Why Your AI Transformation Is Stalling at Middle Management
The C-suite is excited. The developers are excited. The people in between have misaligned incentives.
Your CEO came back from Davos on fire about AI. Your engineers are already using Cursor and Claude and Codex and have been for months. You've got a board slide about "AI-first strategy" and a consulting firm has been paid good money to produce a roadmap.
And yet. Nothing is moving.
The research confirms this is not just your company. PwC's 2026 Global CEO Survey — 4,454 CEOs across 95 countries — found that 56% report AI has delivered neither higher revenues nor lower costs over the past 12 months [1]. Only 12% of those CEOs can claim both cost savings and revenue gains. BCG found that 74% of companies have yet to achieve tangible value from AI investments [2]. Gartner had predicted that 30% of generative AI projects would be abandoned after proof of concept by end of 2025 [3] — and the evidence suggests they were right.
Meanwhile, 88% of organizations say they use AI in at least one function [4]. Sounds impressive until you read the footnote: only 33% have begun scaling AI across the enterprise, and just 6% qualify as actual high performers — organizations where AI measurably moves EBIT. Goldman Sachs, in March 2026, put the macro picture bluntly: "We still do not find a meaningful relationship between productivity and AI adoption at the economy-wide level." [5]
The technology works. In specific, well-scoped deployments where companies actually measure results, McKinsey finds median productivity gains of roughly 30% [4]. The problem isn't whether AI can deliver. It's whether organizations can get it out of the pilot phase and into production.
The bottleneck isn't technology. It's not budget. It's not even culture in the abstract. It's a specific layer of your organization: directors and VPs who have every incentive to slow this down and just enough power to do it.
This is the frozen middle problem. And it's quietly killing more AI transformations than any technical failure ever will.
Why Middle Managers Are Specifically Threatened
Before you dismiss this as organizational cynicism, understand the structural reality.
Middle managers exist to do three things: aggregate information from the front lines, coordinate across functions, and translate between strategy and execution. They are the nervous system of the organization. The people who know what's actually happening and can turn that into action.
AI is extraordinarily good at all three of those things.
Not eventually. Now. Today. LLMs synthesize information across sources. Agentic workflows coordinate handoffs across systems and teams. AI tools translate messy business problems into structured outputs with (almost) zero headcount.
This isn't a 10-year threat. It's a this-quarter threat. And the directors and VPs who've built careers on being the connective tissue of the organization can feel it, even if they can't articulate it.
The organizational irony is brutal: the people most threatened by AI are exactly the ones with the most power to block its adoption. They control budget approvals. They set workflow norms. They gate access to systems. They define what "success" looks like for a pilot. They write the performance reviews of the people trying to push AI forward from below.
This isn't malice. It's self-preservation. And it's completely rational, which makes it hard to fight.
Three Archetypes You've Already Met
You know these people. You've been in meetings with them. Maybe you've worked for one. There are three patterns that show up reliably.
Archetype 1: The Pilot Purgatory Director
This one is enthusiastic about AI. Genuinely! They'll fund a pilot. They'll sit in the demo. They'll nod along to the business case.
Then the pilot ends and... nothing happens. Because before we move to production, we need to address a few more edge cases. And now the success criteria have shifted slightly. And actually, we should probably run it in parallel with the existing process for another quarter, just to be sure.
The pilot lives forever because production means accountability. A pilot that doesn't graduate is just a learning experience. A production deployment that underperforms is a career event. Deloitte's research found executives openly acknowledging this: "PoCs using dummy data create false optimism; real data reveals underlying problems" [7].
So the pilot keeps running. The team that built it gets reassigned. The vendor eventually stops following up. And the director gets to say, truthfully, that they ran an AI pilot.
Archetype 2: The "Not My Budget" VP
The tool costs $50K per year. The business case shows $500K in productivity gains. The math is obvious.
But: "It's not in this year's budget." Budget planning is in Q3. We'll put it on the list for next year. (It never makes the list.)
