Is AI Overhyped? What Leaders Should Actually Believe in 2026
The question I get asked more than any other, on more stages and in more boardrooms than I can count, is some version of this: "Shawn, be honest with us. Is this real, or is it the biggest bubble of our careers?"
It is a fair question. It is also the wrong one.
Because "Is AI overhyped?" bundles three completely different questions into one, and the answer to each is different. Is the technology overhyped? Is the capital overhyped? Is your company's AI programme overhyped? A leader can be right about one and catastrophically wrong about another. The executives I watch make the worst decisions in 2026 are not the ones who believe too much or too little. They are the ones who have not separated these questions at all.
Here is what the evidence actually says, and what I think you should believe when you walk back into your Monday morning meeting.
The short answer
AI is overhyped as a financial story in the short term and underhyped as an operating story in the medium term. The capability is real and improving. The capital deployed against it has almost certainly run ahead of near-term returns. And the overwhelming majority of corporate AI programmes are producing far less value than their sponsors claim, not because the technology failed, but because the organisation never changed.
Both things are true at once. That is the whole point.
The three questions hiding inside "Is AI overhyped?"
Before you can answer, separate these:
Capability hype: Can the technology do what people claim it can do?
Capital hype: Is the money being spent to build it justified by the returns it will generate?
Corporate hype: Is your organisation's AI activity creating measurable value, or performing the appearance of progress?
Confusing these is how a CEO ends up cancelling a genuinely productive AI programme because they read a headline about a market correction. It is also how a board approves a ninth pilot because the technology is "clearly transformative," without ever asking why the first eight produced nothing.
Let me take them in turn.
Question 1: Is the capability overhyped?
Mostly, no. The capability is the strongest part of the story.
Adoption at the top of the funnel is genuinely extraordinary. McKinsey's State of AI survey, drawing on nearly 2,000 respondents across 105 countries, found that 88% of organisations now use AI regularly in at least one function. That is not a niche technology. That is closer to email penetration than to blockchain penetration.
And where AI is deployed properly against a bounded task, the productivity evidence is solid. A Stanford HAI meta-analysis published in April 2026, pooling results across multiple published studies rather than relying on a single experiment, found meaningful gains: roughly 14 to 15% in customer support, 26% in software development, and up to 50% in marketing output.
Those are not vendor claims. Those are measured effects across controlled studies.
*Where capability genuinely is oversold:*
Autonomy. The gap between "an agent completed this task in a demo" and "an agent runs this workflow unsupervised in production, in a regulated industry, at scale" remains enormous.
Reliability at the tail. Systems that are excellent on the typical case are still unpredictable on the unusual one, and the unusual case is where most enterprise risk lives.
Timeline compression. Almost every "within 12 months" claim made about enterprise AI since 2023 has taken longer than promised.
So: capability is real, and the honest caveat is that reliable autonomy is harder than the demos suggest. If you want the plain-English version of what agents can and cannot do, I have written a full breakdown in Agentic AI Explained for Business Leaders.
Question 2: Is the capital overhyped?
This is where the "bubble" argument has the most force, and where I think leaders should be genuinely cautious.
The scale of the infrastructure buildout in 2026 is without modern precedent. Gartner's forecast puts global AI spending at roughly $2.5 trillion in 2026, up around 44% year over year. Estimates for combined capital expenditure by the five largest hyperscalers cluster in the $675 to 800 billion range for 2026, up from roughly $261 billion in 2024, a near-tripling in two years, with projections above $1 trillion for 2027.
To put that in perspective: analysis published by legal and economic researchers in 2026 noted that hyperscaler capex is on a path to consume a larger share of US GDP than peak spending on the Apollo programme, the interstate highway system, or the dot-com-era broadband buildout.
Several structural concerns deserve to be taken seriously rather than dismissed:
Circular financing. Capital flows into AI companies, is spent immediately on compute from cloud providers, and is booked as revenue by those providers, which supports valuations that justify further investment. Money moving in a loop can look like growth from the outside.
Off-balance-sheet commitments. Moody's reported in early 2026 that hyperscalers hold roughly $662 billion in data-centre lease commitments that have been signed but not yet commenced, obligations that sit outside the capex figures most analysts scrutinise.
Cash-flow strain. Free cash flow across the largest spenders has compressed sharply, debt issuance has climbed, and several firms have paused or reduced share buybacks to fund the buildout.
Depreciation assumptions. Hardware is being depreciated over five to six years while its useful economic life may be considerably shorter, which flatters reported earnings.
The Bank for International Settlements has flagged rising leverage among AI firms, a growing footprint in credit markets, and the opacity of private financing arrangements as genuine financial-stability concerns.
