Every few weeks, a new model drops. GPT-5. Claude Opus 4.5. Gemini 3. Llama 4. Grok. The benchmarks improve. The context windows expand. The demos look more impressive. The tweets get more breathless.
And then you talk to a normal person and they say something like, "Yeah, I tried ChatGPT once. It was cool."
There's a gap forming between what AI can do and what people are actually doing with it. That gap is the most interesting thing happening in technology right now. Not because it means AI is overhyped. But because understanding why the gap exists tells you more about the future than any model release ever will.
The Numbers Look Good Until You Look Closer
ChatGPT has 900 million weekly active users and over 1 billion monthly, with 50 million paying subscribers. Claude has over 300,000 business customers processing 25 billion API calls monthly. GitHub Copilot has 4.7 million paid subscribers and is used by 90% of Fortune 100 companies. 65% of organizations now use generative AI in at least one business function, double from ten months prior.
Those are real numbers. They sound like a technology that has arrived.
But look underneath: 54% of organizations are actively deploying AI agents, up from just 12% in 2024, yet only 5% achieve substantial ROI at scale. And here's the one that should stop you: only 7% of organizations that adopted AI achieved strategic integration. That means 93% adopted AI and then... did something with it that didn't fundamentally change anything.
The adoption is wide. But it's about an inch deep.
The Pilot Graveyard
There's a pattern playing out across industries right now. Companies buy AI tools, run pilots, generate internal excitement, and then nothing happens.
42% of companies abandoned most of their AI initiatives in 2025. That's up from 17% the year before. It more than doubled.
A financial services company deployed AI customer service chatbots without training their call center staff. A large portion of employees opposed the system. Adoption stalled, the system sat unused.
A logistics company purchased an AI-powered routing system but didn't integrate it with their existing warehouse software. Project abandoned after an expensive implementation.
Taco Bell rolled out an AI ordering system that became a punchline. One customer accidentally ordered 18,000 cups of water. The AI kept asking customers to add more drinks even after they declined. Rather than speeding up service, the system required constant human monitoring. They ended up switching to a hybrid approach.
These aren't edge cases. They're the norm. MIT reported that 95% of generative AI pilots at companies are failing. Not because the technology doesn't work. Because organizations don't know how to absorb it.
$700 Billion and Counting
Big Tech AI capex is now tracking closer to $700 billion in 2026 — Amazon alone at $200 billion, Alphabet $175-185 billion, Microsoft $145 billion, Meta $115-135 billion. Global AI spending is forecast to surpass $300 billion this year. AI is now 6% of the global SaaS market and growing faster than any software category in history.
So what are we getting for all that money?
According to RAND and MIT, 80.3% of enterprise AI projects fail: 33.8% abandoned outright, 28.4% deliver no value, 18.1% unable to justify costs. Of $684 billion invested in AI in 2025, an estimated $547 billion failed to deliver intended value. Only 5% of companies achieve substantial ROI at scale.
The average abandoned AI project costs $4.2 million. Failed projects that limped to completion cost $6.8 million while returning only $1.9 million. Why? 73% lack clear executive alignment. 68% underinvest in data governance. 61% treat AI as an IT project rather than business transformation.
61% of CEOs say they're under increasing pressure to show returns on AI investments. Big Tech free cash flow could drop up to 90% in 2026 as capital expenditure outpaces revenue growth.
This isn't a bubble thesis. The technology is real. But the gap between what's being spent and what's being earned is wider than it's ever been. And 2026 is the year that gap becomes impossible to ignore.
Where It's Actually Working
It's not all bad. Some use cases are delivering genuine value, and they tend to share a pattern: they're specific, measurable, and integrated into existing workflows rather than replacing them.
Software development is the clearest win. GitHub Copilot now has 4.7 million paid subscribers, with 46% of code written by the average user being AI-generated and 88% of that code staying in the final version. Developers complete tasks 55% faster. Pull request time dropped from 9.6 days to 2.4. This makes sense because code is structured, testable, and the feedback loop is immediate. You know right away if the AI-generated code works.
Call centers are another bright spot. A large-scale study showed 14-15% productivity gains measured by cases resolved per hour. The AI handles routine inquiries, routes complex ones to humans, and provides real-time suggestions during calls. It's augmentation, not replacement, and it works.
Financial services firms are seeing 40-50% time savings on document review and due diligence. Research synthesis is 25-40% faster. Compliance drafting has accelerated significantly.
And when companies commit to serious workflow redesign, not just bolting AI onto existing processes, productivity gains of 35-55% have been reported. But that keyword "serious" is doing a lot of work. Most companies aren't doing serious redesign. They're adding a chatbot and calling it transformation.
