Cloud Gave You Years. AI Gives You Months

The cloud took years to roll out. Everyone hated the wait. The data migrations, the hybrid environments, the phased cutovers that dragged across budget cycles.

That wait was doing something nobody put on a slide. While the technology crawled, the organization had time to catch up. People learned the new tools. Processes got reworked. By the time the technology was fully live, the rest of the company had partly caught up to it.

AI took that time away.

The framework here is old. Harold Leavitt published it in 1965 as the "Diamond Model" in a paper called Applied Organization Change in Industry: people, structure, tasks, technology, all interdependent, so a change in one forces changes in the others. Decades later structure and tasks got folded into "process," and the security writer Bruce Schneier made People, Process, Technology a standard way to think about any IT change. The core idea held the whole time: you cannot move one pillar and leave the other two alone.

I watched this play out for twenty years at Microsoft, the last stretch while running data and AI sales for capital markets. Moving an enterprise to the cloud was sold as a technology project, and the technology was genuinely the heavy pillar. It cost real money and took real months, sometimes years. You migrated data, rebuilt applications from the ground up, and stood up hybrid environments that kept the old and new systems running side by side while you retrained people on a new set of tools. It was painful and it was slow. That turned out to matter more than anyone realized.

Here is the part that gets missed. Cloud migrations failed constantly, and they did not fail on the technology.

McKinsey put roughly $100 billion of wasted cloud-migration spend over three years, with the average company running 14 percent over budget and 38 percent of migrations slipping more than a quarter. Gartner found around 83 percent of data-migration projects fail or blow their budget and timeline. Only about 10 percent of cloud transformations captured their full value. McKinsey's own read on why: 70 percent of transformation failures trace to people-related issues, not technology problems.

So the failures were already a people and process story. But the slowness of the technology pillar gave companies a runway to do that work in parallel. The years it took to migrate data and stand up hybrid environments were the same years a company could spend retraining staff and redesigning how work flowed. The technical lift set the pace, and the pace happened to match how fast an organization can actually change, which is slowly.

The companies that used the runway won. The ones that treated it as purely technical lost, and they usually found out when they measured the wrong thing: cloud success judged by IT output, like the number of applications migrated, rather than business outcomes. That detail matters, because AI is about to repeat it.

AI does not need the runway, because the heavy technical build already happened. It rides the hyperscale cloud and the internet that are already in place.

As one analysis put it, AI "is running on top of the cloud," skipping the twenty-year infrastructure build-out and jumping straight to the deployment phase. The hyperscalers carry the capital, the GPUs, the data centers. Enterprises consume it through the same APIs and marketplaces that delivered cloud software. Every employee already has the device. There is no data migration standing between you and turning it on. Companies are now adopting AI faster than they adopted cloud, mobile, or APIs, precisely because it augments existing workflows instead of ripping them out.

This is where I want to be exact, because the easy version of this point is wrong. The technology is not free. Companies are spending heavily to enable it. What changed is the time. The technology pillar that used to take years to stand up now flips on for everyone in a day.

And the work in the other two pillars did not get any smaller. Reimagining a process still takes the same hard thinking. Training a person still takes the same months. Those are people and process decisions, and they are exactly where AI value is made or lost. The runway that used to cover them is gone. The technology now finishes first and the organization is left standing at the start line.

Here is the trap the compression sets. Leadership writes a large check to turn the technology on. The technology turns on immediately. So leadership asks for the return immediately, before any of the slow people and process work has happened.

MIT's NANDA initiative studied this directly. In The GenAI Divide: State of AI in Business 2025, they found 95 percent of enterprise AI pilots delivered no measurable impact on the bottom line, against $30 to $40 billion invested. Only 5 percent reached production with real value. Their conclusion was blunt: "the major barriers are organizational, not technological." They named a "learning gap," and a "shadow AI economy" of employees using consumer tools their employers never sanctioned.

Ethan Mollick has tracked the same gap from the worker's side. In Making AI Work, he wrote that "AI use that boosts individual performance does not naturally translate to improving organizational performance." The personal gains are real. The company just never sees them, because turning private wins into shared ones is the slow work nobody funded.

The 95 percent is not measuring bad technology. It is measuring organizations that were handed an instant technology pillar and asked for results before the people and process pillars had any time to move.

If the technology is the part that turns on overnight, the people and process work is everything that has to happen after, and now it has to happen fast. Five decisions sit inside those two pillars. The cloud era let companies make them slowly, across years of migration. AI asks for them inside the same months the tools went live.

Decide what you are actually rebuilding. "Roll out AI" is not a strategy. Picking two or three processes you intend to reimagine, and being willing to let them get strange before they get better, is. MIT found most AI budgets went to sales and marketing while the largest returns sat in back-office operations. That is a strategy failure before it is anything else: companies aimed the technology at the most visible function instead of the most valuable one.

Decide who owns it. When AI is handed to IT as a procurement and security problem, it stays a tool nobody reorganizes around. This is an executive decision, not a technology one. Ownership belongs with the people who can redesign how the work flows, and the organization has to be willing to change its shape around the new workflow instead of bolting the new workflow onto the old shape.

Decide how people learn, and how that learning travels. This is where the compression cuts deepest. Retraining a workforce used to have the migration's runway underneath it. Now it has months. And the hard part is not the training itself, it is that the gains are hiding. Mollick's secret cyborgs and MIT's shadow AI are the same people, getting private wins they never report. The learning decision is building the loop that surfaces and spreads what they find: shared prompt libraries, workflows people can copy, and real time set aside to experiment.

Decide how people are rewarded for sharing. A learning loop only works if people feed it. When finding a faster way to do the job means more work or a smaller-looking role, the rational move is silence, and most companies built exactly that incentive without meaning to. People will not hand over the thing that makes them look indispensable if the reward for doing it is being made dispensable. Reworking that is a people decision no model can make for you.

Decide how you will measure it, function by function. The cloud-era mistake was a single IT-output number. The AI-era version is a single ROI number, demanded early and read at the top of the house. It measures the wrong shape, because AI lands differently in every function. In support it changes resolution time. In legal it changes how a contract gets reviewed and by whom. In finance it changes how a model gets built. Real measurement runs inside each functional unit, on what actually moved in that unit's work. That is slower than a board-ready percentage, and it is the only version that tells you whether absorption is happening.

I now teach knowledge workers how to orchestrate AI, and the same compression shows up at the individual scale. People arrive having gotten real personal gains in a weekend, and they have no idea how to make those gains travel to a teammate, let alone a department. The tool was instant. The spread is slow, and the spread is the whole point.

That is the enterprise problem in one person. Everyone has the same models now. The companies that pull ahead will be the ones that spent hard and fast on the people and process work the cloud era let them stretch across years, and did it inside the months AI actually gives them.

Written with Claude Code, drawing on Ethan Mollick's research, the MIT GenAI Divide study, and twenty years of watching enterprise technology meet organizations that change at their own speed.

Thanks for reading through.
I’d love to know how you felt about today’s newsletter. This will help me make the newsletter better.