Two-thirds of IT leaders report AI results, but few would interrupt the CEO's vacation over them

Two-Thirds of IT Leaders Say AI Is Delivering, So Why Would Almost None of Them Interrupt the CEO's Vacation?

By · Published September 26, 2026 · Updated September 26, 2026

Here is a strange picture of artificial intelligence in the enterprise. About two-thirds of IT leaders say AI is producing results for their organizations. That is a big number. It suggests the years of pilots, proofs of concept, and endless slide decks have finally turned into something real.

But then comes the second half of the finding, and it changes the whole mood. When you ask those same leaders whether their AI systems are important enough to interrupt the CEO's vacation over, almost none of them say yes.

That contrast is the most useful thing anyone has said about enterprise AI in a while. It tells us that AI has arrived, but not at the center of the business. It is working. It is just not yet load-bearing.

The Two-Thirds Number Is Real, and It Is Also Vague

Let's start with the good news. Two-thirds is not a rounding error or a marketing claim. It means that in most companies, someone can point to an AI system and say, "that thing is doing work."

What kind of work? Mostly the useful, unglamorous kind:

These are genuine wins. They save hours. They reduce frustration. In some cases, they save money you can actually see on a budget line.

But notice something about that list. Almost every item has a human fallback sitting right behind it. If the AI gets it wrong, a person fixes it. If the system goes down, the work slows but does not stop. That is the difference between a helpful tool and a critical system.

The "Wake the CEO" Test

Think about the things that genuinely justify pulling a chief executive off a beach. A payment system failure. A security breach. A regulatory filing that is about to be wrong. A safety issue on a factory floor. An outage that stops revenue.

Those events pass what we might call the wake the CEO test. They are urgent, they are expensive, and there is no easy backup. Someone has to own the problem right now, at the highest level.

Almost no AI deployment passes that test today. And the reason is not that AI is unimportant. It is that AI still has a safety net. A person can take over. The output can be reviewed. The mistake can be caught. The work is recoverable.

Recoverable systems are wonderful. They are also, by definition, not existential. That is exactly why the two numbers in this finding can both be true at once.

Why "Results" and "Urgency" Drift Apart

There are a few forces pulling these two ideas in opposite directions.

1. "Results" is a very loose word

Ask ten leaders what counts as an AI result and you will get ten answers. For one, it is a working demo. For another, it is a few hours saved per employee per week. For a third, it is a hard dollar figure tied to headcount. All three get filed under the same word. That makes the two-thirds figure encouraging, but hard to act on.

2. Most AI wins are incremental

Big, company-changing technology usually rewrites how the business makes money. So far, most AI gains are improvements layered on top of the way things already work. Fifteen percent faster here. Twenty percent cheaper there. That adds up, but it rarely changes the shape of the company. And things that do not change the shape of the company rarely change the shape of a CEO's holiday.

3. Trust is still under construction

Systems that occasionally make things up, drift over time, or need constant babysitting are hard to put at the heart of a business. Leaders know this. They would rather run AI where a mistake is annoying than where a mistake is catastrophic.

4. Nobody wants to be the person who bet the quarter

There is a career risk in going all-in. If the AI works, you get a modest thank-you. If it fails spectacularly, your name is attached to it forever. That math pushes even enthusiastic leaders toward the middle of the road.

What This Means for the Future of AI

The next phase of enterprise AI is not about getting more companies to try it. Two-thirds are already there. The next phase is about moving AI from helpful to load-bearing, from a tool people use to a system the business depends on.

That shift will not happen because models get smarter. It will happen because the engineering around the models gets better. Specifically, organizations will need:

Here is the twist. Those are exactly the traits that make a system worth interrupting a vacation for. Reliability is urgency. Once AI becomes genuinely dependable and genuinely hard to replace, it stops being a productivity story and becomes an operations story.

Agentic AI, systems that take actions instead of just suggesting them, is the clearest test case. The moment software can move money, change records, or talk to customers without a human checking first, the stakes change completely. The safety net disappears. And the moment the safety net disappears, so does the reason AI never passed the wake-the-CEO test.

Expect the market to split in two. On one side, cheap, invisible AI that everyone uses without thinking about it, the way we use electricity or search. On the other, a small number of mission-critical AI systems that get the full treatment: monitoring, audits, on-call rotations, executive attention. The interesting question for the next few years is not "how many companies use AI." It is "how many AI systems would someone get out of bed for."

What Businesses Should Do Right Now

The gap between reporting results and earning urgency is not a failure. It is a roadmap. Here is how to use it.

Tier your AI portfolio

Stop treating every AI project the same. Split them into three groups: systems the business cannot run without, systems that make people faster, and experiments that might become either. Each tier gets its own rules, its own budget, and its own level of scrutiny.

Ask the "what breaks" question

For every AI system, ask: if this stopped working at 2 a.m., what happens? If the answer is "someone would notice on Monday," it is a tier-two tool. If the answer is "we would lose money or break a promise," it is tier one, and it needs tier-one engineering.

Build the fallback before you scale

If you cannot safely switch a system off, you are not ready to depend on it. Design the human backup first, then grow the AI.

Put a name on everything

Unowned systems never become urgent. Someone has to be responsible, reachable, and empowered to act.

Translate results into money

Convert every claimed win into hours, dollars, or risk reduced. "It's working well" does not survive a budget meeting. "It saved this team 400 hours last quarter" does.

Watch the escalation metric

The percentage of leaders reporting AI results will keep climbing, it is a crowded, easy-to-hit number. The percentage who would interrupt a vacation is the harder, truer signal. Track it. When it starts to move, real change is underway.

What It Means for the Rest of Us

There is a calm, slightly uncomfortable piece of good news buried in this finding. If AI is not yet critical enough to ruin a CEO's holiday, we still have time. Time to build the guardrails, write the rules, and figure out who is accountable when an automated system gets something wrong.

That window will not stay open forever. The same qualities that make a system worth waking someone up for, dependence, speed, scale, no easy backup, are the qualities that make a failure hurt. The technology is arriving before the habits are. The two-thirds number says adoption is here. The vacation number says the discipline is not.

AI's future will not be decided by which models score highest on a benchmark. It will be decided by whether organizations can make these systems dependable enough to depend on. The companies that close that gap first will not just report results. They will be the ones whose phones ring at 2 a.m., and who are ready when they do.

TLDR: About two-thirds of IT leaders say AI is delivering real results, but almost none say those results are important enough to interrupt the CEO's vacation over. That gap is the real story: AI is working, but it is not yet mission-critical, because nearly every deployment still has a human safety net. Closing the gap means treating AI like operations, reliability, evaluation, rollback, and clear ownership, rather than a set of clever experiments. Watch the "would you wake the CEO" number, not the "are you using AI" number. That is the metric that will tell you when AI has truly arrived.