A glossary of AI cost engineering terms
The vocabulary Class1 uses to estimate and gate AI cost before merge, in plain language. Each term links to its own anchor so you can cite a definition directly.
P50 (median monthly delta)The monthly AI-cost change a pull request is equally likely to beat or exceed. Class1 reports it as the central estimate, never as a guarantee.
P90 (tail monthly delta)The monthly cost change exceeded in only about one month in ten. Class1 gates on P90, not the average, because AI cost risk lives in the tail.
P95A stricter tail percentile than P90, exceeded roughly one month in twenty; used when a budget owner wants a more conservative ceiling.
Estimate class (AACE-adapted)A maturity label from Class 5 (order-of-magnitude, new install) to Class 1 (validated), adapted from AACE. It rises only as real estimate-to-actual pairs accumulate, so confidence cannot outrun evidence.
Common random numbers (CRN)A variance-reduction technique: the baseline and the proposed change are simulated on the same random draws, so an unchanged workload has a delta of exactly zero and only the change moves the number.
Paired Monte CarloRepeatedly simulating a month of usage for both the baseline and the change to produce a distribution of cost outcomes instead of a single point estimate.
Tornado sensitivityA ranking of which inputs - token volume, retries, price, model mix - move the cost delta the most, so reviewers can see what actually drives the risk.
Forward escalationA decomposition of forward-looking cost into price, volume and structure effects. It is a forecast of change, not general inflation.
Cost per completed taskCost normalized by successfully completed work instead of per API call. A cheaper model that fails and retries can cost more per completed task than a pricier model that succeeds first time.
Fit-for-purpose model selectionChoosing the model that minimizes cost per completed task for a workload, rather than the one with the lowest per-token price. Cheapest is not the same as correct.
Budget gate (policy gate)A CI check that can fail a pull request when its P90 monthly cost delta exceeds a declared budget. Advisory by default, and it never blocks a change whose delta is zero or negative.
CallsiteA location in code that invokes an LLM. Class1 counts and flags an added callsite but never multiplies it into a fabricated volume; magnitude is learned from actuals, not guessed.
FOCUSThe FinOps Open Cost and Usage Specification, a vendor-neutral format for cloud and AI billing data. Class1 ingests OpenTelemetry-style events and can export FOCUS.
Contingency vs management reserveP90 is contingency for modelable known-unknowns (retries, drift, demand spikes). It is not management reserve for unknown-unknowns like future scope; those re-enter only through the actuals loop.
Frozen basis (effective-dated snapshot)Reproducibility discipline: prices, specs and capability data are pinned to dated snapshots, so the same diff yields the same estimate independent of any live feed.
ReconciliationClosing the loop by comparing a persisted estimate against a real bill for the same deployment window, producing a validation pair that can improve the estimate class.
Shadow modeRunning a newer estimation engine alongside the authoritative one, recording its output for comparison without letting it drive any decision until it earns promotion.
Estimate decayAs the underlying basis ages, the accuracy band widens and the estimate class steps back toward Class 5, so a forecast's confidence cannot outlive the data it rests on.
Retry and fallback tailThe extra cost from repeated attempts, sequential fallbacks and hedged calls. These correlate at the month level and are a primary driver of the P90 tail.
AI carbon footprint (per PR)The carbon, water and materials delta attributed to a change's expected inference, computed from the same Monte Carlo draws as the dollar estimate - a second currency for the merge decision.