NextFin News - Kate Smaje of McKinsey says the debate over whether artificial-intelligence development is slowing is an important one - but it is not the debate happening in boardrooms. Executives who have already committed heavily to AI are now asking a different question: can these tools scale, deliver profitability, and remain sustainable? Her remarks, made in a televised interview on "The Close" on September 14, land against fresh survey data showing that corporate conviction in AI is still running well ahead of the financial returns companies can actually point to.
The Gap Between Conviction and the Bottom Line
The gap between belief and earnings has become the defining feature of enterprise AI in 2026. McKinsey's latest State of AI survey, published August 25 and based on responses from 1,719 professionals and business leaders worldwide, found that 37 percent of respondents attribute at least some earnings impact to their use of AI - about the same share as a year earlier. The proportion of "high performers" - organizations that attribute at least 5 percent of earnings before interest and taxes to AI and describe the impact as significant - has stayed flat at roughly 6 percent.
Yet investment keeps rising. Eight in ten respondents say AI has improved their individual productivity, half say it helps them make better decisions, and a majority still plan to increase AI spending. The report's own assessment captures the tension:
Organizations' conviction in AI is growing faster than the immediate financial returns they can attribute to it.
That is the tension Smaje's comments crystallize. While AI researchers and technology leaders argue publicly about the pace of progress toward artificial general intelligence, the risk of the technology, and the case for patience, corporate buyers have moved on to a harder set of questions. The technology is no longer the story. The economics are.
For investors, the stakes are concrete. Trillions of dollars of capital have flowed into the AI infrastructure build-out - chips, data centers, power, cooling - on the assumption that enterprise adoption will eventually convert into earnings. The survey's flat earnings-impact line is the first hard evidence that the conversion is taking longer than the capital cycle anticipated. It does not invalidate the build-out. But it does put a clock on it.
The ROI Gap Is an Implementation Lag, Not a Technology Failure
The flat earnings-impact number is the single most important figure in the survey, and it cuts both ways. To the skeptics, a steady 37 percent - and a high-performer share stuck at 6 percent - after years of deployment is evidence that AI delivers diffuse productivity gains that never aggregate to the income statement. To the believers, it is exactly what you would expect midway through a multi-year technology cycle: the tools arrive before the workflows that make them valuable.
History supports the second read, with a caveat. Michael Chui, a coauthor of the report and a senior fellow at McKinsey's QuantumBlack unit, has described the delay as a reflection of patterns seen with previous technologies:
It should not be surprising that it has taken time, because it is a reflection of trends we've seen with other technologies. History doesn't repeat itself, but it rhymes.
Enterprise software, cloud computing, and earlier waves of automation all followed the same arc - a long plateau between adoption and measurable return, broken only when companies redesigned processes rather than simply layered new tools onto old ones.
The mechanism matters. AI raises productivity at the individual level first - hence the 80 percent figure - because a worker with a capable model can complete a given task faster. That gain stays trapped at the individual level unless the organization changes who does what, how work is handed off, and how output is measured. The survey's own data points to where the value is hiding: respondents who redesign end-to-end workflows and reimagine entire domains such as marketing and operations report the greatest earnings impact. The difference between a pilot and a payoff is organizational, not computational.
This is a cyclical lag riding on top of a structural shift, and the distinction determines the investment conclusion. The structural part is real and durable: AI is changing how software is built, how knowledge work is allocated, and how firms compete. The cyclical part is the timing of value capture - and that tends to mean-revert, because returns catch up to deployment once the operating model catches up to the tool. If the ROI gap is structural, AI is a capital sink for most adopters. If it is cyclical, it is a deployment problem with a predictable resolution path. The evidence - the workflow redesign correlation, the historical analogy, the still-rising investment plans - points to cyclical.
Agentic AI Is Quietly Reshaping the Software Market
The most consequential number in the survey may not be the ROI figure at all. Forty percent of respondents at large organizations - those with more than $1 billion in annual revenue - report scaling AI agents, up from 27 percent a year earlier. At smaller organizations the share stayed flat at 22 percent, underscoring that the agentic wave is so far a large-enterprise phenomenon.
Within that, coding agents are moving fastest. About two in ten respondents overall are scaling them, rising to 31 percent at larger enterprises. And 32 percent of respondents say their organizations have decided against buying one or more software products or features because the functionality could be built in-house with agentic coding tools.
That is a second-order shock most of the market has not fully priced in. The AI conversation has been dominated by the demand side - who is buying chips, who is building data centers, how much capital is flowing into the infrastructure layer. The agentic coding trend points the other way: it is a substitution story for the enterprise-software layer. When a third of companies start building features they would have bought, the growth assumptions embedded in commercial software valuations come under pressure. This is not just a productivity gain inside the firm; it is a demand shift across the technology supply chain.
The build-versus-buy move also imposes a cost discipline that was largely absent in the first wave of generative AI enthusiasm. Building in-house forces a company to price the tokens, the engineering time, and the maintenance burden against a vendor quote. That accounting is exactly what Smaje's "sustainable" framing points to - and it is why the cost constraint in the survey deserves attention even though it affects a minority of respondents today.