This one is particularly insidious because it's not technically a no. It's a process no. It hides behind real organizational constraints, legitimate ones even, while achieving the same outcome as an outright veto.
Every budget cycle adds another year of delay. The bar for "next year's budget" keeps moving. The tool that would have transformed the workflow in 2025 is now "table stakes" by the time it gets approved in 2027, and the window of advantage is gone.
Archetype 3: The "Security Says No" Deflection
Security concerns are real. Data privacy matters. Governance is important. I am not dismissing any of that.
But "security said no" as a conversation-ender, with no risk assessment, no proposed controls, no path to resolution, is not a security posture. It's a veto dressed up in compliance language.
The tell: when security concerns are raised without any corresponding question of "what would need to be true for this to be approved?"
Real security work is solvable. Data classification frameworks, access controls, approved model lists, output auditing: these are engineering problems with engineering solutions. The deflection version skips all of that and lands at "no" without ever visiting the middle.
When you hear "security says no" with no further texture, you're not looking at a security problem. You're looking at someone using security as a shield.
The Organizational Trap
Here's the dynamic that makes this so hard to unblock from below.
The IC who wants to use AI tools goes to their manager for approval. The manager says not yet. The IC pushes back with a business case. The manager escalates to their VP for a budget exception. The VP says not this quarter. The IC, now months into the process, quietly starts using the free tier of something and not telling anyone.
Shadow IT is not the failure mode. It's the symptom. When AI adoption goes underground, it means the official channels have failed. And underground AI adoption is genuinely riskier than sanctioned adoption, with no governance, no data controls, no visibility. Research suggests 50–60% of workers are already using unsanctioned AI tools [6] — meaning frontline adoption is happening around middle management, not through it.
The frozen middle creates the exact risk it claims to be preventing. And there's a harder dynamic underneath it: Gartner predicts that by 2026, 20% of organizations will eliminate more than half of their current middle management roles via AI [3]. Middle managers aren't imagining the threat. They're responding rationally to a real one. The organizational irony compounds: the more aggressively the C-suite pushes AI, the stronger the survival instinct of the layer being asked to implement it.
Meanwhile, the C-suite is issuing mandates and wondering why the roadmap is behind. The consulting firm is writing a follow-up deck about change management. And the directors and VPs are nodding along in the all-hands and then returning to their desks to find another reason the pilot isn't ready for production.
What Actually Unblocks This
There are strategies that work. None of them are easy and most of them require someone above the frozen layer to actually care about outcomes, not just optics.
Executive sponsorship with teeth, not words.
"The CEO is committed to AI" means nothing if there's no consequence for blocking it. Sponsorship with teeth looks like: a named executive who reviews AI initiative status quarterly, with authority to break budget and approval logjams. Not a steering committee. A person. With accountability.
The frozen middle is rational. It responds to incentives. If the consequence of blocking AI adoption is a difficult conversation with a senior executive who actually follows up, the calculus changes.
Bottom-up pressure through visible wins.
Pilots that never graduate to production stay invisible. Wins that are visible to the organization create pressure the frozen middle can't ignore.
This means: deliberately publicizing AI wins, even small ones, through internal channels. Town halls, team newsletters, lunch-and-learns where teams show what they've built. When peers see other teams shipping real AI workflows and getting recognized for it, the "not my budget" defense gets harder to hold.
Make AI adoption part of how managers are evaluated.
If performance reviews for directors and VPs include no signal on AI adoption, you're rewarding inaction. Full stop.
This doesn't mean punishing people for being careful. It means including adoption metrics in the conversation: What pilots have you run? What graduated to production? What's your team's AI literacy? What tools has your function implemented?
The people who get measured on it will find a way to make it happen. The people who don't will find a way to explain why it isn't their problem.
Pre-approved pilot frameworks.
A significant chunk of "not my budget" and pilot purgatory is friction in the approval process. If every AI pilot requires a custom business case, security review, budget exception, and VP sign-off, you've made piloting so expensive that only the most motivated teams will attempt it.
