Ravi Palwe, Senior Manager, AI/UX Transformation, Financial Services | Capgemini.
getty
Gartner has predicted that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear value and inadequate risk controls. The risk controls get most of the attention, but the interface rarely does, and in my experience, that is the part that fails first.
Most teams design one interface for their AI agent, pick one level of autonomy and ship both. It feels like a finished product. It is actually the core mistake, because an agent that acts on a person's behalf is never a fixed thing. Two variables are always moving underneath it: how sure the agent is about what it is doing and how much the person has come to trust it. A static interface is blind to both, and that blindness is where trust fails.
I have spent close to two decades designing complex application flows, and the last year designing for agentic systems, since that is when institutions started actually adopting them. The pattern I keep seeing is the same. The system is capable, the interface is fixed, and the fixed interface is the reason people either lean on the agent too hard or stop using it altogether.
Confidence Should Change The Interface
Start with the agent's confidence. Most teams treat it as a number to display, such as a small badge in the corner that reads 82%. They believe they have handled it. They have not. A number in the corner changes nothing, because the person reads right past it. Confidence only matters when it changes the action itself.
So, the interface has to move with it. When the agent is sure, the action should recede. It presents itself and gets out of the way, because asking a person to scrutinize a confident, low-risk decision wastes the one thing they have least of, which is attention.
When the agent is unsure, the same action should look different. It should sit higher in the person's view, signal plainly that it is uncertain and ask for a heavier, more deliberate confirmation before it runs.
You are not informing the person of the agent's confidence. You are spending their attention where the risk is and refusing to spend it where there is none. That is a moving interface, and it is the difference between confidence that means something and confidence that is decoration.
Trust Should Change The Level Of Autonomy
The second thing that moves is the person.
Trust in an agent is not granted all at once. It is earned in increments, and a fixed level of autonomy ignores that completely. Set it too high on day one and the person is unsettled by what the agent does before they have any reason to believe in it. Set it too low and the agent is a glorified suggestion box that never earns the right to do more. Either way, you have forced a bad bet at the exact moment you know the least about how much this person will trust the system.
The fix is to let autonomy migrate. Give the person a dial, from review only, to suggest and edit, to fully automated, and make it effortless to move in both directions. As the agent proves itself on the small decisions, granting it more should take a single step. The moment something feels wrong, pulling it back should be just as fast. Trust is not a setting you configure on the user's behalf. It is a slope a person climbs at their own pace, and the interface exists to make both the climb and the retreat easy.
Calibrate Human Attention To Risk
This matters more as the stakes rise. In a low-stakes tool, a static interface is merely clumsy. In finance or insurance, it is dangerous in two directions at once.
An interface that treats a risky, low-confidence action exactly like a routine one will let a person wave through something they should have caught. An interface that demands the same heavy confirmation for everything will wear that person down until they stop reading at all, which is its own kind of failure. Movement is what keeps trust calibrated. It surfaces the decisions that deserve a human and hides the ones that do not.
The Bottom Line
So, here is what I would leave you with. Do not ship a fixed interface for a system that changes every time it acts. Two things are always in motion, the agent's confidence and the person's trust, and an interface that ignores both will feel wrong long before anyone can say why. Design it to move.
The agents people keep using are not the ones that did the most on day one. They are the ones whose interface met the person where they actually were, and kept meeting them there as that changed.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?