The Cost Wall Is Coming Into View
About 20 percent of respondents say AI-related operating costs, including token costs, have constrained their use of the technology. A minority for now - but a leading indicator, not a lagging one. Costs constrain usage before they show up in earnings, and usage constraints are how ROI dreams quietly die.
The pattern is familiar from earlier technology cycles. The first wave of adoption is budget-inelastic because nobody knows what the steady-state cost looks like. The second wave is when finance gets involved, unit economics get modeled, and projects that looked free at pilot scale get repriced at production scale. McKinsey's finance leadership has signaled that the year-over-year token-cost increase seen in 2026 is not affordable again into 2027 - the moment the industry crosses from experimentation into budgeting.
This is where the sustainability question in Smaje's boardroom framing bites hardest. Scaling is not just a technical challenge of model reliability and latency; it is a financial challenge of whether the marginal value of an AI-driven action exceeds its marginal cost. For high-volume, low-margin use cases, that equation is unforgiving. It is also why the high performers - the 6 percent - matter more than the average: they are the proof that the unit economics can work when the use case, the workflow, and the cost control line up.
Labor Expectations Are Running Ahead of the Evidence
One more survey figure deserves scrutiny. Thirty-nine percent of respondents expect AI-related declines in their organization's total employment in the coming year, up from 32 percent in the 2025 survey. Only 43 percent expect no change.
The report itself flags the gap between expectation and reality: workforce reductions in 2025 "fell well short of what respondents in last year's survey had anticipated." In other words, companies have been expecting AI-driven job cuts for two consecutive years, and the cuts have not arrived at the scale forecast.
That matters for the profitability question. If AI were delivering labor substitution at the pace the market expects, earnings impact would be easier to find - headcount is one of the largest and most measurable cost lines on an income statement. The fact that expectations of employment decline are rising while reported earnings impact stays flat suggests companies are still in the augmentation phase: workers are more productive, but the workers are still there. The earnings benefit of that configuration is real but modest - better output per dollar of payroll, not a smaller payroll.
The Strongest Case Against the "On the Road to ROI" Narrative
The bear case is not that AI does not work. It is that it works in a way that does not translate into shareholder value for the companies paying for it. Productivity gains at the individual level may be captured by workers as reduced effort rather than by firms as reduced cost or higher output. The 6 percent high-performer share has not budged in a year, even as spending has risen sharply - a pattern consistent with diminishing returns at the margin. And the build-versus-buy shift, while real, may simply move spending from software licenses to engineering headcount and cloud bills, with no net margin improvement.
This counter-thesis is backed by the survey's own central finding: conviction growing faster than attributable returns. It is also consistent with a broader body of research showing that technology diffusion often enriches the suppliers of the technology more than the adopters. The chipmakers, the cloud providers, and the model developers are capturing a large share of the economic surplus today; there is no guarantee the adopters will catch up.
The answer to the bear case is the workflow evidence and the time-horizon split. Individual productivity gains are the first derivative; earnings impact is the second, and it arrives only after reorganization. But the burden of proof is on the believers, and the proof is not yet in the data.
The falsifying signal is concrete: if the share of respondents attributing at least 5 percent of EBIT to AI does not rise materially above the current 6 percent in the 2027 survey - say, to 10 percent or more - then the "on the road to ROI" framing is wrong, and the bear case that AI is a supplier-enriching capital sink becomes the base case. A secondary signal: if the 20 percent of respondents citing cost constraints rises above one-third in the next cycle, the scaling story stalls on economics before it stalls on technology.
What to Watch and Where the Value May Move
Short term, watch the cost line. Token prices, inference costs, and the share of firms reporting budget constraints will determine whether deployment continues at the current pace. A rise in cost-constrained respondents above one-third would be the first clear sign of an investment pause.
Medium term, watch the high-performer share and the workflow metric. The 2027 survey should show whether companies that redesigned workflows have pulled ahead of those that merely deployed tools - and whether the 6 percent becomes 10 percent or stays stuck. This is the single most important read on whether AI is a general-purpose technology in the economic sense or a productivity tool with a ceiling.
Long term, watch the build-versus-buy shift. If the 32 percent of firms building instead of buying grows toward half, the enterprise-software market faces a structural demand shock that would reverberate through software valuations, cloud growth, and the economics of the AI infrastructure build-out itself.
The base case is that AI value realization follows the classic technology-diffusion arc: a long plateau, then a step-change as workflows redesign around the tool rather than the other way around. The upside case is that agentic coding and autonomous workflows compress that timeline, delivering earnings impact faster than the historical analogy suggests. The downside case is that the cost wall and the flat high-performer share mark the top of the cycle, and the 37 percent EBIT-impact figure proves to be the ceiling rather than the floor.
For investors, the implication is asymmetrical. The infrastructure layer - chips, data centers, power - has already been repriced on the conviction story. The next leg of returns, if it comes, belongs to the companies that reorganize around AI, not the ones that merely buy it. And the software layer faces a demand question that most valuation models have not yet confronted.
The AI debate has split into two conversations - one about how smart the machines are getting, and another about whether they pay for themselves. For the next leg of the market, only the second one matters.
Explore more exclusive insights at nextfin.ai.
