A framework for governing your AI supply chain, not just the bill. KYI evaluates every inference system across five dimensions, so a board can see what creates value, what carries risk, and what to do about it.
The industry optimises per-token cost. Production inference is where a model’s outputs meet real users, operating costs, and reliability requirements. Just as electricity's impact came not from cheaper lighting but from the industries it enabled, inference's value lies in more valuable, reliable, governable capabilities, not a smaller unit price.
KYI scores every inference use case across performance, economics, integration, strategy, and risk, then rolls them into a single weighted score, a confidence level, and a recommendation. Pick a use case to see its profile.
o10 routes to the cheapest model that clears your quality floor. Evals are what make that floor real. They replay your traffic against every candidate model and score it, so "good enough" is measured, not asserted. The cheapest model that passes is the one o10 routes to.
Per use case, not a global average. A support bot and a code copilot clear at different bars. Set each from a real eval suite.
Before any switch, evals show the cheaper model clears your floor on your own traffic. The savings number ships with its proof.
Models and prompts shift. Continuous evals flag regressions the moment a routed model slips below the floor, and re-route.
Inference is a supply chain: capacity sourced across venues, routed in the path, scored for value and risk, and reported to the board. o10 is the control plane; KYI is the layer that makes it governable and sustainable.
Cost per outcome, the KYI score, the recommendation, and an immutable audit trail. A defensible record the board and the regulator can sign.
Every use case scored across five pillars, against a floor. Sustainable, governable, and tied to value creation rather than unit price.
In the request path: routes every call to the cheapest eval-passing model, holds the budget envelope, and records an immutable per-call ledger.
The quality floor, made measurable. Evals replay traffic against every candidate model, prove equivalence before a switch, and run continuously to catch drift.
o10 watches which prompts run on which models, classifies every call by purpose, and maps your real AI supply chain. Purpose → model → venue. Then it re-sources each purpose to the cheapest model that clears its floor. Toggle to see o10's recommended architecture; click any purpose to trace it.
KYI isn't a one-off engagement or a slide deck. It runs inside the control plane. Every routed call and every eval feeds the score, so the recommendation is current the moment a board asks.
Every call in the path streams cost, latency, policy, and eval scores. Continuously, not sampled after the fact.
→Performance, economics, integration, strategy, and risk update from real evidence as conditions change.
→The composite score, confidence, and recommendation are live. Ready for the board or the regulator on demand.
→A use case that slips below its floor gets auto-rightsized or capped at the control plane. No ticket, no sprint.
↻The composite KYI score maps to a clear recommendation. The line between an investment that defends its economics and one that gets rightsized or capped at the control plane.
Expanded methodology and governance detail. Beyond the interactive scorecard above.
Know Your Inference (KYI) is o10’s framework for reviewing an AI inference system across performance, economics, integration, strategy, and risk. It organizes evidence for a decision; a composite score is not a certification or a guarantee of safe operation.
Score a defined workload and version, not AI adoption in the abstract.
For performance, record task acceptance criteria and evaluation results. For economics, measure fully loaded cost per accepted outcome. For integration, document dependencies, failure handling, and the owner responsible for operation.
For strategy, explain the business outcome and the alternative to the proposed system. For risk, document permissions, data handling, failure consequences, and the controls that have actually been verified.
The framework weights performance and economics at 25% each, integration and strategy at 20% each, and risk at 10%.
Keep the evidence and confidence level beside each component score. An average can conceal a serious weakness: an unresolved access-control failure should not be canceled out by a strong cost score.
Document mandatory release criteria separately from the composite. Record the decision owner, accepted limitations, and the conditions that would trigger a new review.
Repeat the relevant evaluation after material changes to models, prompts, tools, traffic, or data.
Preserve the previous configuration and results so differences can be explained. A routing receipt helps trace a request, but it does not by itself establish that an output is correct or a workflow is compliant.
Use the scorecard as an input to your organization’s operating process. Verify current product controls independently; jurisdiction and residency routing are roadmap capabilities rather than current enforcement guarantees.
o10 framework explanation. Weights are framework design choices, not independently validated predictors of business outcomes.
No. KYI is a framework created by o10 to structure evaluation and governance discussions. It does not independently certify legal compliance, model safety, or business value.
No. Define non-negotiable requirements separately. A weighted score summarizes dimensions; it should not override an unmet permission, quality, or data-handling requirement.
Put KYI in the path on your top use cases. The score, the risks, and the lever, continuously.
See it on your traffic →