Rohit Kedia is a software architect turned CEO of Xoriant, a AI-native Applied Intelligence digital engineering company.
I have spent the better part of the past year inside enterprise AI rollouts across different industries and different technology stacks. There is the same pattern I see no matter where I look: The models themselves keep getting better, but the value companies actually derive barely moves.
That gap is not just something I feel from where I sit. Leading analysts have put a number on it, too, showing the staggering number of AI projects that never deliver a meaningful business outcome.
Not Just A Technology Gap
Raw AI capability has been advancing at an almost exponential pace while enterprise adoption has followed a shallow, linear line by comparison, indicating an enterprise value gap. But the more pilots I have watched succeed (and then fail), the more convinced I have become that there is a second gap sitting underneath this. I call it the “human agency gap.”
Most companies treat AI adoption as something that happens to the organization rather than something the organization deliberately directs. When leadership announces an AI mandate, almost always, no one sits down and decides what the AI actually owns versus what a human is still accountable for. So while the AI pilot looks promising in a demo, it stalls once it meets the mess of production.
The technology gap gets all the attention because it shows up in dashboards and postmortems, while the human agency gap is the one that actually explains why so many technically sound pilots never make it to scale.
Where HI Meets AI
That brings me to the principle I keep coming back to: In the world of AI, humans have the ultimate agency. And the organizations that internalize this are the ones building something durable, which is exactly why I think of the real work here as human ingenuity (HI) plus AI, rather than AI on its own.
I do not mean “HI plus AI” as a comforting slogan designed to make people feel better about automation. I mean it as an engineering fact about the kind of systems we are actually dealing with. AI systems today are probabilistic by nature, which means they do not give you the exact same answer every time you ask. This is a very different approach from the deterministic software most enterprises grew up running. A probabilistic system needs a human anchor at the points in the process that actually matter. Take that human anchor away, thinking you are buying more autonomy in exchange, and what you actually get is drift.
How Human Agency Actually Moves As AI Matures
I have watched organizations mature in how they use AI, and there is a recognizable arc to it. But the more interesting insight is not the arc itself; it is what has to be true underneath it for human agency to relocate safely instead of just evaporating:
1. The first is whether the data underneath the AI can actually be trusted.
2. The second is whether the actual work has been redesigned around what agents can do, rather than agents being bolted onto workflows that were built for a different era of software.
3. The third, which is the one I think gets the least attention relative to how much it matters, is whether there is real operational discipline once the agents are live.
There is a useful way to think about which parts of this should stay close to a human and which parts should not:
• Some of the work is fundamentally about domain knowledge and intent—deciding what the business thesis even is, prioritizing what matters and encoding the context that gives an agent any hope of getting the nuance right. That work does not scale by removing the human from it, because the human’s judgment is the entire value of it.
• Other work is fundamentally about technical execution at scale—running the same well-defined pattern reliably, over and over, across thousands of instances. That work can scale even by removing the human from the loop.
Early on, agency looks like an individual using an AI tool with no real coordination across the organization. As the workflow gets redesigned and the data foundation solidifies, agency shows up as a human validating output before it goes anywhere. Later, once the operational discipline is mature enough to trust at scale, you start seeing agent chains run entire processes with very little manual intervention. And at the most mature stage, signals and events trigger the work automatically, with humans setting intent upfront and holding the AI accountable to that intent.
From the outside, it may look like humans are stepping back at every stage. But look closer, and you will see the opposite is true. Human agency does not shrink as this progresses. It relocates, moving from “doing” the task to “deciding” what the task should even be in the first place, and then to judging whether the outcome actually makes sense.
What This Looks Like In Practice
At a Fortune 100 bank we worked with, agentic AI took over the cross-document reconciliation work sitting behind regulatory reporting. Roughly 1,100 analysts were freed from that grind, and the system reached 95% accuracy on root-cause analysis.
But the more important point that actually matters for this conversation is what happened to those analysts afterward. They did not become less important to the business. They moved from “checking” whether numbers matched to “interpreting” what the numbers actually meant for the bank’s risk posture.
Where The Real Advantage Will Come From
I do not think the enterprises that win the next phase of this will be the ones with access to the best underlying models. The enterprises that win will be the ones that figured out exactly where human agency creates the most value inside their specific business, and then built their organization to protect that deliberately. That is the real work of applying intelligence. Not simply using AI, but applying it with a very clear and deliberate sense of where the human still holds the wheel.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?