Why Adoption Isn’t Diffusion
A recent review by Alex Imas and Madhav Shukla provides a useful roundup of who is using AI and how. But while the analysis inside the frame is thorough, the frame itself deserves scrutiny.
The authors implicitly assume that AI has passed what I call a Filter 1 test: demonstrating clear, firm-level financial gains at market prices, independent of subsidised pilots or venture-backed ‘free tiers.’1
If we look at the history of transformative technologies, the winners didn’t just diffuse—they pulled themselves into existence by solving a problem so clearly that firms couldn’t afford not to adopt them.
The Myth of Subsidized Productivity
Most studies cited in AI productivity literature measure performance in ‘lab’ conditions: pre-configured tools, directed tasks, and controlled environments. This misses the crucial test of whether the technology pays for itself when the firm has to foot the bill for the infrastructure, the tokens, and the organizational overhead.
Contrast this with two historical giants that cleared Filter 1 almost instantly:
- Containerization: Delivered a 90% cost saving on the first use, alongside a massive reduction in theft.
- Early Electrification: Provided 20% fuel savings and removed the headache of coal, water, and boiler maintenance before factories were even reorganized.
These weren’t ‘General Purpose Technologies’ (GPT) because a whitepaper said so; they were GPTs because they delivered unambiguous first-use gains. They solved the ‘prior question’ of value before anyone worried about the ‘catch-up’ or ‘lock-in’ of diffusion.
The Kenyan ‘Tell’
The most revealing study in the Imas / Shukla review is the field experiment with Kenyan entrepreneurs. It’s the only study where participants chose how and whether to use AI on problems they actually cared about.
The result? An average treatment effect of zero.
The authors suggest that low performers selected harder problems, leading to a ‘catch-up’ failure. But there’s a more grounded interpretation: AI may be excellent at solving a narrow class of well-structured ‘entrepreneurship-as-task’ problems, but it fails on the open-ended, messy, and idiosyncratic problems that define real business.
If the technology only works in narrow, well-scaffolded contexts, the lack of adoption isn’t a ‘diffusion problem.’ It’s a sign that the technology isn’t yet a general-purpose tool.
Narrative vs. Reality
We often forget that General Purpose Technology is a post-facto narrative, not a predictive category. We saw this with nanotechnology and blockchain—enormous amounts of organisational ‘scaffolding’ (consultancies, frameworks, adoption roadmaps) were built to help firms ‘adopt’ technologies that ultimately didn’t solve enough real problems at a sustainable cost, and was simply discarded.
Rather than asking ‘How do we get firms to adopt AI?’, we could ask ‘Are unsubsidized firms, with free choice of tools, renewing and expanding their use at scale?‘
Commercial infrastructure is pulled into existence by demand, not pushed by subsidies.2 You can’t scaffold your way to a transformation if the underlying technology hasn’t cleared Filter 1. Until AI proves it can handle the open-ended ‘unstructured’ reality of the firm without a venture-capital safety net, the adoption gap isn’t a lag—it’s a signal.
- Evans-Greenwood, Peter. “Technologies Don’t Wait.” Substack newsletter. The Puzzle and Its Pieces, February 3, 2026. https://thepuzzleanditspieces.substack.com/p/technologies-dont-wait. ↩︎
- Evans-Greenwood, Peter. “The Electrification Productivity Puzzle.” Substack newsletter. The Puzzle and Its Pieces, January 13, 2026. https://thepuzzleanditspieces.substack.com/p/the-electrification-productivity. ↩︎