Skip to content

The Confidence Gap: 92% of Executives Are Certain AI Is Working. Half of Them Don't Know Who Is Using It.

Blog  ✺  AI  ✺  Strategy  ✺  Governance

There is a kind of executive confidence that should make you nervous. Not the kind that comes from a dashboard. The kind that comes from a feeling. A study published in February put a number on it: 92%.

That is the share of C-suite leaders expressing full confidence in AI's impact on their organization. In the same survey, nearly half admitted they do not know who is using AI or how. Both numbers came from the same people. About the same company.

TLDR: A January 2026 study of 365 senior leaders at companies with 1,000+ employees (TrendCandy for Larridin) found 92% of executives fully confident in AI's impact, while 58.2% cite unclear or fragmented ownership as their primary barrier, 62% lack a comprehensive inventory of the AI applications in use, and 75% have no fully implemented governance program. Organizations run an average of 23 AI tools. Half do not measure outcomes at all or rely on surveys instead of real data. Three-quarters expect ROI within six months. The gap between conviction and evidence is not a measurement problem waiting for better tooling. It is a governance vacuum that measurement cannot fill.

The number that does not belong

Take the survey findings one at a time and each is unremarkable. Governance lags adoption. Ownership is fuzzy in the early phase of any capability. Tool sprawl happens. These are the ordinary growing pains of a technology moving faster than the organization absorbing it.

Put them next to the confidence number and something breaks.

An executive who reports that 58.2% of the barrier is unclear ownership, that no one has an inventory of the 23 tools in play, and that outcomes are measured by asking people how they feel, is describing an organization with almost no observability into its own AI estate. That executive then reports full confidence in the impact.

Confidence in what, exactly? Not in a measured result. Only 25% of organizations in the study said they could reliably measure end-to-end AI adoption. The confidence is not resting on evidence. It is resting on the absence of contradicting evidence, which is a very different thing and behaves very differently under pressure.

I wrote in R(O)AI that the industry cannot prove its AI spending is working. This study suggests something sharper. Most organizations have not built the apparatus that would let them find out.

What confidence is actually made of

The study does offer some real signal. 82% report faster project delivery. That is a genuine finding and worth taking seriously.

Then look at the magnitude. Nearly 86% experienced time savings of under 10 hours per month per employee. Ten hours a month is roughly half an hour a working day. That is real. It is also the kind of gain that disappears entirely into the noise of a P&L, because saved minutes are not recovered capacity unless someone deliberately redeploys them. Nobody does. The half hour dissolves into slightly longer coffee breaks, slightly more Slack, slightly less pressure. This is not a criticism of employees. It is how organizations work when no one owns the conversion of time saved into value created.

So executives have a true fact — things feel faster — and they are extrapolating from it to a conclusion the fact cannot support. Faster delivery is an input. Impact is an output. Between them sits workflow redesign, which McKinsey's datahas repeatedly shown is the single organizational attribute most strongly correlated with EBIT impact, and which most organizations still have not attempted.

Confidence is being manufactured from the input side because the output side is dark.

The ownership problem is the whole problem

30.5% of respondents named unclear responsibility for measurement as a barrier. 27.7% named fragmented ownership across teams. Those two figures describe the same failure from opposite ends, and together they account for the 58.2% headline.

Here is why that matters more than the governance statistic everyone will quote. A missing governance program is a document problem. You can write the document. A missing owner is a structural problem, and it produces a specific pathology: every function can start an AI initiative and no function is accountable for whether it worked.

That is how you get to 23 tools. Nobody bought 23 tools. Marketing bought three, operations bought four, IT sanctioned two, finance ran a pilot, and several arrived inside SaaS products nobody evaluated as AI purchases at all. Each decision was locally rational. The portfolio is the accident.

And an unowned portfolio cannot be measured, because measurement requires someone with the authority to define what counts as success and the standing to be wrong about it. 50% of organizations do not measure outcomes or substitute surveys for real-time data. Surveys are what you use when you have no owner. You ask people whether they liked it, because nobody is empowered to say what it was supposed to do.

The six-month clock

Three-quarters of these executives expect AI ROI within six months.

Set that against the rest of the picture. No inventory. No owner. No governance. No measurement. And a six-month deadline on a return nobody has defined and nobody can detect.

There are only two ways this resolves. Either the deadline passes and the confidence quietly persists, sustained by the same absence of evidence that produced it. Or someone senior asks for the number, discovers it does not exist, and the swing from 92% confidence to blanket disillusionment happens in a single quarter. S&P Global already recorded the shape of that swing: the share of businesses scrapping most of their AI initiatives rose to 42% in 2025 from 17% the year before.

Unmeasured programs do not fail gradually. They fail all at once, on the day someone asks.

red and white stop sign
Photo by Levi Meir Clancy / Unsplash

What the 25% are doing differently

The organizations in the study that could reliably measure end-to-end adoption were not distinguished by better analytics tooling. They were distinguished by having answered three questions that the other 75% have not.

First, who owns this. Not who sponsors it, not who is excited about it, but which single named executive carries the outcome in their objectives and loses something if it does not materialize. Ownership without downside is sponsorship.

Second, what would falsify it. Every AI initiative should ship with the condition under which it would be shut down. If you cannot state that condition, you have not defined the outcome, you have defined an aspiration. Aspirations cannot be measured because they cannot be wrong.

Third, where does the saved time go. Ten hours a month per employee is either a rounding error or a redeployable asset, and which one it becomes is a management decision made in advance, not a benefit that accrues automatically. The organizations extracting value decided beforehand what the freed capacity was for.

None of these three require a platform. All three require someone to accept accountability, which is why they remain undone.

The pattern underneath

We have been here before, and the previous rounds are instructive because they resolved.

Nobody today runs an ERP program without a named business owner, a defined benefits case, and a post-implementation review. Nobody runs a cloud migration on vibes. Those disciplines exist not because the technology demanded them but because a generation of expensive failures made them non-negotiable. The governance came after the wreckage, as it always does.

AI is not exempt from this sequence. It is simply early in it, and MIT's GenAI Divide study — which found only 5% of implementations producing measurable P&L impact — is what the middle of it looks like. The 92% confidence figure is what the beginning looks like from the inside, and it will not survive contact with the first serious board question about return.

The uncomfortable part is that the fix is not technical and never was. It is the least glamorous work available: naming an owner, defining a kill condition, writing down what the time saved is for. Work that no vendor can sell you and no model release makes easier.

Confidence is not a result. It is what you have instead of one.

Mehr

The Jagged Frontier: AI Makes You Faster at the Easy Stuff. It Doesn't Help With the Hard Stuff.

The Limiting Reagent: Why Altman's Own Warning Undercuts OpenAI's Enterprise Vision