A global investment bank launched more than 250 projects linked to language models. A food giant deployed AI in 185 countries. A fashion company automated more than 18,000 financial processes. The result? Almost no real transformation of the business. It sounds absurd, but this is exactly what senior executives described at a closed-door summit at Harvard Business School.
The D³ (Digital Data Design) research institute and Microsoft jointly convened representatives from dozens of global companies in healthcare, banking, and industry. The goal was simple: to find out why AI transformation in companies had reached a dead end. And the answer that emerged is surprisingly human. The problem is not the technology. The problem is us.
What is the “last mile”?
The term is borrowed from logistics. The last mile is the most expensive, most complicated stage of delivering a shipment—the stretch from the warehouse to the customer’s door. In AI transformation, it is the moment when technical capability collides with organizational reality. An AI agent can draft a complex contract in seconds. Then that contract waits in a queue for two weeks for a manual legal review because no one has redesigned the internal processes. The bottleneck has not ceased to exist. It has merely moved.
Harvard researchers identified seven specific points of friction blocking this last mile:
- A proliferation of pilots without results. Companies are literally “pilot-rich but transformation-poor.” Local successes exist, but no one knows how to turn them into a genuine operating standard across the entire organization.
- Productivity that disappears without a trace. More than 99% of employees at one global payment network actively use AI assistants. One industrial manufacturer recorded double-digit productivity gains among thousands of engineers. But where does that show up in the accounts? The time saved is quietly swallowed up by more meetings and unnecessary emails. No one has deliberately redirected it toward more valuable work.
- Legacy process debt. AI is an excellent diagnostic tool. But what does it uncover? Years of accumulated, broken processes. At one health insurance company, AI surfaced discrepancies faster than the company could resolve them. One consulting firm discovered that the same process was carried out in dozens of different ways depending on which country you happened to be in.
- Knowledge trapped in people’s heads. This may be the most sensitive point. Experienced employees carry years of accumulated know-how in their heads, with no written record of it anywhere. And why would they share it? Because that very know-how gives them status and makes them indispensable. This is not about retraining; it is about a crisis of professional identity. AI asks people to externalize their judgment and encode it into systems. For many, that is an existential challenge.
- Governance and oversight that cannot keep up. Traditional control mechanisms work for isolated cases. But what if you have 100 AI agents acting autonomously and coordinating with one another? One major bank admitted that its accountability structure was breaking down. Who is responsible for a mistake made by an agent? These are questions the IT department cannot answer. They are more of a topic for PO.
- Technological patchwork. Most companies run AI across several cloud platforms at once. Connecting agents in SAP, Microsoft, and Google environments so that they communicate reliably with one another? It took one fashion giant months. And by the time it was finished, new models had arrived, and the team was tempted to start over.
- The efficiency trap. Companies originally marketed AI internally as a cost-cutting tool. A new form of outsourcing. The result? Middle management began to resist, and leadership’s ambitions narrowed. An analytics firm warned at the summit: an obsession with efficiency can destroy precisely those human capabilities, such as judgment and storytelling, that create real value.
How can it be solved?
Companies that have overcome these obstacles share several common approaches.
Instead of grafting AI onto old processes, they started from scratch. They asked: “How would we design this process today if we knew what AI could do?” They mapped where an agent could work independently and where a human had to make decisions. They began capturing tribal knowledge systematically. They did not ask experienced experts to hand over their know-how, but to build a legacy. They positioned it as an opportunity to encode their unique judgment into a system that would free them from routine work and allow them to focus on truly complex challenges.
They stopped viewing AI agents as software and began managing them like digital workers. With their own performance dashboards, security permissions, and clear accountability. And what about people? Their roles are shifting from execution to orchestration and interpretation. New job descriptions seek the ability to learn and deep domain expertise, not just technical skills.
What does this mean?
The Harvard report shows that if your company is struggling with AI transformation, you are not behind because of poor technology or a lack of data. You are behind because of things that can be changed: processes, structures, culture, and leadership’s willingness to genuinely rebuild the company. The technology is ready. The question is whether you are ready. The companies that understand this first and dare to redesign not only their processes but also how they view their people’s roles will gain a lead that will be difficult to overcome.
