Article content
The Year of Reckoning: Why 2026 Will Be Critical for Enterprise AI
42% of companies have already abandoned most of their AI projects. The data suggests the worst may not be over.
TL;DR: 42% of companies abandoned AI projects in 2025, double the previous year. Stanford HAI predicts 2026 will be the year of reckoning: less hype, more demand for concrete proof. Brynjolfsson’s employment data shows the impact already: -20% for junior developers, +8% for senior. For investors, the implications are clear: metrics defined before launch, not after; vendor solutions (67% success rate) vs internal development (33%); attention to go-live timelines, which kill projects more than technology.
In mid-December 2025, nine faculty members from Stanford Human-Centered Artificial Intelligence published their predictions for 2026. This is not the usual academic futurology exercise, but a collective statement with a clear message: the party is over.
James Landay, co-director of HAI, opens with a phrase that sounds almost provocative in an era of triumphalist announcements: “There will be no AGI this year.” The point, though, is what he adds immediately after: companies will begin publicly admitting that AI has not yet delivered the promised productivity increases, except in specific niches like programming and call centers. And we’ll finally hear about failed projects.
This is not a prediction about the future. It’s a snapshot of something already happening.
The Numbers No One Wants to Look At
In July 2025, the MIT Project NANDA published a report that generated considerable debate for a single statistic: 95% of enterprise AI projects generate no measurable return. The number has been contested, the methodology has its limitations, the definition of “success” is debatable. But it’s not an isolated data point.
During the same period, S&P Global found that 42% of companies abandoned most of their AI initiatives in 2025. In 2024, the percentage was 17%. The abandonment rate has more than doubled in a year. On average, the surveyed organizations threw out 46% of proof-of-concepts before they reached production.
According to the RAND Corporation, over 80% of AI projects fail, double the failure rate of traditional IT projects. Gartner reports that only 48% of AI projects reach production, and over 30% of GenAI projects will be abandoned after the proof of concept by end of 2025.
The causes are always the same: insufficient data quality (43% according to Informatica), lack of technical maturity (43%), skills shortage (35%). But beneath these numbers lies a deeper pattern. Companies are discovering that AI works in demos but not in production, generates enthusiasm in pilots but not ROI in balance sheets.
It’s these numbers that explain why Stanford HAI, an institution hardly known for technological pessimism, is shifting the conversation. No longer “can AI do this?” but “how well, at what cost, for whom?”.
Canaries in the Coal Mine
If failure rates are the symptom, Erik Brynjolfsson’s work offers a more precise diagnosis. “Canaries in the Coal Mine”, published in August 2025 by Stanford’s Digital Economy Lab, is among the most rigorous studies currently available on AI’s impact on the job market.
The paper uses ADP payroll data, the largest payroll service provider in the United States, covering over 25 million workers. The goal is to track employment changes in AI-exposed professions.
What emerges is clear. Employment for software developers ages 22-25 has declined 20% from the peak of late 2022, roughly since the launch of ChatGPT, through July 2025. This is not an isolated data point: early-career workers in the most AI-exposed occupations show a relative decline of 13% compared to colleagues in less exposed roles.
The most interesting finding, though, is the age divergence. While young workers lose ground, workers over 30 in the same high-exposure categories have seen employment growth between 6% and 12%. Brynjolfsson puts it this way: “It appears that what young workers know overlaps with what LLMs can replace.”
It’s not a uniform effect, but a realignment: AI is eroding entry-level positions faster than it creates new roles. The “canaries in the coal mine”—young developers and customer support staff—are already showing symptoms of a larger change.
When Brynjolfsson predicts the emergence of “AI economic dashboards” that track these shifts in near-real-time, he’s not speculating. He’s describing the infrastructure needed to understand what’s happening, infrastructure that doesn’t exist today but could become urgent in 2026.
The Divergence Between Adoption and Results
There’s a paradox in 2025 data that deserves attention. AI adoption is accelerating: according to McKinsey, the percentage of companies claiming to use AI rose from 55% in 2023 to 78% in 2024. Use of GenAI in at least one business function more than doubled, from 33% to 71%.
