# The AI Bill: Itemised — Bill investigation worksheet

Use one worksheet for one workload and billing period. Write unknown when evidence is missing. Short notes fit a one-page print layout at /listener-tools.

## Workload and task definition

Feature: __________  Period: __________  Owner: __________

One original task means: __________  Original task count: __________

## Evidence and token categories

Invoice/log/estimate sources and dates: __________

Average tokens per billed attempt: fresh input ____; cache read ____; cache write ____; output ____.

Prices (USD / 1M tokens), with source/date: fresh ____; read ____; write ____; output ____.

Fresh/read/write are exclusive categories. If cached tokens are included in total input, subtract them to obtain fresh input. Count reasoning once when already included in output. Separately billed components require their own amount.

## Cache and retries

Repeated prefix: ____ tokens. Cache hit rate: ____% of ____ cacheable requests.

Cache lifetime, storage/write rules and possible causes of misses: __________

Attempts per original task: ____. Additional billed attempts per original task: ____.

Are failed attempts charged? Yes / No / Unknown. Billing policy evidence: __________

Main retry reason: __________

## Fees and invoice scope

Other charges in period: $____. Includes: __________

Currency, credits, taxes and excluded costs: __________

## Acceptance, latency and engineering time

Accepted original tasks: ____ / ____ originals = ____%. Judgment/test and evidence: __________

Quality change: Better / Same / Worse / Unknown. Source: __________

Median latency: ____; P95: ____; limit that changes the decision: ____; source: ____.

Engineering hours: ____. Work covered: ____. Still worth doing at twice the effort? Yes / No / Unknown.

## Totals and change

Token spend: $____; other charges: $____; total with known charges: $____.

Cost per original task: $____; cost per accepted original task: $____.

Expected cost at current prices: $____. Changes: volume $____; output $____; fresh input $____; cache $____; retries $____; fees $____.

Unexplained difference: $____. Missing evidence: __________

## One next check

Evidence that would change my mind: __________

Action: __________  Owner: __________  Review date: __________

## Method and limits

Token cost per billed attempt = sum of each exclusive token category × its price / 1,000,000. Billed attempts = original tasks × (1 + additional billed attempts per task). Token spend = billed attempts × token cost per attempt. Total = token spend + other charges. Cost per accepted task = total / accepted original tasks; unavailable when acceptance is unknown or zero.

This assumes all billed attempts use the same average token mix. Keep different currencies and workloads separate. Check provider-specific billing, quality, latency, reliability and engineering time. A modelled result is not a measured saving. A low cache hit rate is a reason to investigate, not proof of waste.

## Worked example (fictional, not a customer result)

A before period has 10,000 tasks, 4,000 fresh-input and 600 output tokens per attempt, and prices of $2/$10 per million. With no cache, charged retries or other fees, token spend is 10,000 × (4,000 × 2 + 600 × 10) / 1,000,000 = $140.

The after period has 16,000 tasks, 4,000 fresh-input and 1,200 output tokens per attempt, at $1/$5 per million. Token spend is $160. The old workload at the new prices costs $70; extra volume adds $42 and longer output adds $48. The token residual is $0. These invented numbers explain the arithmetic, not any real team's spend.

If acceptance were 85% before and 80% after, cost per accepted original task would be $140/8,500 = $0.01647 and $160/12,800 = $0.01250. Those acceptance rates are illustrative assumptions, not measurements. Quality, latency and engineering effort remain unknown in this example.