Inside OpenAI, the Message Is Clear: The AI Bottleneck Isn't the Models, It's Adoption

Maggie Hott, a go-to-market leader at OpenAI, posted something on LinkedIn this week that had nothing to do with model benchmarks or GPU shortages. Her argument was that the real constraint on AI right now is people who know how to help other people use it.

"Most organizations don't have a model problem," she wrote. "They have a people problem." Her advice to new graduates was blunt: spend less time trying to become the world's best prompt engineer and more time learning how to help organizations adopt AI.

What she's seeing in the field

Hott says she has spent nearly every week this year meeting with healthcare executives, an industry that should be one of AI's biggest winners given how much of the work is administrative. Their top request is almost never another model demo. It's help driving adoption.

That's a telling thing to hear from inside one of the leading AI labs. The companies building the models are finding that their hardest customer conversations are not about what the technology can do. They're about getting thousands of employees to change how they work, which is slow, unglamorous, and has very little to do with the software itself. A model can be deployed in a weekend. Changing daily habits, workflows, and trust across a workforce takes quarters or years, and it doesn't happen just because leadership bought licenses.

The numbers tell the same story

Research from this year keeps landing on the same conclusion.

WRITER's 2026 enterprise survey found that 79% of organizations face challenges adopting AI, up double digits from the year before, even though most companies in the survey are spending over $1 million a year on AI. More than half of the C-suite executives surveyed said adoption is creating serious internal friction.

PwC's January 2026 research found 56% of CEOs report zero measurable ROI from AI despite having deployed it. Change management and workflow redesign now rank above the technology itself as the main constraints.

Deloitte's 2026 survey of more than 3,200 leaders named the skills gap as the top barrier to integration. The follow-through is weak, though. While 80% of tech-focused organizations say upskilling is the best way to close the gap, only 28% plan to actually invest in training programs over the next few years.

Then there's the trust problem. WalkMe surveyed 3,750 enterprise workers in 2026 and found that only 9% of frontline employees trust AI with complex, business-critical decisions. Among executives, that number is 61%. The people approving the rollouts are the most confident, and the people who have to live with them are the most skeptical.

None of these stall points are technical. Unclear ownership, no training, tools that don't fit the actual workflow, employees who never got an answer to "what's in it for me." These are human problems, which is exactly Hott's point.

The labs are putting money behind this thesis

Here's the part that makes Hott's post more than just career advice: her employer, and its biggest rival, have both spent this year building entire businesses around the adoption problem.

In May, OpenAI launched the OpenAI Deployment Company, a majority-owned subsidiary that raised over $4 billion from 19 investors including TPG, Goldman Sachs, and SoftBank. Its whole purpose is to embed engineers inside client organizations to build AI systems within their real infrastructure, compliance constraints, and messy legacy environments. OpenAI acquired UK consulting firm Tomoro to staff it with roughly 150 deployment engineers on day one, and just this month added a second acquisition, Northslope, expanding the bench to hundreds.

A week earlier, Anthropic announced its own version: a $1.5 billion enterprise services venture with Blackstone, Hellman & Friedman, and Goldman Sachs, built on the same premise of embedding engineers inside companies and redesigning workflows around AI rather than shipping software and walking away.

Both are borrowing from Palantir's forward deployed engineer playbook, and both are effectively an admission of the thing Hott said out loud. If the models sold themselves, none of this would be necessary. The labs would not be raising billions to build consulting arms and competing with Accenture and Deloitte if deployment were a solved problem. The fact that the two most prominent AI companies in the world decided, within days of each other, to get into the implementation business tells you where they think the bottleneck actually is.

What this means for the job market

If the labs are right, the most valuable emerging role in AI may not be the ML researcher or the prompt specialist. It's the person who understands the technology well enough to know what it can do, and understands organizations well enough to get people to actually do it.

This resembles the change management discipline that grew up around ERP and cloud migrations, with one difference. AI capability is compounding faster than any previous enterprise technology, so the gap between what the tools can do and what organizations actually do with them keeps widening. Roles like AI enablement lead, adoption strategist, and forward deployed engineer are showing up in job postings across healthcare, financial services, and manufacturing, often reporting into the business rather than IT, because adoption gets won or lost inside individual teams' workflows.

The takeaway for leaders

Buying AI is not adopting AI. Organizations that treat deployment as the finish line are the ones filling out the 79% that are struggling and the 56% seeing no return. The ones pulling ahead treat adoption as an ongoing capability, with dedicated enablement staff, continuous training, redesigned workflows, and executives who use the tools themselves instead of just mandating them.

Models will keep getting smarter. Everyone agrees on that. Hott's closing line is the one worth remembering: the people who help others grow alongside those models may be the hardest to replace.