Anthropic’s labour-market research tracks where large language models are actually being used in work, versus where theory says they should be useful. The result isn’t what the hype cycle promises.
The observed gap
In the sectors the research covers, the average job has a theoretical suitability score for AI — a measure of how much of the work could, in principle, be done by a current model. Observed usage is then measured against that score. For most sectors, observed usage trails suitability by a noticeable margin.
For Architecture & Engineering, the trail turns into a chasm. The sector scores high on suitability and low on usage. In absolute terms, it’s one of the largest gaps in the dataset.
Callout · verify before publishing
This post cites adoption-gap statistics from Anthropic’s research. Exact figures are pulled in from the primary source — not recalled from memory — before this goes live.
Reading the sector name as “architects designing buildings” misses half the point. The category covers the full construction stack: civil, structural, M&E, electrical, specialty contractors. Anything where work has to be described once, costed, executed, tested, handed over, and maintained.
Why the gap exists
The received wisdom is that construction is “resistant to change.” That’s lazy. The real answer is that the current generation of tools digitised each stage but never connected them. Adding an AI assistant to a form still leaves you filling out the same form.