Picture a senior executive at a well-known public tech company. She runs roughly a thousand engineers. Nearly every one of them uses AI. And yet, as Azeem Azhar recounted in Exponential View, the needle won't move on her bottom line.
This isn't an isolated story. MIT research found that roughly 95% of generative AI pilots fail to scale to production or deliver measurable P&L impact. Tools everywhere, ROI nowhere. It's become the defining management frustration of 2026.
The standard diagnosis blames the usual suspects: bad data, wrong tooling, immature models. Those problems are real, but they're not why most rollouts stall. Canva's AI Discovery Week (in which the company paused normal work for 5,300 employees to spend an entire week doing nothing but learning AI) found something far more humbling: even among employees who were already AI-fluent, the biggest blockers were human ones.
Not technical. Human.
That's the insight executives keep treating as an afterthought. The gap between a successful pilot and enterprise-wide ROI isn't a model problem or an integration problem. It's a behavior problem. And until leadership treats it like one, companies will keep funding pilots that flatter the board deck and disappear from the financials.
The Pilot Trap
Here's how it plays out. A small team (usually motivated, technically curious) runs a pilot. The conditions are nearly perfect: everyone opted in, the scope is contained, the use case is cherry-picked, and the participants want it to work. Results look strong. Leadership approves a broader rollout.
Then the real organization shows up.
The rollout lands on people who didn't choose to participate. Managers who weren't involved in the pilot now have to absorb it into their reporting structures. Employees who weren't in the room see the new tools as something that happened to them, not something they helped build.
The pilot graveyard fills up fast. Weekly active users plateau at embarrassing numbers. The tool gets a polite mention in all-hands. Nobody has the heart to call it dead.
Pilots aren't the problem. Pilots are valuable. The mistake is believing that a successful pilot proves the rollout will work -- that once the technology is validated, the hard part is done. It isn't. Proving the technology works is the easy part. Getting 800 people who didn't run the pilot to actually change their daily behavior? That's the hard part.
Three Human Blockers Executives Underestimate
1. Why Fluency Doesn't Spread
Canva's experiment was unusual in a specific way. The company didn't just offer training sessions -- it stopped regular work entirely. For a week. Speakers from OpenAI, Anthropic, and Google were flown in. Teams ran hackathons. Employees could actually tinker without feeling guilty about their inboxes.
What they found was counterintuitive. Employees who became genuinely skilled with AI tools often couldn't transfer that excitement to colleagues who weren't in the room. Technical fluency and the ability to evangelize AI to peers are two completely different skills. One doesn't produce the other automatically.
This matters because research from Microsoft shows that peer influence is the single strongest driver of sustained AI adoption -- far more powerful than top-down mandates or formal training programs. Workers whose colleagues actively use and discuss AI tools are nearly 9 percentage points more likely to become heavy AI users themselves. That multiplier sits mostly untapped in most organizations, because nobody builds the bridge between "person who's great at AI" and "person who can make others great at AI."
Most enterprise AI programs identify power users, hand them slightly better tools, and call it an ambassador program. That's not enough. What those power users actually need is communication coaching: how to run a demo that lands for a skeptical audience, how to translate a productivity gain into terms a nervous colleague finds reassuring, how to pull a non-user into a genuine experiment rather than a sales pitch.
What to do: Build a formal peer-teaching layer. Select internal AI advocates not only for technical skill but for their ability to teach and translate. Give them time on team meeting agendas -- thirty minutes, not a full workshop. Build a shared library of short, role-specific demos (five minutes max, tied to that team's actual work). The goal isn't an internal lecture series. It's creating enough "wait, this actually works for me" moments that adoption becomes self-sustaining.
2. AI as a Performance Review
There's a reason employees quietly reduce their AI use when they think others are watching, even when it hurts their output. One study found AI usage dropped 14% and task accuracy fell 3.4% when workers knew their AI use was visible to evaluators. The fear wasn't job loss. It was judgment: using AI might signal a lack of confidence in their own abilities, marking them as someone who needs help.
AI tools increasingly handle the exact tasks that workers built their professional identities around. A senior analyst who takes pride in synthesizing data doesn't want to see a language model do it in 40 seconds. A lawyer who spent a decade learning to draft contracts doesn't want to feel like those hours are suddenly worth less. Research published in HBR found that workers experience AI not just as a threat to employment but to expert status and professional recognition -- a referendum on whether their skills still matter.
The "AI shaming" data makes this concrete. Employees judge colleagues who visibly rely on AI as less competent and lazier, even when the AI-assisted output is better. That social cost is real, and it creates quiet, rational resistance that's almost invisible in a survey but very visible in adoption metrics.
