Working definition: improve the total efficiency of recurring AI work while preserving the agreed result and verification standard. Token count matters, but it is one measure among context, cost, time, waiting, human intervention and accepted output.
View the full infographic → · View this article
Optimise the workflow, not a single number
A smaller request can cost more overall if it needs repeated calls, more review or a recovery run. A fast reply is not useful until it passes the acceptance check. Compare equivalent tasks over a declared window and retain failures in the denominator.
What to measure
- Provider and exact model, where available.
- Input, cached-input and output usage using provider definitions.
- Request count, retries and context-boundary changes.
- Machine active time, provider waits and human handling separately.
- Verified jobs and their quality/acceptance criteria.
When an item is unavailable, label it UNKNOWN. When a figure is calculated from entered assumptions, label it CALCULATED or MODELLED and show formula and source date. Never convert a missing count into zero just to complete a ratio.
A practical cycle
- Define one task and its acceptance test.
- Measure the request and workflow baseline.
- Change one factor at a time: context, tool choice, model route or retry policy.
- Verify that the result still passes.
- Compare cost and total system time over a suitable sample.
- Publish setup, failures and limitations with the conclusion.
The methodology explains how observations are scoped. The tools configure rules and calculate scenarios, but do not create workload evidence for you.