My honest read: an infrastructure overbuild is more likely than not. That is what happened with railways, with electricity, and with fibre. It is also worth remembering what happened after each of those overbuilds: the capacity did not disappear. It became cheap, and the second wave of companies built on top of it profitably. A capital correction is not a technology refutation. Confusing the two is the single most expensive mistake a leader can make in the next 24 months.
Question 3: Is your company's AI programme overhyped?
Almost certainly yes. And this is the question you can actually do something about.
Three findings, from three independent research programmes, tell you everything you need to know:
Finding
Source
Only 39% of organisations report any enterprise-level EBIT impact from AI, and only around 6% qualify as high performers
McKinsey State of AI
56% of CEOs report AI has produced neither revenue growth nor cost savings to date
PwC 29th Global CEO Survey, roughly 4,500 CEOs
More than 80% of AI projects fail to deliver intended business value
RAND Corporation
Here is what I take from that, and it is not the obvious reading.
The obvious reading is that AI does not work. It is wrong. Look at who is being surveyed. These are CEOs and enterprise leaders reporting on portfolios, not engineers reporting on tools. What is failing is not the technology. It is the translation layer between a capability and a business outcome, and that layer is entirely made of things leaders control: workflow design, measurement, sponsorship, and the willingness to stop doing things that are not working.
The supporting evidence points the same way. A separate global survey found 42% of companies abandoned most of their AI initiatives in 2025, up sharply from 17% a year earlier. The widely quoted claim that around 95% of generative AI pilots produced no measurable profit impact has been contested on methodology, and I would not lean on it alone. But the direction is corroborated from too many independent angles to dismiss.
The single most telling number I have seen is this one. Across a global survey of 639 senior enterprise AI leaders, the share of organisations whose AI returns had failed to outpace their investment sat at 57%, unchanged across two consecutive years, even as production capabilities improved substantially.
Read that again, because it is the whole argument. Capability went up. Value did not follow. The gap between them is a leadership problem wearing a technology costume.
Why the value does not follow
Across the research, the failure patterns are remarkably consistent, and almost none of them are technical:
No agreed definition of success. Roughly 73% of failed AI projects had no shared definition of success before launch. Projects with quantified success metrics defined upfront succeed at around 54%; those without, around 12%.
No one checks afterwards. A significant share of enterprise AI projects were approved on projected ROI that was never measured post-launch.
The workflow was never redesigned. Organisations reporting significant financial returns are roughly twice as likely to have redesigned the workflow before selecting the tool. Most do the reverse.
The data foundation was not ready. Gartner has forecast that 60% of AI projects unsupported by AI-ready data would be abandoned through 2026.
Expecting too much, too fast. In a Gartner survey of 782 infrastructure and operations leaders, 57% of those who experienced AI failure attributed it to unrealistic speed expectations.
No discipline to kill projects. If your organisation has cancelled zero AI initiatives in the last twelve months, you do not have a portfolio. You have a collection.
Executive sponsorship fades. Enthusiasm at kickoff, absence at the six-month mark.
I have written at length about the risks leaders systematically under-weight in What Are the Real Risks of AI Adoption That Most Organizations Ignore, and about why so much of the training spend evaporates in Why Most Corporate AI Training Programmes Are a Waste of Money.
The scaling ceiling nobody talks about
Here is the statistic that should reframe your entire 2026 planning cycle.
The same McKinsey research shows that while 62% of organisations have begun working with AI agents in some form, only 23% have scaled agents in at least one function, and in no single business function do more than 10% of organisations get past the experimentation stage. Software engineering and IT are furthest along. Everything else trails badly.
Meanwhile, Gartner projects that more than 40% of agentic AI projects will be cancelled by the end of 2027 on grounds of escalating cost, unclear value, and inadequate risk controls.
So when someone tells you that "everyone is running autonomous agents now," they are describing a demo economy, not an operating one. The bottleneck is not model quality. It is procurement cycles, legal review, data residency, security governance, and the simple fact that redesigning how humans work is slower than deploying software.
That is not a reason to slow down. It is a reason to understand what you are actually racing against, and it is precisely why I argue that leadership-led AI strategy outperforms bottom-up experimentation.
What leaders should actually believe in 2026
Seven positions I would defend in front of any board:
1. Believe the capability, discount the timeline. Assume the technology will do what is claimed. Assume it will take twice as long to work reliably inside your organisation as the vendor implies.
2. Separate the market story from your operating story. A correction in AI equities tells you almost nothing about whether AI should be handling your claims triage. Do not let a headline set your operating plan.
3. Treat AI as an organisational-design problem, not a procurement problem. The evidence is overwhelming: the winners redesigned the work first. The losers bought the tool first.
4. Measure like a sceptic, invest like an optimist. Baseline before you deploy. Use holdout groups. Insist on total cost transparency, including the human review time nobody puts in the business case.
