Companies have poured billions into generative AI. They have rolled out tools like Copilot and ChatGPT to tens of thousands of employees. They have launched hundreds of pilots. Yet the promised revolution in organizational productivity remains stubbornly out of reach.
The problem, according to new research from Harvard Business School’s Digital Data Design Institute (D³) and Microsoft, is not the models. It is not the data. It is the last mile of transformation — the messy, human, organizational stretch where technical capability meets real operating models.
In a major new study published in Harvard Business Review, professors and practitioners Karim R. Lakhani, Ali Hosseini, and Sara Mahdavi lay out why AI adoption stalls at scale.
Drawing on the Frontier Firm Initiative and insights from senior leaders across banking, healthcare, manufacturing, and professional services, they identify **seven critical frictions** that keep organizations “pilot-rich but transformation-poor.”
Here are the seven barriers — and the concrete blueprint executives must follow if they want to move from isolated productivity wins to enterprise-wide AI-native operating models.
1. The Proliferation of Pilots
Most companies are swimming in successful small-scale experiments — yet almost none scale into standard operating practice.
A global investment bank built more than 250 LLM-connected applications. A global apparel company automated over 18,000 finance processes. A food-and-beverage giant ran AI pilots in 185 countries. The results stayed local. Without a repeatable path from proof-of-concept to enterprise rollout, pilots become expensive science projects.
2. The Productivity Gap
Individual workers report impressive time savings, but organizations rarely see corresponding gains in headcount, cycle time, or output.
A global payments network achieved 99 % employee adoption of copilots — yet could point to zero measurable improvement in key financial metrics. Freed-up time simply gets re-absorbed into more meetings and email. Without deliberate redesign of roles, budgets, and incentives, the productivity dividend evaporates.
3. The Burden of Process Debt
AI shines a harsh light on decades of fragmented, exception-ridden workflows built through acquisitions, geographic variations, and legacy systems.
A large healthcare insurer discovered that AI exposed inconsistencies faster than the organization could resolve them. A professional services firm operating in 170+ countries ran the “same” process dozens of different ways depending on location. You cannot simply layer AI on top of broken processes; you must re-architect them from the ground up.
4. The Identity Problem of Tribal Knowledge
The deepest expertise in most organizations still lives in the heads of long-tenured experts — and those experts are often reluctant to externalize it.
AI demands that judgment be encoded into systems, which threatens the very identity of people whose status came from being “the one who knows.”
An engineering consultancy described this as a classic identity crisis: people protected their expertise because it conferred power. Until organizations address the human side of knowledge capture, AI will hit a wall of silent resistance.
5. Governance in an Agentic World
When AI agents move from answering questions to taking actions, traditional governance collapses.
A global bank watched its human-in-the-loop controls break down under multi-agent architectures. Another institution is already running 100+ agents and planning for tens of thousands.
Questions that used to belong to HR — Who creates agents? Who owns their performance? How do you “onboard,” evaluate, or retire them? — now sit unanswered at the heart of operations.
6. Architectural Complexity
Most large companies operate a patchwork of platforms across multiple hyperscalers and vendors. Integration becomes a nightmare.
A global apparel company spent months just getting agents across SAP, Microsoft, and Google to communicate reliably. Platform updates frequently force teams to reset entire initiatives. The tension between “single-platform speed” and “multi-vendor flexibility” is real and unresolved.
7. The Efficiency Trap
When leaders frame AI primarily as a cost-cutting tool (the modern version of offshoring), they trigger defensive behavior and dramatically narrow ambition.
Boards demand quick ROI, so companies optimize low-value tasks performed by $20-an-hour employees instead of reimagining how the enterprise creates and captures value. The risk: hollowing out the very human capabilities — judgment, storytelling, creativity — that differentiate great companies.
The Blueprint: Seven Strategic Pillars for the AI-Native Firm
The authors do not stop at diagnosis. They offer a clear set of counter-measures that together form the foundation of an AI-native operating model:
1. Clean-Sheet Process Redesign
Stop patching legacy workflows. Design processes from scratch with AI agents as first-class participants. Use agent-centric mapping to define exactly where humans orchestrate and where agents execute.
2. Strategic Knowledge Capture
Treat tribal knowledge as a corporate asset. Pair domain experts with “AI process architects” and knowledge stewards to externalize both rules and nuanced judgment. Frame it positively: “We are building your legacy into the system so your expertise scales forever.”
3. Managing the Digital Workforce
Treat agents like employees. Build centralized “agentic control planes” — dashboards for monitoring performance, security, accuracy, and compliance — just as you manage human teams.
4. Role Redesign and Career Pathing
Shift humans from execution to design, orchestration, and interpretation. Create new career paths built around learning agility, systems thinking, and agent management. Some organizations are already assigning dedicated managers to their digital workforce.
(The research notes that the remaining three pillars complete a synchronized transformation covering governance, incentives, and technology architecture — all of which must move in lockstep.)
Also read:
- Forced Piracy: When Paying for Content Becomes Impossible
- YouTube’s Biggest Thumbnail & Title Study Ever: 323,000 Videos, 62.6 Billion Views, and the Brutal Truth About What Actually Gets Clicks
- Streaming Services Continue Their Hunt for Creators: Tubi Launches Creatorverse Incubator with TikTok
The Bottom Line
The AI “last mile” is not blocked by technology. It is blocked by unresolved questions of operating models, governance, and human identity.
Organizations that treat AI as just another productivity tool will continue to collect impressive pilot results and disappointing enterprise outcomes. The winners will be the ones that accept the harder truth: crossing the last mile requires rewriting the organizational operating system itself.
The blueprint exists. The question for every executive is simple — and urgent: Are you willing to redesign the enterprise for an agentic world, or will you remain pilot-rich and transformation-poor?
The research is clear. The last mile is organizational. And it is yours to close.
