Juan Piaggio · 2026-08-10 · 7 min read · ai · cost · operations
For most of the last decade, the expensive part of AI was building the model. Training was a capital event: large, planned, and finite. Running the thing afterwards was a rounding error.
That relationship has inverted. Organizations now spend more running models than creating them, and for teams like ours — who never trained anything and never will — the flip is total. Our AI cost is entirely inference, it recurs every day, and it grows with usage rather than ambition. That is not a bigger version of the old problem. It is a different category of problem.
A training run is a decision you make once, with a number you approve in advance. An agentic workflow is a cost per event, forever, and the volume is set by your users rather than your roadmap.
The distinction has teeth because of how agentic work consumes tokens. A chatbot turn is one call. A multi-step agent that reads a ticket, checks for duplicates, evaluates a policy, and drafts a transition is several calls, each carrying context, some of it re-read at every step. Add a specialist mesh and the multiplier compounds — three agents sharing context can cost more per item than the monolith they replaced.
None of that shows up as a spike. It arrives as a slope: fine in the pilot, uncomfortable at ten times the volume, alarming at a hundred. And because usage growth is the thing you were hoping for, the cost curve is correlated with success. You cannot budget your way out of it annually; you have to instrument it continuously.
Total monthly spend is the wrong metric — it moves with volume, so it tells you nothing about whether the system is healthy. The number worth watching is cost per accepted decision: what you paid, divided by the proposals a human actually approved.
That denominator does real work. It means an agent that generates a hundred cheap proposals nobody accepts scores worse than one that generates ten good ones, which is exactly the ranking you want. It makes waste visible, where waste is any inference that produced something a human threw away.
Once you can see that number, the levers are obvious in a way they were not before:
Operating costs need operating controls, and the two that matter are the ones that bound the worst case rather than optimizing the average.
A spend cap per tenant, enforced daily. Not an alert — a cap. Alerts inform you after the money is gone. In Meshworq, AI spend is tracked per workspace against a daily ceiling, so a single tenant's runaway usage is bounded by design instead of by someone noticing a dashboard.
A circuit breaker per agent. The expensive failure in agentic systems is rarely one costly call. It is a loop: an agent retries, or two agents hand work back and forth, and a workflow that should have cost cents costs hundreds before anyone looks. A per-agent breaker that trips on repeated failures or an abnormal call rate turns an unbounded incident into a bounded one. Autonomy without a breaker is just an unbounded loop with good intentions.
A cost cap is not pessimism about your agents. It is what lets you give them more autonomy without flinching.
Both controls have a second benefit that is easy to miss: they make the cost of a proposed feature knowable in advance. When every agent runs under a ceiling, "what happens if this goes wrong" has a number, and that number is what makes the conversation with finance short.
Token counts mean nothing outside engineering. Translating them into a stable internal unit — we use credits, at a fixed ratio to currency — turns AI spend into something a non-engineer can reason about, forecast, and compare across teams.
The framing that lands is per-outcome, not per-month: this workflow costs roughly this much per ticket triaged, against roughly this much of a person's time it replaces. That comparison is defensible in both directions. Sometimes it says keep going. Sometimes it says this particular agent is not worth running, which is a genuinely useful thing to learn early rather than at renewal.
AI spending has moved from a finite project cost to a recurring operating cost that scales with your success, and agentic workflows multiply it by making several calls where a chatbot made one. Track cost per accepted decision rather than total spend, use the confidence floor and model tiering as economic levers, and put a per-tenant daily cap and a per-agent circuit breaker in place before you need them. Then express it in units your finance team can hold. Inference is a line item now. Treat it like one, and it stops being the reason you hesitate to expand.