Sam Altman recently revealed that OpenAI's internal benchmark, GDP-Val, shows GPT-5.2 Pro now beats or ties human experts 74% of the time across 40+ business tasks. Six months ago, GPT-4o hit 38%. That's a near-doubling of expert-level performance in half a year.
And yet, according to Altman himself, most companies extract only about 10% of available AI value.
That's the "capability gap" Altman keeps referencing: the chasm between what AI can do and what organizations are actually using it for. His prediction for 2026 is that this gap will narrow, but "probably not as fast as you'd expect."
Here's what bothers me about this framing: it puts the burden on capability. As if the problem is that AI isn't good enough yet, and once it gets a little better, adoption will follow.
The data says otherwise.
The Real Gap: Deployment, Not Capability
A 2025 MIT study found that despite $30-40 billion in enterprise AI investments, 95% of generative AI pilots produce no measurable ROI. Only 5% of pilots ever make it into production workflows.
Let that sink in. We're not talking about cutting-edge experimental projects. We're talking about the vast majority of enterprise AI initiatives failing to deliver value, not because the technology isn't capable, but because organizations can't operationalize it.
S&P Global's 2025 data tells the same story from a different angle: 42% of companies abandoned most of their AI initiatives this year, up from just 17% in 2024. The average organization scrapped 46% of proofs-of-concept before production.
This isn't a capability gap. It's a deployment gap. And the distinction matters because it changes what enterprises should be doing about it.
The Dabbling Problem
Here's what no one is saying: the biggest risk for enterprises right now isn't falling behind on AI adoption. It's adopting AI wrong.
I call it the "dabbling problem." Organizations launch AI experiments. They spin up pilots. They run proofs-of-concept. They check the box that says "we're investing in AI." And then they wait for the technology to get good enough that the rest takes care of itself.
It doesn't work that way.
Cal Newport's analysis of why AI failed to "join the workforce" in 2025 captures this perfectly. The tech leaders promised AI agents would materially change company output. What actually happened? As Newport puts it, products that "fell laughably short," with agents spending fourteen minutes trying to select a value from a dropdown menu.
But the failure isn't technical. Gary Marcus called the underlying models "clumsy tools on top of clumsy tools." Andrej Karpathy acknowledged "overpredictions going on in the industry." The real lesson is that we've been measuring the wrong thing. Capability keeps improving. Adoption stays stuck.
Why Pilot Purgatory Happens
The CIO research on enterprise AI identifies four critical misconceptions derailing adoption:
The Organizational Readiness Illusion: Leaders equate acquiring technology with building organizational capability. They underestimate how disruptive AI is to structures, power dynamics, and workflows. The research notes that "firms struggle to capture value not because technology fails, but because people, processes and politics do."
AI Expectation Myths: Leaders assume AI can solve any business challenge with guaranteed quick returns. In reality, AI delivers measurable value only in targeted, well-defined use cases. This creates "pilot paralysis": many experiments, but few reach production.
Data Readiness Bias: Organizations prioritize data volume over quality and governance. While 91% acknowledge that reliable data is essential for AI success, only 55% believe they actually have it.
The Deployment Fallacy: Treating AI like traditional software ignores that AI systems are probabilistic and require continuous lifecycle management, not set-and-forget approaches.
These aren't capability problems. They're organizational problems. And organizations solve them through commitment, not experimentation.
The EOD Lesson: You Don't Dabble in Explosive Ordnance
In Navy EOD, we didn't "experiment" with techniques on live devices. We didn't run "pilots" to see what might work. We executed with a plan, every time.
The difference between a bomb technician and someone who used to be a bomb technician often comes down to whether they treated every operation as a fully committed mission or as an opportunity to "try something."
I see the same dynamic in enterprise AI adoption. Organizations that treat AI as something to explore, to experiment with, to test when convenient, end up in pilot purgatory. They accumulate proofs-of-concept without accumulating capability.
Organizations that treat AI as an operational commitment, something that requires restructuring workflows, retraining staff, redesigning processes, and reallocating resources, actually capture value.
Altman himself made this point, though it often gets buried: "The biggest wins come from redesigning work around agents, rather than bolting AI onto old tools."
Redesigning work isn't dabbling. It's commitment.
What the Data Says About What Works
The MIT research revealed a telling split. Purchasing AI tools from specialized vendors succeeds about 67% of the time. Internal builds succeed only 33% of the time. The difference isn't about technical competence. It's about commitment level. Buying a specialized tool means you've decided to integrate it into operations. Building internally often means you're still figuring out whether AI is valuable.
The study also found that more than half of generative AI budgets go to sales and marketing tools, yet the biggest ROI exists in back-office automation: eliminating business process outsourcing, cutting external agency costs, streamlining operations.
Organizations are dabbling where it's visible (sales and marketing) instead of committing where it's valuable (operations). That's not a capability gap. That's a strategy gap.
The Capability Gap That Actually Matters
Altman talks about the gap between what AI can do and what people use it for. The framing implies that once AI gets capable enough, adoption will follow.
The real capability gap is organizational: the gap between recognizing AI's potential and building the organizational capacity to realize it.
This gap doesn't narrow by waiting for better models. It narrows through:
Strategic commitment over experimentation. Pick specific use cases with measurable outcomes. Staff them properly. Hold leaders accountable for results, not for "innovation."
Workflow redesign over tool adoption. Don't bolt AI onto existing processes. Redesign processes around AI's actual capabilities. This requires disruption, which is why most organizations avoid it.
Operational integration over proof-of-concept. The MIT data shows that only 5% of pilots make it to production. That's because most pilots are designed to demonstrate possibility, not to prove operational viability. Change the success criteria.
Manager reinforcement over technology access. The most damaging assumption of 2025, according to multiple industry analyses, was that "once the tools are good enough, adoption will follow." In practice, behavior follows incentives, muscle memory, and manager reinforcement, not tool quality.
The Risk of Continued Dabbling
I wrote about the implementation gap in The AI Safety Gap No One Is Talking About. The argument there was that enterprises have enough evidence to act on AI security but aren't doing it.
The same dynamic applies to AI adoption more broadly. The evidence is clear:
- AI models perform at or above expert level on many business tasks
- 95% of pilots fail to deliver measurable ROI
- Organizations that commit to workflow redesign capture value
- Organizations that dabble end up in pilot purgatory
The risk of continued dabbling isn't just wasted investment. It's opportunity cost. While your organization runs another proof-of-concept, competitors who committed to operational integration six months ago are compounding their advantage.
The Cisco AI Readiness Index shows only 13% of companies get consistently measurable returns from AI. That 13% isn't waiting for better models. They're operationalizing the capabilities that already exist.
What This Means for 2026
Altman predicts the battle for enterprise AI will shift from "which model is smartest" to "which platform handles my company's data, agents, and workflows best."
That's the right frame. But it requires organizations to stop thinking about AI adoption as a technology acquisition problem and start thinking about it as an organizational change problem.
The capability gap Altman describes won't close because AI gets better. It will close when enterprises stop dabbling and start committing.
For organizations still in pilot purgatory, the question isn't "when will AI be ready for us?" The question is "when will we be ready for AI?"
The technology has been ready. The organizational capability hasn't. And no amount of model improvement changes that.
