Model price is the number everyone quotes and almost never the number that decides whether agentic work pays. The unit a business actually buys is an accepted outcome — an artifact a reviewer signs off on, a decision a human approves. This page lays out the full cost stack behind that unit, how governance shapes it, and what to measure. No savings percentages, no benchmarks — a way to count honestly.
Per-call model price tells you what a request costs. It says nothing about how many requests, retries, reviews, and repairs it takes to produce one output your organization actually accepts.
The honest unit. An accepted artifact or an approved decision — output that cleared quality gates and human review, and got used. Everything on this page divides by that.
Eight lines sit above the divide. Most cost analyses count the first one and ignore the rest — the rest is usually where the money goes.
total cost per accepted outcome =
( model usage + evidence processing + tools & APIs + infrastructure
+ human review + retries & loops + failed runs + downstream rework )
÷ accepted outcomes
The controls that make agents trustworthy are the same controls that shape the cost curve. Each mechanism in the governed-autonomy ladder is designed to attack a specific line in the stack.
Two kinds of control: routing that keeps each task on the cheapest model that clears the quality bar, and execution bounds that keep a workflow from spending indefinitely.
The platform routes each task to the most cost-effective model that meets the quality bar — frontier models for the hardest reasoning, efficient hosted models for everyday work, and local inference where it fits. Routing is configuration, not code.
Routing mechanics are described on the Architecture page.
You do not need a complicated model. Four measurements, taken inside your own deployment, answer whether a governed workflow is paying for itself.
| Measure | What it tells you |
|---|---|
| Cost per run | The baseline unit economics of the workflow — every run, accepted or not. |
| Cost per accepted artifact | The real production cost of a deliverable that cleared quality gates and review. |
| Cost per approved decision | The full cost of getting a human to a confident yes or no — evidence, evaluation, and review included. |
| Baseline comparison | The same outcome priced against the pre-automation process — the only comparison that answers whether it was worth it. |
Measured within your deployment, against your acceptance criteria. The gap between cost per run and cost per accepted artifact is your failure-and-retry overhead — the number governance is designed to shrink. Acceptance is defined by the evaluation gate — see evaluation & reliability.
In a walkthrough we map one of your workflows onto this cost stack and show where the governance mechanisms sit — so you can measure it yourself.
Request a technical walkthrough → See the architecture →