Pre-approved pilot frameworks flip this: a pre-cleared list of tools, a standard data classification review, a budget envelope that managers can access without an exception process, and a graduation checklist that defines what "production ready" actually means.
When the path to yes is shorter, more people take it. You won't eliminate the frozen middle entirely, but you can reduce the surface area it has to block.
The Bottom Line
Your AI transformation is not stalling because the technology isn't ready. It's not stalling because your people don't get it. It's stalling because a layer of your organization has a structural incentive to slow it down and the organizational power to do so.
The frozen middle is not a culture problem. It's an incentive problem. Culture is downstream of incentives.
The fix requires C-suite leaders to do something harder than issuing mandates: they have to redesign the incentive structure for the layer below them, measure what matters, and be willing to have uncomfortable conversations when adoption stalls.
The alternative is another year of pilot purgatory, another budget cycle where AI tools don't make the list, and another all-hands where someone asks why the AI roadmap is behind.
The technology is ready. The question is whether the organization is willing to get out of its own way.
Are you a middle manager who sees other hurdles to adoption or success? I'd love to get your perspective -- let's discuss in the comments. Regardless of role, what pattern have you run into most: pilot purgatory, budget deflection, or the security veto? Drop it in the comments. I'm curious whether one of these dominates, or whether it's different by industry.
References
- PwC (January 2026). 29th Global CEO Survey: Leading Through Uncertainty in the Age of AI. Survey of 4,454 CEOs across 95 countries. https://www.pwc.com/gx/en/issues/c-suite-insights/ceo-survey.html
- BCG (October 2024). AI Adoption in 2024: 74% of Companies Struggle to Achieve and Scale Value. Survey of 1,000 CxOs and senior executives across 59 countries. https://www.prnewswire.com/news-releases/ai-adoption-in-2024-74-of-companies-struggle-to-achieve-and-scale-value-302285294.html
- Gartner (July 2024). Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept by End of 2025. https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025
- McKinsey & Company (September 2025). The State of AI: Global Survey 2025. Survey of 1,993 participants across 105 countries. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
- Goldman Sachs / Fortune (March 2026). Goldman finds no meaningful relationship between AI and productivity at the economy-wide level. https://fortune.com/2026/03/03/goldman-earnings-ai-anxiety-no-meaningful-impact-productivity-economy-30-percent-in-2-areas/
- SecureWorld (2025). Frozen in the Middle: The AI Bottleneck. Citing Salesforce data on unsanctioned AI tool usage. https://www.secureworld.io/industry-news/frozen-middle-ai-bottleneck
- Deloitte Global (October 2025). AI ROI: The Paradox of Rising Investment and Elusive Returns. Survey of 1,854 senior executives across 14 countries. https://www.deloitte.com/global/en/issues/generative-ai/ai-roi-the-paradox-of-rising-investment-and-elusive-returns.html
If this resonated, here are some related articles:
- Are companies really doing layoffs "for AI"? (the structural threat middle managers feel is real — this article examines what's actually happening when companies cite AI as a reason for workforce reductions): LinkedIn | Substack
- What the Beer Game teaches us about AI adoption (the bullwhip effect hits AI rollouts too, and the frozen middle is exactly where the distortion originates): LinkedIn | Substack
- Dunning-Kruger, now available at enterprise scale (why organizations confidently underestimate what AI adoption actually requires — and how overconfidence at the top feeds resistance in the middle): LinkedIn
- Situational leadership applied to AI (the same incentive logic that freezes middle managers applies when deciding how much autonomy to grant your AI collaborator): LinkedIn | Substack
Keith MacKay is a technology strategy consultant and CTO in EY-Parthenon's Software Strategy Group (SSG), specializing in AI disruption and technology diligence for private equity and corporate clients. SSG's AI Disruption Lab conducts rapid assessments of how AI transforms and threatens existing business models and value chains. Keith teaches at Northeastern University and writes about strategy, management, and AI/technology, with an AI collaborator.