HiddenN de The Possimpible

Strategy

The state of enterprise AI in 2026: what's working, what isn't, and what the experts actually say

· 9 min read · Hidden Node

Here's the strange thing about AI in business right now. Almost everyone is using it, and almost nobody can show you the money.

Stanford's latest AI Index says organizational adoption has reached 88%. In the same breath, McKinsey's global survey found that only 39% of companies report any impact on EBIT at the enterprise level, and MIT's NANDA initiative reported that about 95% of generative AI pilots it studied produced no measurable impact on profit and loss.

So which is it: a revolution or an expensive hobby? Having read through the big studies of the last eighteen months, our answer is that both are true. It just depends on what a company actually does with the technology.

The gap between "using" and "getting value"

The numbers line up with almost suspicious consistency:

Notice what's not on that list: "the models weren't smart enough." MIT's researchers put it bluntly. The failures come from brittle workflows, weak contextual learning and misalignment with day-to-day operations, not from model quality.

What the winners do differently

McKinsey separated out a small group of "AI high performers", about 6% of respondents, who attribute at least 5% of EBIT to AI. The single habit that stood out: they were nearly three times as likely as everyone else to fundamentally redesign their workflows when they deployed AI.

That matches what we'd expect. Bolting a chatbot onto a broken process gives you a broken process with a chatbot. The companies getting results change how the work flows: who checks what, which steps disappear, where a person still decides.

The same survey found that 80% of companies set efficiency as their goal, but the ones seeing the most value usually aimed for growth or innovation as well. Cost-cutting alone tends to produce cost-cutting-sized results.

What the market is getting right

It's easy to read all this as doom. It isn't. A few things are going genuinely well.

AI is getting dramatically cheaper. Stanford's 2025 AI Index measured the cost of reaching GPT-3.5-level performance falling from $20 to $0.07 per million tokens between November 2022 and October 2024, roughly a 280-fold drop. Use cases that didn't make financial sense two years ago now do.

Narrow agents work. Andrew Ng showed that wrapping an older model in an "agentic" loop (plan, draft, check, revise) lifted GPT-3.5's score on a coding benchmark from 48.1% to as high as 95.1%, better than GPT-4 working alone. Good process design can matter more than which model you buy.

People are already using it. Microsoft and LinkedIn found 75% of knowledge workers using AI at work back in 2024. Demand isn't the problem. Channelling it safely is.

What the experts are saying

The most interesting voices right now are the ones building the technology, and they're more cautious than their marketing departments.

Sam Altman, OpenAI's CEO, agreed in August 2025 that investors are overexcited about AI: "When bubbles happen, smart people get overexcited about a kernel of truth," he told reporters, while still calling AI the most important technology in a very long time.

Satya Nadella, Microsoft's CEO, argued that self-declared AI milestones mean little. "The real benchmark is: the world growing at 10%," he said. In other words, judge AI by economic output, not demos.

Andrej Karpathy, a founding member of OpenAI, pushed back on the "year of agents" hype, calling it a decade-long project and saying today's agents "just don't work" as replacements for employees yet. He still uses them every day.

Gartner went further and predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, estimating that only about 130 of the thousands of vendors selling "agentic AI" actually offer it.

Put those together and a clear picture emerges. Nobody serious thinks AI is a fad. But almost everyone serious thinks the hype has outrun what most companies can actually deploy today.

The 15 problems we'll be writing about

Over the next two weeks we're publishing a deep dive on each of the problems these studies keep pointing to, with the evidence and what actually fixes it:

  1. Pilots that never reach production
  2. AI that confidently makes things up
  3. Data that isn't ready
  4. Nobody can prove the ROI
  5. "Agentic AI" that isn't
  6. Prompt injection and AI security
  7. Shadow AI: employees using tools you didn't approve
  8. Productivity gains that turn out to be imaginary
  9. Replacing people too fast
  10. Bias you didn't notice until it was too late
  11. India's new rules: DPDP and the RBI's AI framework
  12. Adding AI without changing the workflow
  13. No way to test whether it's working
  14. Surprise costs, and why prices keep falling
  15. "Is this a bubble? Should we wait?"

One honest observation before we start: none of these problems is really about artificial intelligence. They're about data, process, people and judgement. That's good news, because those are things a company can actually fix.

Sources

Want help applying this in your organization?

Get a free readiness assessment

More from the blog

Strategy · 7 min read

Problem 1: Why most AI pilots never make it to production

Studies from MIT, IBM and Gartner agree that most AI pilots stall. The reason is rarely the model. Here is what goes wrong and how to design a pilot that survives.

Read → Governance · 7 min read

Problem 2: AI that confidently makes things up (and the real bills it has run up)

An airline, a law firm and a Big Four consultancy all paid for AI hallucinations. What happened in each case, and the practical safeguards that would have caught it.

Read →