Yet, in parallel, project abandonment rates are growing instead of declining. S&P Global shows a jump from 17% to 42% in a single year. The MIT NANDA report speaks of a “GenAI Divide”, a clear division between the 5% extracting real value and the 95% that remain stalled.
Many companies have gone through the phases of enthusiasm, pilots, impressive demos, and then crashed against the wall of real production. They discovered that the model works in a sandbox but not with their data; that integration into existing workflows is more complex than expected; that the ROI promised by vendors doesn’t materialize.
Angèle Christin, a communication sociologist and HAI senior fellow, puts it plainly: “San Francisco billboards saying ‘AI everywhere!!! For everything!!! All the time!!!’ betray a slightly manic tone.” Her prediction: we’ll see more realism about what we can expect from AI. Not necessarily the bubble bursting, but the bubble might stop inflating.
The Measurement Problem
One of the most concrete, and potentially most significant, predictions comes again from Brynjolfsson. He proposes the emergence of high-frequency “AI economic dashboards”: tools that track, at the task and employment level, where AI is increasing productivity, where it’s displacing workers, where it’s creating new roles.
Today we have nothing like that. Labor market data arrives months late. Companies measure AI adoption but rarely its impact. Industry reports capture hype but not results.
If these dashboards do emerge in 2026, they’ll change how we talk about AI. The debate will shift from the generic “does AI have an impact?” to more precise questions: how fast is this impact spreading, who’s being left behind, which complementary investments work.
It’s an optimistic vision: better data leads to better decisions. But it’s also an implicit admission: today we’re navigating blind.
Healthcare and Legal: The Test Sectors
Two sectors emerge from Stanford predictions as particularly relevant testbeds.
Nigam Shah, Chief Data Scientist at Stanford Health Care, describes a problem that anyone in the sector will recognize. Hospitals are flooded with startups wanting to sell AI solutions. “Every single proposal can be reasonable, but in aggregate they’re a tsunami of noise.”
According to Shah, 2026 will see systematic frameworks emerge for evaluating these solutions: technical impact, the population the model was trained on, ROI on hospital workflow, patient satisfaction, quality of clinical decisions. This is work Stanford is already doing internally, but it will need to extend to institutions with fewer technical resources.
Shah also signals a risk. Vendors, frustrated by hospitals’ long decision cycles, might start going directly to end users. “Free” applications for doctors and patients that bypass institutional controls. This is already happening: OpenEvidence for literature summaries, AtroposHealth for on-demand answers to clinical questions.
In the legal sector, Julian Nyarko predicts a similar shift. The focus will move from “does this model know how to write?” to more operational questions: accuracy, citation integrity, exposure to privilege violations. The sector is already working on specific benchmarks, like those based on “LLM-as-judge”, frameworks where one model evaluates another model’s output for complex tasks like multi-document summarization.
Healthcare and legal share a characteristic: they’re highly regulated, with severe consequences for error. If AI must prove its value anywhere, it’s where the test will be hardest. And most significant.
Track Record: How Reliable Are These Predictions?
Stanford HAI publishes annual predictions going back several years. It’s worth asking how accurate they’ve been.
At the end of 2022, Russ Altman predicted for 2023 a “shocking rollout of AI way before it’s mature or ready to go”. It’s hard to find a more accurate description of what happened: ChatGPT, Bing Chat, Bard launched in rapid succession, with accuracy problems, hallucinations, embarrassing incidents. Altman had also predicted a “hit parade of AI that’s not ready for prime time but launches because driven by an industry too zealous.” Exactly right.
Percy Liang, also at the end of 2022, predicted that video would be a focus of 2023 and that “we might reach the point where we can’t tell if a human or computer generated a video”. He was a year early (Sora arrived in February 2024) but the direction was correct.
For 2024, Altman predicted a “rise of agents” and steps toward multimedia systems. Both came true, though agents are still more promise than production reality.
