AI and the Great Divergence

The White House Council of Economic Advisers published a short report, “Artificial Intelligence and the Great Divergence” (January 21, 2026). I read it and wanted to share how I see it.

The core message is straightforward: AI is increasingly compared to the Industrial Revolution, and if the impact is truly industrial scale, the leaders in investment, capability, and adoption speed can lock in the gap. What I liked is that the report leans on measurable signals we can observe now, not speculation about some distant future.

Why the Industrial Revolution analogy works, if you think in stages

The Industrial Revolution was not “one innovation.” It unfolded in stages: first the foundation (energy and infrastructure), then scaling (factories, standards, processes), then maturity (mass production and logistics). The gap widened because some moved through the stages faster, not because they used the word “revolution” louder.

With AI, I see the same structure

Stage 1 today is the foundation: compute infrastructure, data, security, access control, observability, cost discipline, and most importantly, the ability to embed AI into operations, not slide decks.

Stage 2 is scaling: the value does not come from a “chatbot for show,” it comes from redesigning processes, KPIs, and quality control.

Stage 3 is reshaping the product and the economics: new unit costs, new speed, new service models.

Why I’m cautious with “revolution” narratives

Over the last years we’ve heard many “new industrial revolutions”: Blockchain, Smart Contracts, Web3, NFT, the Metaverse, IoT. In every wave the technology was real, but the narrative almost always ran ahead of market maturity and a company’s ability to turn technology into operational normal.

IoT was the most vivid example for me. When I worked at Microsoft, responsible for strategy and technology, the ecosystem pushed “IoT = the new industrial revolution” very aggressively: calls, events, pressure not to fall behind, glossy success stories, and the feeling that if you do not jump in now, you will be late forever.

The technology was real. But the gap between the narrative and real time-to-value for most companies was massive.

I’ve seen a very concrete lesson. One strong, mature executive from a Microsoft partner company believed the IoT story so deeply that he stepped away from the core business and built a separate company with fancy name and marketing, focused purely on IoT. And it did not take long to see that the story was far from reality, because the market, customer readiness, and monetization speed were not where the presentations suggested. To be fair, this happened in Georgia (Country), where the market lags behind developed markets. But that makes the point sharper, not softer. If even mature leaders can buy a “revolution” narrative where the foundation and demand are not ready, how many such decisions get made globally, especially across peripheral markets?

A “real IoT” example, without illusions: In the elevator industry, predictive maintenance has clean math: downtime is expensive, telemetry is natural, and the installed base is huge. For example, thyssenkrupp collected data from elevators and escalators, sent it to Microsoft Azure, analyzed it, and helped prevent failures and reduce downtime.

At the same time, IoT was not a gimmick. It delivers real value where the economics are clear and operations are mature.

The “revolution” there happened not because the word sounds good, but because three things aligned: measurable downtime, a clear data model, and the operational ability to act on signals, not calendar schedules.

Where the line is between AI as an industrial cycle and AI as the next hype

It is very easy to turn AI into theatre: give everyone “a chat,” build a showcase RAG, launch a couple of “agents,” publish a strategy. This repeats the IoT pattern: lots of activity, little outcome. Industrial logic is different: foundation first, then scale, then a true model shift.

Since I use the term RAG, means Retrieval-Augmented Generation. In simple terms: you do not “train the model” on company documents. Instead, you retrieve the relevant pieces from your internal knowledge base at query time, and the model answers based on those snippets. It can be faster and cheaper than training, but it requires strong data discipline and security: what can be retrieved, by whom, what is logged, and how leakage is prevented.

If you do not want to lose the “divergence” race, the question is not “do we have an AI strategy.” The questions are:

how fast are you building the foundation (data, security, cost, adoption into paths);
how fast do you turn experiments into operational normal;
how fast do you change processes and product economics, not just the user interface.

Speed through the stages, not the volume of the narrative, will determine who gets the compounding advantage and who will be catching up later and more expensively.

How to tell real AI adoption from a showcase

First: there is measurable impact. Not “innovation.” A KPI with a baseline and an owner: cost reduction, revenue lift, cycle-time reduction, downtime reduction.

Second: AI is embedded in the process. If you remove the model, the process visibly degrades and the metrics show it.

Third: unit economics are controlled. Clear cost per request or per case, budgets, limits.

Fourth: data and access rights are in order. Known sources, quality, accountability.

Fifth, and critical: security and data classification. If a company has no data classification, people end up feeding “everything” into AI. Then a simple question can surface what the model should never reveal: internal confidential information or material that is classified. Real adoption starts with basics: what is allowed into the model, what is forbidden, who approves, what is logged, and how leakage via answers is prevented.

Conclusion

AI may indeed become an industrial cycle, not just another trend. That is exactly why balance matters: speed without discipline becomes risk and leaks; security without speed becomes paralysis and loss. The goal is not “to do AI.” The goal is to industrialize it properly: build the foundation, scale only what proves measurable value, and protect data as strictly as money and reputation. Then AI stops being a showcase and becomes a competitive advantage.

Originally published on X. Blog date: 23 January 2026.