When you push AI tools into a team that hasn't been part of the framing conversation, you're not just asking them to learn software. You're asking them to accept an implicit argument about their own replaceability. Most people don't do that warmly.
What to do: Change the narrative before deploying the tool. Position AI adoption as the acquisition of a new professional skill, not evidence that the old one is obsolete. Leaders should talk openly about what they personally use AI for -- including the tedious tasks it saves them from -- to normalize AI use as a professional capability rather than an admission of weakness. Canva's approach framed their week around helping employees discover what AI could do for their specific role, rather than broadcasting generic capability claims from the top.
Designate "no judgment" learning sprints where the only goal is experimentation. When people feel psychologically safe to look uncertain in front of AI, they actually learn to use it.
3. Middle Managers Do the Math
Middle managers are the make-or-break layer in any enterprise rollout, and most AI strategies treat them as passive conduits rather than active decision-makers with their own incentive structures.
Team-level AI adoption is 2.1 times higher when the direct manager actively supports and models AI use, according to Gallup. Only 28% of employees in AI-implementing organizations strongly agree their manager actively supports AI use. That's a wide gap between what's needed and what's happening.
The reason isn't reluctance. It's math. Most managers are measured on quarterly outputs, project delivery, and team performance. Asking them to dedicate team time to AI experimentation and learning is asking them to absorb short-term friction for a benefit that may show up in someone else's KPIs -- or whose timing is genuinely uncertain.
The middle management AI veto rarely looks like overt resistance. It looks like a manager who keeps pulling the AI champion back to "urgent" work. It looks like team learning sessions that get canceled when a deadline appears. It looks like someone who says the right things in steering committee and never creates actual space for their team to experiment. Passive non-participation is just as lethal to a rollout as active resistance. It's just quieter.
Cultural encouragement doesn't change the math. The incentive structure itself has to change.
What to do: Tie AI adoption to manager performance metrics. Not vague "supports AI culture" language -- specific observable behaviors. Did the manager run at least two workflow demo sessions this quarter? Did team AI tool usage hit the adoption threshold? Is there at least one AI-assisted process change the team owns? Put these in the performance framework with real weight, not buried in a footnote.
Also, cut the preparation burden for managers wherever possible. Pre-built learning resources -- thirty-minute, role-specific modules requiring no manager preparation -- reduce the effort from "I have to design a training session" to "I have to press play." Canva had the luxury of shutting down for a week. Most organizations can't do that. But they can make adoption low-effort enough that a busy manager will actually follow through.
The Mid-Market Version: Same Ingredients, Smaller Dose
Canva is worth tens of billions of dollars and can fly speakers in from OpenAI. Most companies aren't Canva. The good news is you don't need the week-long immersion to create the same behavioral shift. You need the same ingredients at a smaller scale.
Protected time. The most important thing Canva gave employees wasn't content -- it was permission to stop and think. Replicate this with a monthly "AI sprint afternoon": two to three hours, no regular work, with a structured experiment prompt. Small, but protected from cancellation.
Role-specific relevance. Generic AI training fails the same question every time: "What does this actually do for my job?" Build or curate at least five short demos tied to your most common roles. Sales, customer success, finance -- each team gets different use cases and scripts. The goal is shortening the path from abstract capability to personal "aha."
Visible peer success. Two or three people per team -- not dozens -- who are openly using AI to do work that colleagues recognize as good. This is the social proof mechanism that Microsoft's research identifies as the primary adoption driver. These people don't need a title or budget. They need time on the team meeting agenda.
Manager accountability. One AI-related behavior metric per manager per quarter. Make it visible to their manager. Start simple -- "ran one team demo session" -- and build from there.
Safe-to-fail framing. Say it explicitly before anything else: using AI imperfectly is fine, running a prompt that doesn't work is fine, not knowing the answer is fine. The identity threat is real, and it needs to be pre-empted with actual language, not assumed away by a vague "we encourage experimentation" in a slide deck.
None of this requires a transformation office or a dedicated budget line. It requires a leadership decision to treat behavior change as first-order work, not a footnote to the technology deployment.
The Real ROI Question
When AI pilots succeed and rollouts don't, the instinct is to fix the technology. Buy a better model, upgrade the integration, hire AI engineers.
Those might be the right calls in specific situations. But in most organizations, Azeem Azhar's diagnosis holds: the tools work, adoption is real in small pockets, and the company is still carrying infrastructure costs while waiting for organizational change to catch up. The gap between investment and return isn't technical. It's human.
The companies that crack this won't have the best models. They'll be the ones who treated the communication gap, the identity threat, and misaligned manager incentives as serious organizational problems -- with the same seriousness and follow-through they bring to deploying the technology itself.
Most executives still think it's a tooling problem. It isn't.