requires moving with both urgency and intention, starting with asking the right questions, right now.

The Forrester report, The State of Agentic AI, 2026, goes deeper into this topic. Get complimentary access. 

Information in Forrester publications is based on Forrester’s efforts to compile and analyze the best resources reasonably available to Forrester at any given time. Opinions reflect judgment at the time and are subject to change.  This report is part of a broader collection of Forrester resources, including interactive models, frameworks, tools, data, and access to analyst guidance. 

---

Frequently Asked Questions:

What should CIOs consider before deploying AI agents?

CIOs should start by identifying a specific business problem for AI to solve, then assess whether the organization has the controls, architecture and internal capabilities to deploy agents safely. They should also consider the cost of waiting, since delaying AI adoption while competitors move forward can create a widening competitive gap.

What controls should enterprises have in place for AI agents?

Enterprises should clearly define what AI agents are and aren’t allowed to do, establish mechanisms to stop agents when necessary, and maintain human oversight until agents have demonstrated that they can operate reliably. Technical controls should enforce these boundaries rather than relying on governance policies alone.

Why can AI agents take actions that weren’t intended?

AI agents are designed to pursue goals and determine how best to achieve them within the boundaries they are given. Because their reasoning is probabilistic, insufficient or imprecise constraints can lead an agent to pursue a course of action its designers didn’t anticipate or intend.

How can enterprises deploy AI agents safely?

Safe AI agent deployment requires both strong governance and technical controls. Organizations should pair AI’s probabilistic reasoning with deterministic execution that restricts agents to approved actions, while also establishing appropriate identity, access controls, auditability and human oversight.

Should enterprises wait until AI agents are less risky to deploy them?

Waiting carries its own risk. AI agents can already deliver benefits such as operational efficiency, cost reduction, faster modernization and improved decision-making. Rather than choosing between moving quickly and waiting, enterprises can move deliberately by starting with well-defined business problems, putting appropriate controls in place and expanding agent autonomy as capabilities mature.