One area where serious integration is happening fast: agentic commerce. AI agents that don't just recommend products but negotiate, purchase, and manage transactions autonomously. Google launched its Universal Commerce Protocol (UCP) in January 2026 to enable autonomous shopping across platforms. Visa and Mastercard announced authentication standards for autonomous AI transactions. Shopify launched agentic storefronts. Stripe built an entire commerce suite around it. Morgan Stanley projects nearly 50% of online shoppers will use AI agents by 2030, representing about 25% of ecommerce spending. This is one we'll be writing about in more depth soon — the infrastructure layer being built underneath it is worth its own piece.
The Productivity Paradox Nobody Wants to Talk About
UC Berkeley published research in February 2026 that hit a nerve. They found that as AI tools improved employee productivity, workers didn't work less. They took on more work. The time savings that AI was supposed to create got absorbed immediately by increased workload.
The researchers put it simply: "You had thought that maybe because you could be more productive with AI, then you save some time, you can work less. But then really, you don't work less."
On top of that, many organizations are dealing with what's been called "workslop," low-quality AI output that employees spend hours fixing. This hidden cost rarely shows up in ROI calculations. The AI generates a draft in seconds, and then a human spends 45 minutes making it not sound like it was written by a machine. That's not a productivity gain. That's a different kind of work.
The Trust Problem
You'd expect that as people use AI more, they'd trust it more. The opposite is happening.
Regular AI usage among workers jumped 13% in 2025. Confidence in AI use plummeted by 18% in the same period. The people using AI the most are becoming less confident in it, not more.
And there's a massive gap between how executives and the general public feel about this technology. 93% of corporate leaders believe AI will have a net positive impact in the next five years. Only 58% of the general public agrees. 50% of Americans say they're more concerned than excited about increased AI use, up from 37% in 2021.
Among workers using AI tools, only 22% feel comfortable and 15% feel excited. Meanwhile, 21% feel overwhelmed and 23% feel stressed. 52% cite lack of knowledge as the biggest barrier.
This is a technology that's being pushed on people faster than they can learn to use it. And the result isn't adoption. It's anxiety.
The Jobs Question
Nobody has a clean answer for this one yet.
40% of jobs globally face meaningful AI exposure, rising to 60% in high-income countries. About 78,000 tech job losses were attributed to AI in the first half of 2025 alone. Software developer employment for ages 22-25 declined 20% from its late 2022 peak. 37% of business leaders now anticipate replacing human workers by end of 2026, and employee concerns about job displacement jumped from 28% to 40% in just two years.
But zoom in and the picture shifts. Occupations with higher AI exposure experienced larger unemployment rate increases between 2022 and 2025. Computer and mathematical occupations, the most AI-exposed category, saw some of the steepest rises. The displacement isn't happening evenly. It's concentrated in specific roles, and it's accelerating.
The long-term projections say AI could displace 92 million jobs globally while creating 170 million new ones, a net positive of 78 million. But "net positive at the macroeconomic level" doesn't describe how smoothly an individual worker transitions from a declining role to a new one. 3.9% of US workers, roughly 5 to 6 million people, face high AI exposure with low adaptive capacity. The gap between the macro data and the lived experience is where the real tension sits.
One number worth sitting with: workers with AI skills command a 56% wage premium over comparable roles. The path isn't clear, but the direction is. Learning to work with AI is becoming job insurance.
So What's Actually Happening?
Somewhere between the hype and the doom, there's a more honest picture.
The models are genuinely better every quarter. That's not in question. But the gap between what AI can do in a demo and what it can do inside a real organization, with real people and real constraints, is enormous.
The bottleneck isn't the technology. It's us. 38% of knowledge workers now use generative AI daily, up from 11% in 2024. But 52% of organizations still say they lack the skills. 46% of tech leaders cite the skills gap as their number one obstacle, above technology or budget. Companies that invest in workforce capabilities see productivity gains. Companies that don't see nothing.
Nearly 7 in 10 organizations have adopted generative AI. Almost 5 in 10 have moved to agentic AI. But only 5% achieve substantial ROI at scale. That's the gap. And it's not closing quickly because it requires something harder than buying software. It requires changing how organizations think and work. That has never been fast.
The models will keep improving. That's the easy part. The hard part is everything else: training, trust, workflow redesign, organizational culture, and the willingness to actually do things differently rather than bolting a chatbot onto the same old processes.
AI isn't overhyped. But it is underabsorbed. The technology is ready. Whether people are is a different question entirely.
Subscribe on Substack for future research