# The AI Bill: Itemised — Decision log

Record one intervention per entry. Keep the current setup among the options. Missing evidence stays unknown.

## Date

__________

## One decision

__________

## Options considered

1. __________
2. __________
3. __________

## Evidence, billing period and source

Invoice/log/test: __________

What was measured, when, and how: __________

## What is still unknown

__________

## Owner

__________

## Review date

__________

## Measured outcome after review

Measurement, period and source: __________

## Keep, change or reverse — and why

__________

If the review arrives without evidence, write unknown and set a new review date. Do not backfill a forecast as an outcome. Check quality, latency, reliability and engineering effort alongside the token bill. Do not bundle a model change, a prompt change and a volume change in one experiment and attribute the result to only one of them.

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

- Decision: test a shorter summarisation instruction on one task type.
- Options: keep the current instruction; test a shorter instruction; test a different model later.
- Evidence: in an invented comparison, output grew from 600 to 1,200 tokens per attempt while monthly tasks rose from 10,000 to 16,000. Quality and engineering effort are unknown. This supports an investigation, not a conclusion that longer output is waste.
- Unknowns: acceptance under the same quality rule, latency and maintenance effort.
- Owner: assign one person before starting.
- Review date: set a date before starting the experiment.
- Outcome: unknown; no experiment has been run in this fictional example.
- Follow-up: record measured acceptance, token spend and latency for the same task population, then choose keep/change/reverse. Do not claim the forecast as a saving.