Not all predictions came true. Expectations of U.S. Congressional action were disappointed: Biden’s Executive Order happened, but the new administration changed direction. Overall, though, Stanford HAI’s track record is reasonable: they tend to be cautious rather than enthusiastic, and technical predictions are generally well-founded.
This doesn’t guarantee that 2026 predictions will come true. But it means they’re worth taking seriously.
What It Means for Decision-Makers
If Stanford predictions and failure rate data converge on anything, it’s this: 2026 will be the year when enterprise AI must show results, not demos.
For those managing tech budgets, the implications are concrete.
On the metrics front, AI projects must have success criteria defined before launch, not after. Not “let’s explore AI for customer service” but “reduce average ticket resolution time by 15% within 6 months, with cost-per-interaction below X”. Projects without clear metrics have a disproportionate likelihood of ending up in the 42% of abandonments.
On the make-or-buy front, the MIT NANDA report indicates that solutions bought from specialized vendors have a 67% success rate, against 33% for internal development. This doesn’t mean internal development is always wrong, but it requires skills, data, and infrastructure that many organizations overestimate having.
On timing, mid-market enterprises move from pilot to production in about 90 days, according to the same report. Large enterprises take nine months or more. Bureaucracy kills AI projects more than technology does.
Finally, a matter of honesty. The shadow economy of AI (90% of employees use personal tools like ChatGPT for work, according to MIT NANDA) indicates that individuals already know where AI works better than official enterprise initiatives. Instead of fighting it, organizations could learn from this spontaneous adoption.
What’s Missing
Stanford predictions have clear blind spots.
None of the experts mention energy consumption and AI’s environmental impact. Christin hints at it (“tremendous environmental costs of the current build-out”) but the topic isn’t developed. Yet AI data centers are becoming one of the world’s biggest energy consumers, and this will eventually factor into ROI calculations.
There’s also a lack of serious discussion about market concentration. Frontier models are developed by a handful of companies. This creates dependencies, influences pricing, determines who can compete. It’s a strategic factor that anyone planning AI investments should consider.
Landay alludes to “AI sovereignty”, countries wanting independence from American providers, but the topic remains superficial. This is rapidly evolving, with significant geopolitical implications, that deserves deeper analysis.
A Shift in Tone
More than individual predictions, what strikes you about the Stanford article is the tone. There’s no industry-typical enthusiasm. No promises of imminent transformation. There’s caution, demand for proof, emphasis on measurement.
When the co-director of a Stanford AI institute opens by saying “there will be no AGI this year,” he’s taking a stand against a dominant narrative. When economists like Brynjolfsson publish data on young workers losing employment, they’re documenting costs, not just benefits.
This doesn’t mean AI is overvalued or that projects should stop. It means the phase of uncritical adoption is ending. Whoever continues to invest will need to do so with calibrated expectations, defined metrics, ability to admit failure when it occurs.
2026, if these predictions are correct, will be the year when we discover which AI projects were sound and which were built on hype. For many organizations it will be a painful discovery. For others, an opportunity: whoever has already learned to measure, iterate, and distinguish value from promise will have a competitive advantage that generic enthusiasm cannot buy.
Sources
Brynjolfsson, E., Chandar, B., & Chen, R. (2025). Canaries in the Coal Mine: Six Facts about the Recent Employment Effects of AI. Stanford Digital Economy Lab.
McKinsey & Company. (2024). The State of AI in 2024: Gen AI adoption spikes and starts to generate value. McKinsey Global Institute.
MIT Project NANDA. (2025). The GenAI Divide 2025. Massachusetts Institute of Technology.
RAND Corporation. (2024). The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed. RAND Research Reports.
S&P Global Market Intelligence. (2025, October). Generative AI Shows Rapid Growth but Yields Mixed Results. S&P Global.
Stanford HAI. (2025, December). Stanford AI Experts Predict What Will Happen in 2026. Stanford Human-Centered Artificial Intelligence.
