Enterprise AI impact isn’t about the most powerful model—it’s about the smartest steering
This is something you have probably noticed several times—at least, I certainly have. When you use an AI chatbot for personal questions, decision-making, research, and the like, it works amazingly well. But when you try your company’s AI . . . well, the results tend to be much less impressive, unless we are talking about purely “administrative” uses.
Why is this happening? Basically, because when you use it, you yourself are doing the steering, the corrections, the reframing, the “forget about this, give me more of that,” until you decide that the answer is good enough. Doing all that supervisory work is pretty much invisible, because it feels natural, like when you do it with a coworker or a subordinate.
But when you switch from personal help to running a process, the steering function has to appear someplace else, and that’s not necessarily easy. The move from assistants to human-agent teams, and later to “human-led, agent-operated” workflows is much more problematic, much less natural, and generates many more issues.
Understanding the distinction between assistance and autonomy
The essence is in the nature of the roles: While copilots propose, humans typically judge. That—in the context of corporate processes that can span from several hours to several days or more, and can involve different systems and departments—creates an essentially different managerial concern: Who checks what? Is this sequence of actions still heading toward the outcome we wanted?
These long-running agents clearly create an open problem, since even highly capable ones from frontier models tend to struggle over long horizons, with a lot of sessions involved and different context windows, unless we have some sort of surrounding system that helps them maintain progress and continuity in an efficient manner.
We can imagine a lot of examples. An AI running a customer retention process, for instance, has to draft messages, choose or make offers, schedule follow-ups, and update the corporate CRM at every step of the way. But if a discount campaign that works to improve renewals is damaging margins or causing customers to churn six months later, who’s going to notice that? That’s precisely what steering means: not just generating the next action, but making continuous judgments and adjustments on whether the whole process still moves toward our goals.
The model alone is not enough
Right now, I believe that the main challenge is no longer to prove that agents can work, but to provide them with the layer of reliability that high-value workflows require: policies, permissions, monitoring, escalation rules, and production feedback. Even OpenAI has created a product, Presence, designed for these concerns.
As a CEO, would you give a brilliant employee total freedom with no objectives, no constraints, no supervision, and no feedback? Your answer (and mine too) will probably be “no way, José.” However, companies are now talking about autonomous agents as if autonomy itself were the ultimate goal, when in fact, it is not. The real goal is to get better outcomes, whereas autonomy is only useful when it helps the company move closer to these outcomes. And yes, I know “autonomous” is the fashionable word right now.
According to Gartner, more than 40% of agentic AI projects will get canceled by the end of 2027 due to cost concerns, unclear business value, or inadequate risk controls, and many of the current projects are being misapplied or are still in a proof-of-concept phase. And as I see it, the problem is not a lack of enthusiasm, but a lack of a reliable control layer around autonomous work.
Governance, in plain English
Someone has to decide what an agent may do, what it may not do, when it must stop, and when a human has to take over. The more autonomy that agents get to have and the less direct human oversight they receive, the more room there is for misunderstanding or for unintended actions. And this is where I think we are confusing two completely different things: Intelligence is not the same as control.
Sure, a model can be great when it comes to generating ideas, interpreting language, or choosing possible actions, but control means something completely different: staying aligned to the original goal regardless of how much the context or the circumstances may change. That’s the essence of management. Great companies don’t just hire smart people and walk away. They also define goals, budgets, decision rights, escalation rules, incentives, and review cycles.
McKinsey’s research clearly supports the organizational side of this idea. The firm usually associates workflow redesign with the bottom-line impact of generative AI, and CEO oversight of AI governance appears to be correlated with higher self-reported EBIT impact. Success, according to McKinsey, is not about having smarter models at the bleeding edge of the frontier and dropping them into our existing processes, but about redesigning these processes—how work is directed and controlled.
Microsoft provides another interesting point. Their considerations on “companies becoming learning systems” point toward constraints that are organizational, not individual. In fact, many employees are now moving faster than their own organizations.
Let’s ask CEO-level questions
Whenever I discuss this with executives in my classes, the instinctive response is usually to ask which model is best. I increasingly think that is the wrong question. The CEO-level questions are not “how smart is the model?” or “how many agents can we deploy?” but something akin to “when this system is acting without someone supervising every step, what is keeping it pointed toward the business outcome we really care about?”
We have spent three years making AI astonishingly capable. Now we are beginning to discover that capability without steering is not autonomy; it is drift. The next enterprise breakthrough will come from putting something competent in the empty seat.