What cloud’s lost decade teaches us about adopting AI

In 2006, Amazon launched a service that let you rent a server by the hour. The promise was true. Two decades later the cloud is the substrate of the modern economy — and if you ask the CFO of almost any enterprise that migrated between 2010 and 2020, you will hear the same story: longer than promised, costlier than budgeted, benefits years later than the pitch.

This is not a story about the cloud being a bad bet. It was the right bet. It is about how organizations adopt transformative technology, and about a mistake an entire generation of companies made in near-unison. We are at the start of that same curve with artificial intelligence.

The failure was never technical. It was a learning failure. Cloud demanded that organizations learn to manage elastic, metered infrastructure. AI demands that they learn to govern algorithmic reasoning, pay for tokens instead of servers, ground agents in a shared definition of “revenue,” and rebuild work around humans and agents. Skip that learning and you do not merely overpay. You scale ambiguity, lock-in, and regulatory exposure at the speed of an API.

Procurement is not a capability

The cloud was sold as a procurement decision: stop buying servers, start renting them, book the savings. A procurement program has a completion date. A capability shift has a competence date. Only one of those was ever on the steering committee slide.

Companies moved applications designed for fixed, owned capacity into a world of variable, metered capacity — then were shocked when the bill behaved like a metered utility. The technology arrived on schedule. The understanding of how to use it lagged by a decade. That distance is the bill.

The overspend was a thousand small, rational mistakes: lift-and-shift without re-architecture, fleets left running at peak, no cost attribution, reserved capacity bought to paper over waste. FinOps exists because an industry learned it in production. The tools were available on day one. The knowledge of how to wield them was not.

Every one of those mistakes was made against a predictable meter. A virtual machine costs the same per hour whether it serves one request or ten thousand. Cloud FinOps is an inventory problem. AI does not bill you for things. It bills you for behavior.

Tokens, not instances

Models charge per token, and several times more for what they generate than for what you send. The bill is driven by prompt length, retrieved context, verbosity, and which model you pointed at the problem.

A simple classification job at modest volume cost about $90 a month on GPT-3.5 Turbo in 2023 and about $2,700 on GPT-4 — thirty times, selected by one string in a config file. Against August 2026 list prices, nano to frontier is eighty-two times. Cheap tokens did not remove the decision. They made it easier to stop thinking about.

Model routing is the new right-sizing. Classification and extraction are jobs a small model does at full quality. Frontier models exist for genuine multi-step reasoning. The cloud-era instinct was to provision for the hardest hour. The AI-era version is to provision for the hardest prompt and pay for it on every prompt.

The same curve, steeper

AI is being sold the way the cloud was sold: buy the licenses, plug in the API, book the productivity. Lift-and-shift becomes bolting a chatbot onto the old workflow. Provision-for-peak becomes a frontier model for every task. No cost attribution becomes no evaluation harness. Commitments on top of waste become an enterprise-wide rollout before a single validated use case.

The curve is steeper. Cloud misuse mostly cost money. AI misuse costs money and trust. Shadow AI takes thirty seconds and a browser tab. Lock-in is arriving at the agent platform, not the model. Sovereignty is no longer a residency clause you sign later — inference is where sensitive data flows, so the constraint has to be decided at design time.

A frontier model arrives knowing an enormous amount about the world and nothing about your business. Ask an agent for “Q3 enterprise revenue” without a governed definition and it guesses: syntactically valid, confidently wrong. Count the places “active customer” is defined. If the answer is greater than one — and it is always greater than one — that work must precede the agent rollout.

Your people are ready. Your company is not.

McKinsey surveyed 750 employees and leaders in early 2026. Seventy percent of employees feel personally prepared to use AI. Only 27 percent of leaders believe their organizations are ready to change around it. Organizational readiness accounts for 48 percent of the difference between leaders who capture enterprise value and those who do not; personal readiness, 25 percent. Licenses move the smaller number.

Leaders are 5.3 times more likely to report enterprise value when workflows have been redesigned than when they were left unchanged — 32 percent against 6 percent. Not a better model. Not more seats. The workflow.

Independent analyses put median time to cash-flow positive at 18 to 24 months. Deloitte found most executives report satisfactory ROI in two to four years; only 6 percent saw payback inside a year. Meanwhile 80 percent of people using AI say it improved their individual productivity, and organizations attributing at least 5 percent of EBIT to AI sit at 6 percent. Individual productivity is not the bottleneck. The organization is.

Spreading thirty pilots does not reduce risk. It guarantees that no single one gets the talent, data, or executive attention required to change anything. Pick a few end-to-end domains and rebuild those. One domain done properly produces a number at eighteen months. Thirty pilots produce thirty anecdotes and no mandate.

Front-load the learning

The lesson from the cloud is not “go slower.” It is “front-load the learning so you can go faster safely.”

1. 01Invest in structured learning before scale — including executives. Value capture is 3.9 times likelier where leadership is genuinely AI-fluent.
2. 02Pick a few domains and take them end to end.
3. 03Redesign the process; do not decorate it.
4. 04Build attribution, semantics, and guardrails in from the start.
5. 05Preserve your exit. Keep one workload on an alternative runtime, and know your sovereignty obligations before the architecture hardens.

The trough is real — roughly eighteen to twenty-four months — and the organizations that come out the other side will be the ones whose boards understood the shape of the curve before they signed. Treat learning as the first line item. Point everything at two or three domains that matter. Deploy on evidence, govern the meaning of your data, and know what every token is buying. Do that and you will not merely adopt AI. You will compound it.