Your AI Was Right. Did Anyone Do Anything About It?

This voice experience is generated by AI. Learn more.This voice experience is generated by AI. Learn more.Subrata (Subu) Biswas is the co-founder and CEO of Cimba, a company building proactive AI for business operations.

gettyThe conversation I keep having with operations leaders is not about whether the software was right. It’s about a recommendation somebody saw, agreed with and then didn’t act on. The reasons are rarely dramatic. Usually the person who would have had to make the change was three meetings deep in something else, or the number landed on a Thursday when the call had already been made on Tuesday. Sometimes someone got burned moving fast last quarter and wasn’t going to do it twice. What stays with me is how rarely any of that gets counted as a system failure. The model was right, so the system gets its credit and the quarter still ends the way it was going to end anyway.

For the past three years, the pitch on enterprise AI has overwhelmingly been about access. Give the team a forecast instead of a hunch. Give the ops manager a flag where there used to be nothing at all. The assumption underneath was that people were making poor calls because they lacked information, and that handing them better information would change the output.

McKinsey’s 2026 global survey shows how that bet is playing out. Forty-four percent of organizations now report scaling AI across the enterprise, up from 38% a year earlier, and 80% of respondents say it has made them personally more productive. But the share attributing any EBIT impact to AI sits at 37%, about where it was the year before, and only 6% clear McKinsey’s bar for high performers. So, deployment went up, but the earnings line didn’t move with it.

Ask why the 94% who fall short of that bar have not seen it show up, and the answer is rarely that the model was wrong. The more useful question is one that almost no vendor evaluation asks: What happens after the system is right?

Most operational AI closes its feedback loop at the prediction. It logs whether the forecast matched reality and stops there, rarely tracking whether anyone acted on it and almost never asking why not. A recommendation nobody acts on produces no business outcome, however accurate it was.

This is not unique to enterprise software, and the starkest version of it comes from medicine. A 2024 meta-analysis in the Health Informatics Journal pooled 16 studies and found that physicians override drug interaction alerts roughly 90% of the time. The explanation the literature gives is alert fatigue. So many of the alerts carry little clinical value that dismissing them becomes reflex, and the ones that matter get dismissed along with the rest. Trained clinicians, in a setting where the cost of being wrong is measured in harm, still wave the system away nine times out of 10. If accuracy on its own doesn’t change behavior there, it won’t change behavior in a merchandising or logistics workflow.

Two things had to become true at once, and neither one was true five years ago.

Predictive models alone were never going to solve it. A probability is not a decision, and forecasting tools spent a decade getting more accurate while the distance between a forecast and an action stayed the same. Language models alone weren’t the answer either. They got very good at answering the question someone asked, which did nothing for the much larger set of questions nobody thought to ask, or asked too late to matter.

What is new is putting the two together, so a system can watch dozens of signals at once, weigh which ones are worth a person’s attention today or this week, and phrase what it finds as something specific enough to act on rather than another chart. Neither half does that alone. It also makes the whole thing measurable because there is now a discrete recommendation to accept or decline.

None of this requires new procurement. Start inside the evaluations already running. Vendors get asked about accuracy constantly and almost never about acceptance rate, completion rate or what happened to the recommendations nobody acted on. A vendor who cannot answer that has told you something.

The second test is internal. Pick one report or alert your team already produces and, for two weeks, track how often anyone does something with it. When Gartner forecast in June 2025 that more than 40% of agentic AI projects would be canceled by the end of 2027, it named escalating costs, unclear business value and inadequate risk controls. Unclear business value is the one that surfaces in a two-week test, and it surfaces long before anyone writes the cancellation memo.

This does not work everywhere. A decision that comes around once or twice a year, like a site selection or an annual vendor consolidation, never produces enough repetitions for a feedback loop to learn anything useful from. And the calls that carry real human consequence, headcount among them, should not be reduced to an acceptance rate in the first place.

If you change one thing next quarter, make it the simplest thing available to you. Pick the report your team trusts most and find out what share of its recommendations anyone acted on. Most teams have never measured it. The number is usually lower than anyone expects, and finding it out does not require a purchase order.

The next time someone hands you an accuracy number, ask what happened after the system was right. Did anyone do anything about it? That’s the metric missing from most evaluation criteria, and it is the one that decides whether any of this pays for itself.

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?