Seventy years in five waves: the evolution to AI transformation

AI transformation is the fifth wave of a seventy-year movement, and each wave moved the unit of change further from a single task towards the whole operating model. Two lessons recur across all five: the technology arrives years before the value, and the value arrives only when work is reorganised around it.

Wave 1: rules and expert systems (1956 to the 1990s)

The field took its name at the 1956 Dartmouth workshop. Its first commercial form was the expert system, which codified specialist judgement as rules. These systems worked in narrow domains but were brittle, costly to maintain and unable to learn, and two funding collapses, the "AI winters", followed overpromising. The unit of change was a single expert task.

Wave 2: process automation and enterprise systems (1990s to 2010s)

Enterprise resource planning, workflow engines and, later, robotic process automation digitised and automated whole processes with deterministic rules. This wave produced the clearest warning for today. In 1987 Robert Solow observed that computers were visible "everywhere but in the productivity statistics". Paul David (1990) answered by comparing computers with the electric dynamo, whose productivity gains arrived decades after its invention, once factories were physically redesigned around distributed power rather than a central shaft. The unit of change was the process.

Wave 3: machine learning and predictive analytics (2012 to 2022)

Deep learning's breakthrough in image recognition in 2012, followed by cheap cloud computing and large datasets, made prediction a commodity. Organisations embedded models in specific decisions such as fraud, credit, demand and churn, and built data platforms and model-operations practices to run them. The unit of change was the decision, and the binding constraint became data quality and the skills to use it.

Wave 4: generative AI (2022 to 2024)

The transformer architecture (Vaswani et al., 2017) made models that could read and write language at a general level. ChatGPT, released in November 2022, reached an estimated 100 million users in two months (Reuters). For the first time AI touched knowledge work directly, and organisational use rose from 20% in 2017 to 78% in 2024 and 88% in 2025 (IDM summary of McKinsey). The unit of change became the knowledge worker's task, and most organisations met it with adoption programs: licences, training and pilots.

Wave 5: agentic AI (2025 onward)

Agentic systems plan, use tools and act across multi-step workflows with limited supervision. In 2025, 62% of organisations were at least experimenting with agents and 23% were scaling them somewhere in the enterprise (IDM summary of McKinsey). Because agents take actions rather than produce drafts, the unit of change is the workflow and the role, and governance moves from reviewing outputs to setting decision rights.

The pattern

Each generation of technology was adopted quickly and paid off slowly, and the payoff came from complementary changes in process, skills and organisation. That is the economic logic of the productivity J-curve (Brynjolfsson, Rock and Syverson, 2021), and it is why AI transformation, not AI adoption, is the object worth managing.

Sources

• Brynjolfsson, E., Rock, D. and Syverson, C. (2021). The productivity J-curve. American Economic Journal: Macroeconomics.
• David, P. (1990). The dynamo and the computer. American Economic Review.
• Vaswani, A. et al. (2017). Attention is all you need. arXiv 1706.03762.
• Reuters: ChatGPT user growth
• IDM: McKinsey State of AI 2025 (secondary report of the survey).