Why AI spend is harder to forecast than SaaS seats

In an opinion piece, Connor Linehan, principal consultant at Proxima, argues that the problem with enterprise AI spend is not the pricing model but predictability. Traditional SaaS licensing tracks headcount; AI cost is spread across compute, API calls, storage and data pipelines, and token consumption rises and falls with how people and automated workflows use it.

Because business units can buy AI tools on their own, spend builds up in silos that finance cannot see until the bills arrive. Cost also does not track value: a busy customer chatbot can be cheap per query, while a low-volume internal analytics tool using large context windows can be expensive.

His suggested controls: inventory AI tools across departments and consolidate overlapping contracts; set real-time guardrails such as rate limits, token caps and alerts; restructure contracts so vendors share risk, for example a flat base fee with consumption tied to outcomes; and check whether claimed productivity gains actually show up in revenue or reduced contractor spend.