FeedDiscussion
Well said. “The deployment decision preceded the cost model” is probably the line that stood out most to me.
We're seeing a familiar pattern from the early cloud era playing out with AI: teams optimize for capability first, then try to understand consumption after the system is already running at scale.
The interesting part is that the solution isn't necessarily using cheaper models. It's building the architecture to match the right level of intelligence to the right task.
Routing, caching, cost attribution, governance and observability need to become part of the platform—not decisions every engineer has to make independently.
AI consumption isn't the problem. Unmanaged consumption is
The organizations that build that orchestration layer early will have a very different cost curve as agentic workloads scale.
Worth noting upfront: the author's own consultancy is promoted at the end as the solution to exactly this problem, doesn't make the numbers wrong, but worth factoring in.
The "deployment decision preceded the cost model" point is the sharpest one, piloting on single queries then shipping agentic workflows that use 5-30x more tokens per task means the ROI math was validated against a completely different consumption pattern than what actually shipped. Not careless, just a measurement-timing failure.
The cloud-2018 parallel is apt but incomplete, cloud waste was caught by visibility alone (idle resources). Token waste needs an actual routing decision made correctly upfront; a dashboard tells you where money went, it doesn't route the next query to a cheaper model on its own.