Adoption Patterns and What Usage Data Shows

Because exposure measures technical possibility while usage reflects where capability, verification cost and workflow fit are jointly favourable. Categories rated highly exposed but rarely used are typically those where verification is expensive, the work is embedded in systems the assistant cannot see, or process intervenes.

The verification asymmetry. Code can be checked mechanically, so an error costs little to catch and a correct suggestion pays off immediately. Writing is prominent for a related reason: recognition is cheaper than generation and the output is directly editable.

Because augmentation changes how a task is done while automation changes who does it, so the two have different employment implications at the same level of adoption. Published analyses find a majority of usage still augmentative, with the automation share growing as agentic capability improves.

It mixes personal and educational use with occupational use; it reflects one provider's user base, pricing and capabilities rather than the market; and it misses enterprise deployment through APIs and internal tools entirely, which is a different distribution and much harder to observe.

It establishes where systems are applied, at scale and without self-report bias, which is genuinely valuable. It cannot establish effect: a category with heavy usage and no productivity benefit is entirely possible and looks identical in the data to one with a large benefit.

Because exposure measures technical possibility while usage reflects where capability, verification cost and workflow fit are jointly favourable. Categories rated highly exposed but rarely used are typically those where verification is expensive, the work is embedded in systems the assistant cannot see, or process intervenes.

The verification asymmetry. Code can be checked mechanically, so an error costs little to catch and a correct suggestion pays off immediately. Writing is prominent for a related reason: recognition is cheaper than generation and the output is directly editable.

Because augmentation changes how a task is done while automation changes who does it, so the two have different employment implications at the same level of adoption. Published analyses find a majority of usage still augmentative, with the automation share growing as agentic capability improves.

It mixes personal and educational use with occupational use; it reflects one provider's user base, pricing and capabilities rather than the market; and it misses enterprise deployment through APIs and internal tools entirely, which is a different distribution and much harder to observe.

It establishes where systems are applied, at scale and without self-report bias, which is genuinely valuable. It cannot establish effect: a category with heavy usage and no productivity benefit is entirely possible and looks identical in the data to one with a large benefit.