Cost Tracking and Budgets

This standard governs how the operating cost of the Company's AI services is measured, reviewed, and kept within budget. It is used by AI Operators, who run the cost checks; by the AI Operations Lead, who owns the budget and the response to a spike; and by the Finance function, which reconciles the measured cost monthly. Because the Company operates its own models through its own gateway, the invocation log is the complete meter: every call the services make is recorded there, with its usage, and cost control is a matter of reading that meter correctly and acting on it in the right order.

Roles and duties

• AI Operator: runs the scheduled checks below and reports breaches of the alert thresholds the same day.
• AI Operations Lead: investigates every threshold breach, decides the response, and records what was found.
• Finance function: reconciles invocation-based cost against the operating cost of the AI infrastructure monthly.

What the meter records

Each invocation record carries the feature, the model, the prompt version, input and output usage counts, the computed cost where cost has been calculated, latency, the result, and supporting detail. Two reading rules govern everything in this standard

1. Failed calls are not usage. An error record with no usage counts and very low latency records an attempt in which nothing was computed. Cost shall be computed only on successful records with recorded usage. Summing cost across all records without that distinction produces a wrong number.
2. Latency and usage are independent signals. A slow call with small usage is a slow service, not a long prompt; a fast error is a fast failure. A latency spike shall not send an investigator into the prompts when the usage counts say the request was small.

The usage profile of each call

Operators shall know the normal profile of each feature, because drift from it is the earliest cost signal. On the Pawzly path, per customer turn: triage is the largest input (the instruction, the customer's message, knowledge material, and history); the product selection step is the cheapest call and the highest-volume one (observed at about 170 input and 20 output units per call, against its 120-unit output cap); the grounding rewrite costs more per call (observed at about 265 input and 103 output units, against its 260-unit output cap); and the conversational reply (400-unit output cap) runs only where triage has not already produced the reply. The classroom tutor is metered in the same log on the same basis, so one set of checks covers both services.

The standing checks

Until these checks are presented in a dashboard, they are run as queries against the invocation log on the cadence stated

1. Usage by feature, per day. Successful records grouped by feature, with output usage as the standing proxy for cost until per-call cost is computed for every record: output usage is the larger price component in the Company's cost model.
2. Error rate by feature. The share of records that are not successful, over a rolling 24 hours. A sustained share above a few percent on any feature is investigated the same day.
3. Latency distribution by feature. The middle and 95th-percentile latency of successful records. The observed spread on the selection step is wide relative to its small output; that spread is service variance, and it is what the customer-experience budget must absorb.
4. Usage drift. Average input and output usage per feature per day. A prompt change that lengthens outputs appears here before it appears in total cost. Output caps bound the worst case per call; average drift is what moves the monthly figure.
5. Seller attribution. Where sellers use AI features, usage grouped by seller. A seller at ten times the median is either receiving exceptional value or running an integration out of control; the record distinguishes the two before the cost does.

Budgets and alert thresholds

The following thresholds are procedure, applied by the AI Operator on a daily check

1. Daily AI operating cost above three times its 7-day trailing average: investigate before noon. Causes, in the order they are checked: a prompt change lengthening outputs (usage drift); knowledge material returning more context (input usage); a traffic increase (record counts: the per-address abuse limit caps anonymous misuse, not genuine growth); a seller integration out of control (seller outliers).
2. Any feature's error share above 5% over one hour: check the gateway first, then the credential: a wave of authentication failures is the signature of a rotation performed out of order, under Model Inventory and Lifecycle.
3. The first response to a spike is measurement, not restriction. A feature is restricted only once the cause is known and recorded. The abuse limit is an abuse control, not a cost tool, and tightening it punishes legitimate customers first.

When alerting is automated, each alert shall record its threshold, window, scope (feature, seller, or platform), who is notified, and the action: notify, restrict, or record only: and the thresholds above are the definitions it implements. Thresholds actually used shall be written down as they are adopted, so later automation inherits recorded judgement rather than inventing its own.

Controls

Cost figures quoted internally shall state their basis: computed cost, or output usage as proxy: and the period. No operator shall adjust usage or cost records to meet a budget figure. Budget changes are made by the AI Operations Lead with the Finance function and recorded with their reason.

Records

This standard produces the standing check results, threshold-breach investigations with their findings, monthly reconciliations, and the threshold register. Reconciliations and investigation records are retained for seven years with the Finance function's records.

Escalation

A threshold breach that remains unexplained after one business day, or a single-day cost above five times the trailing average, goes to the AI Operations Lead and to support@aapawz.com the same day. Suspected misuse of a seller integration is handled under AI Disclosures and Agent Registration and, where legal process is involved, legal@aapawz.com.

Review cadence

The standing checks run daily; usage drift and seller attribution are reviewed weekly; budget and thresholds are reviewed monthly with the Finance function.