Three layers of AI diffusion in organisations
A recent NBER (U.S. National Bureau of Economic Research) working paper maps how AI diffuses through organisations across three distinct layers: firm-wide adoption, deployment into specific business functions, and use by workers in their daily tasks.
The research draws on nationally representative US Census data and identifies patterns that are not visible when adoption is measured as a single data point.
Three layers of AI diffusion
The first layer is whether an organisation uses AI at all. The second is which business functions it reaches, such as sales, finance or IT. The third is whether workers are actually using AI in their day-to-day tasks and whether it is changing how work gets done.
Worker-level task use does not follow automatically from firm-level adoption or functional deployment.
In 36% of organisations where workers are using AI, there is no formal firm-level adoption on record. In 19% of firms that have formally adopted AI, there is no evidence workers are using it at all. The direction of diffusion also varies: some AI use spreads top-down from organisational decisions, some bottom-up from individual workers using tools independently of any mandate.
Five profiles of AI-using organisations
The research clusters AI-using firms into five profiles based on how broadly AI is deployed across business functions:
- Minimalist adopters (37%): AI present but narrow — one or two functions, no clear strategy.
- Marketing specialists (31%): AI concentrated in sales and outreach, limited internal reach.
- Administrative integrators (15%): AI in finance, HR, and management — operationally useful, strategically limited.
- Technical strategists (12%): AI in R&D, IT, and strategy — high capability, limited frontline reach.
- Comprehensive adopters (4%): AI deployed broadly across all functions.
Examining the third layer
Most organisations aren't measuring employee level task use alongside headline AI adoption figures.
To better understand how AI is being use on tasks, and if its really adding value, start with these four questions:
- In the functions where AI is deployed, can you name the specific tasks workers now do differently?
- Is that output measurably better, or only faster to produce
- Where use is low, is it a capability gap, a workflow gap, a process gap or a trust gap?
- Or is low use the right answer, because the task is better done by a person?