The Forty-Year Problem Harvard Is Trying to Compress into Ten

When Microsoft gave Copilot to its 62,000-person sales organization, daily active usage peaked at 22.7 percent and fell to 5.1 percent in just over a month, despite executive enthusiasm and an extensive internal campaign. That collapse, documented in a Harvard Business School case, is the reason HBS now runs a four-day executive program called Generative AI Strategy and Execution. The program's premise is that history's most reliable lesson about general-purpose technology is about to be tested on a compressed timeline, and most companies are failing the test the same way they failed it the last time.

The last time was electricity. Edison's Pearl Street station opened in 1882, yet electric motors did not pass half of American factory drive power until around 1920, and the productivity payoff only arrived in the 1920s. The economist Paul David's explanation became a classic: firms swapped steam engines for electric motors and kept the old factory layout, so little improved until they redesigned the work itself. The technology was necessary; the redesign was decisive.

Artificial intelligence is rerunning the experiment with one variable changed: diffusion speed. Within two years of ChatGPT's launch, 39 percent of American working-age adults had used generative AI, according to research by Alexander Bick, Adam Blandin, and David Deming published through the St. Louis Fed. The internet stood at half that share at the same age; the PC took three years to reach it. Goldman Sachs projects a 7 percent lift to global GDP over a ten-year horizon. The forty years electricity needed were mostly spent waiting for the technology to spread and for managers to redesign work around it. The first wait is over. Only the second remains, which is why the plausible window is now a decade, and why the sorting will be decided by management practice rather than model quality.

The program, chaired by Professors Rajiv Lal and Suraj Srinivasan, is built to close what the faculty call the AI strategy gap, the distance between what boards announce about AI and what organizations execute. It runs on the case method: participants prepare each case individually, test their thinking in small morning discussion groups, then argue the decision in a full classroom. The four-day arc follows the journey most companies are on. It opens with opportunity, through Coursera's early moves in generative AI, then turns to disruption and the agentic future: Salesforce building Agentforce, and Adobe deciding whether generative AI is an opportunity or a threat to its core business.

The center of the program is execution. The Microsoft case supplies the cautionary tale; its resolution supplies the doctrine. Usage recovered only when Copilot was embedded into the daily workflow and one use case, meeting recap, gave salespeople a reason to return each morning, sustained by constant enablement that could never pause without usage slipping. Unilever's case examines the talent and capabilities an AI-ready organization requires. Gamma, which reached $50 million in recurring revenue profitably with 30 employees, shows what an AI-native operating model looks like when there is no legacy workflow to defend. Sessions on workflow redesign, commerce in the age of AI through Criteo, Harvey's path from legal copilot to core workflow, agentic AI twins built on the WorkFabric case, and Raffaella Sadun's session on the future of work complete the arc before a closing discussion of governance. Participants also build with AI directly, in hands-on agent sessions and a workshop redesigning one of their own workflows.

Three principles run through the curriculum. Business strategy comes first and AI second: Unilever's global digital director states in the case that generative AI is an enabler, not the strategy, and programs that begin with a business problem outperform programs that begin with "we need an AI strategy." Adoption is change management, not deployment: what sustains usage is senior sponsorship that stays visible, use cases tailored to specific roles, and enablement that never stops. And uncertainty is managed by experiment, not conviction: the program teaches the discipline GitHub and Google applied when they ran randomized controlled trials on AI coding assistants, with control groups, real metrics, and planning horizons of 6 to 12 months.

The honest counterargument is that the revolution is not yet visible in the data. The same researchers who measured AI's record adoption estimate its current effect on US labor productivity at 0.1 to 0.9 percent. But this is precisely what the electricity precedent predicts. In 1900, eighteen years after Pearl Street, electrification's measured effect was also negligible, because the redesign had not happened. Adoption without redesign produced the first productivity paradox; it is producing the second one now. The gap between AI's availability and AI's payoff is not a mystery to be explained. It is a queue of management decisions not yet taken.

That is the wager behind the Harvard program: the companies that treat AI as workflow redesign, run honest experiments, and put accountability with business owners rather than technology functions will compress their own forty years into a competitive weapon. On the historical evidence, the firms that learned the electricity lesson early did not merely adopt a technology. They took the market from the firms that did not. This time the sorting is scheduled for a decade, not a generation.

Sources: Paul A. David, "The Dynamo and the Computer," American Economic Review, 1990; Bick, Blandin and Deming, "The Rapid Adoption of Generative AI," NBER Working Paper 32966 and Federal Reserve Bank of St. Louis; Goldman Sachs Research, 2023; HBS cases "Microsoft Customer and Partner Solutions: The Deployment of Copilot (A)" (626-065) and "Gamma: Slides in the Blink of AI" (826-001); HBS Executive Education, Generative AI Strategy and Execution program materials, June 2026.