5. Build a portfolio, and kill things. If nothing has been cancelled, nothing is being evaluated. Killing a failing pilot is a sign of governance maturity, not failure.
6. Concentrate rather than sprinkle. The organisations capturing value are not the ones with the most pilots. They are the ones that took two or three workflows and rebuilt them completely.
7. Assume the capacity outlives the correction. Even if the capital cycle breaks, the compute, the models, and the talent do not evaporate. They get cheaper. Plan for the world after the correction, not just through it.
A five-question diagnostic for your own programme
Ask these in your next executive meeting. The silences will tell you more than the answers:
What percentage of our AI portfolio has crossed from pilot into scaled production with documented business outcomes?
What did we baseline before we deployed, and against what control?
How many AI initiatives have we killed in the last twelve months?
Which workflows have we genuinely redesigned, as opposed to layered a tool onto?
If our AI budget were cut by half tomorrow, which initiatives would we protect, and can we articulate why in financial terms?
If your team cannot answer these crisply, your AI programme is currently a communications exercise. That is fixable. But not by buying more software.
So, is AI overhyped?
Here is where I land, and it is the same thing I say from the stage whether I am in front of a Fortune 500 executive team, a government department, or an association audience.
The hype is real. So is technology. The mistake is treating them as the same question.
The companies that will look brilliant in 2030 are not the ones that predicted the market correctly. They are the ones that used this period, while everyone else was arguing about bubbles, to do the unglamorous work: rebuilding workflows, fixing data, training people, killing bad projects, and concentrating their bets. That work is boring. It is also the entire game.
Disruption almost never announces itself as a serious threat. It shows up looking underwhelming, overpriced, and slightly ridiculous, which is exactly why disruption always starts as a joke. The leaders who get this right are not the biggest believers or the loudest sceptics. They are the ones bold enough to act on an uncertain thesis while remaining honest enough to measure whether it is working.
That combination, boldness paired with intellectual honesty, is the through-line of everything I write about in The Bold Ones, and it is the core of the keynote I deliver on Innovation in a World of AI.
The bubble question is a spectator's question. The operator's question is better: what are we going to rebuild while everyone else is watching the market?
Frequently Asked Questions
Is AI overhyped in 2026?
Partly. The technology's capability is real and measurably improves productivity in bounded tasks. Studies show gains of roughly 14 to 26% in support and software engineering. What is overhyped is the near-term financial return: independent research consistently finds that the large majority of enterprise AI pilots deliver no measurable profit-and-loss impact, and only around 6% of organisations qualify as genuine AI high performers.
Is there an AI bubble, and will it burst?
There are legitimate structural concerns: hyperscaler capital expenditure has roughly tripled since 2024, roughly $662 billion in data-centre leases sit off balance sheet, free cash flow at the largest spenders has compressed, and international financial bodies have flagged rising leverage and circular financing. An infrastructure overbuild is plausible. However, historical precedent from railways, electricity, and fibre suggests that a capital correction does not erase the underlying capacity: it makes it cheaper for the next wave of builders.
What percentage of AI projects actually fail?
Estimates vary by definition and methodology. Widely cited research puts the figure at around 95% of generative AI pilots showing no measurable P&L impact, while RAND has found that more than 80% of AI projects fail to deliver intended business value. S&P Global reported that 42% of companies abandoned most of their AI initiatives in 2025, up sharply from 17% the previous year.
Why do most enterprise AI projects fail?
Almost never for technical reasons. The dominant causes are: no agreed definition of success before launch, no measurement afterwards, deploying tools without redesigning the underlying workflow, weak or fragmented data foundations, unrealistic speed expectations, no discipline to cancel failing initiatives, and executive sponsorship that fades after kickoff.
What is the difference between AI adoption and AI value?
Adoption measures usage; value measures financial impact. In 2026 these two curves are running at very different speeds: around 88% of organisations use AI regularly, but only 39% report any enterprise-level EBIT impact. High adoption with low value is the defining pattern of the current cycle.
Should we pause AI investment until the market settles?
Pausing entirely is usually the wrong response, because the constraint on value creation is organisational learning, not technology access, and learning cannot be bought back later. A better approach is to concentrate: reduce the number of initiatives, deepen the ones with a clear baseline and owner, and redirect spend from tool licences toward workflow redesign and capability building.
How do I know if my company's AI programme is real or performative?
Ask what percentage of the portfolio has reached scaled production with documented outcomes, what was baselined before deployment, how many initiatives have been cancelled in the past year, and which workflows have genuinely been redesigned. If the answer to "how many have we killed" is zero, no real evaluation is happening.
What should leaders prioritise for AI in 2026?
Workflow redesign over tool procurement, portfolio concentration over pilot proliferation, measurement discipline over anecdote, and capability building over training-day theatre. Organisations that redesign work before selecting tools are roughly twice as likely to report significant financial returns.