The prioritized roadmap for improving Token Yield. Every item maps to a lever, carries its evidence, and names an owner. Work, not suggestions.
Escalation prompts average 4,100 tokens where ~800 would carry the same answer; agents paste full ticket history instead of retrieving the relevant slice.
RonaSwap pasted history for retrieval on the escalation template. Cortex estimates the impact from 214 observed sessions; no workflow change for agents.
Six people across Sales and Support asked variations of the same onboarding question 142 times this month. Each time, the AI re-derived the answer from scratch.
needs ownerA Knowledge Object has been drafted from the highest-rated completion. Approving it routes the answer ahead of model calls: the question gets answered once, then remembered.
Contract-review triage runs a simple labeling step on the most expensive model. A cheaper path matched its output on 96% of sampled cases.
ShaiTrial the cheaper path on the classification step only. The label step is 31% of this Work Unit's token cost; reviewers see no difference.
The incident agent retries a failing tool call up to six times per run. The retries burned $3.1k this period and changed nothing.
needs ownerValidate the tool schema before the call instead of after. Failures began the day the schema changed (May 18). This is a harness fix, not a model problem.
Policy answers retrieve from a handbook that was deprecated 14 months ago. 23% of policy Work Units end with a human correcting the answer.
needs ownerPoint retrieval at the current handbook. Correction rate should fall back to the 8% org baseline.
Incident triage follows the same seven steps across 41 Work Units. The first four steps are deterministic; they don't need a model at all.
MichaelA runbook draft was generated from the five cleanest traces. Approving it routes steps 1–4 without a model call and keeps the model for the judgment calls.
Contract review routes 88% of calls to frontier models, well above the 62% org average, without measured evidence the quality requires it.
needs ownerRun a paired-quality trial before changing anything. This is a policy review candidate, not an immediate switch.
A schema update doubled the tool-definition payload to 4.4k tokens per call. Every one of 203 runs this period paid that overhead.
needs ownerRemove unused tool definitions from the harness manifest. Payload measured directly; the yield impact is modeled.
The dashboard is currently optimized for screens wider than 1024px. Please open it on a larger display to access the full experience.
The dashboard is currently optimized for screens wider than 1024px. Please open it on a larger display to access the full experience.
