Check an AI trial's cost per accepted deliverable
The bill covers attempts, while the team wants to know what each usable result cost. Those are different counts.
What you'll make
a checked usage-cost calculation with a clear denominator and the costs it leaves out.
What you'll need
ChatGPT in your work account, a few trial records and a calculator. Allow 15 minutes. Use fictional or shareable records; real bills may contain account identifiers. This calculation covers the costs you supply, not a supplier price or a forecast.
Example result — illustrative calculation, using invented costs:
| Measure | Calculation |
|---|---|
| Recorded usage | £2 + £3 + £1 = £6 |
| Accepted deliverables | 2 |
| Usage cost per accepted deliverable | £6 ÷ 2 = £3 |
1. Define an accepted deliverable
Select Chat › New chat in ChatGPT. Say what counts as one accepted result. A second attempt at the same result isn’t a second deliverable.
Calculate recorded AI usage cost per accepted audio-tour script. Add the costs of every attempt, including rejected attempts, and divide by the number of distinct accepted scripts. Show the working. Keep other costs separate; do not estimate them.
Fictional trial records:
Script A, attempt 1, rejected, usage cost £2.
Script A, attempt 2, accepted, usage cost £3.
Script B, attempt 1, accepted, usage cost £1.
Review time and seat costs were not recorded.Why this works: the numerator counts spending on the job; the denominator counts results somebody accepted.
What you should see: £6 of recorded usage, two accepted scripts and £3 per accepted script.
2. Recalculate both counts yourself
Add the three costs with a calculator. Count the distinct accepted script names. Then check whether every rejected attempt remains in the total.
Why this works: a plausible unit cost can hide the wrong denominator or missing failed work.
What you should see: the same £6 total and count of two.
3. Ask for the incomplete-cost note
Write a two-sentence note beside the checked calculation. Say exactly which recorded costs it includes and which missing costs prevent calling it a fully loaded cost. Do not estimate labour, seat allocation or savings.Why this works: the label prevents a narrow usage calculation becoming a claim about the whole process.
What you should see: review time and seat allocation listed as missing, with no invented saving.
Check the result
- Are rejected attempts included in the cost total?
- Are accepted scripts counted once each?
- Is the result labelled usage cost, with missing costs named?
The usual failure: the answer divides by all three attempts and reports £2 per deliverable.
Correct the denominator. There are three attempts but only two distinct accepted scripts. Keep all £6 of recorded usage cost, divide by two, and report £3 usage cost per accepted script. Review time and seat costs remain unrecorded.Worked example
Fictional example.
Script A: attempt 1 rejected, £2; retry accepted, £3.
Script B: first attempt accepted, £1.
Two scripts accepted. Three attempts charged. No labour data recorded.Script A needs a rejected £2 attempt and an accepted £3 retry. Script B is accepted after a £1 attempt. The weak result is £6 ÷ 3 = £2.
The repaired result is £6 ÷ 2 = £3 per accepted script. It counts failed work in spending but doesn’t count that work as an accepted result. No labour cost or business saving has been calculated.
Go deeper
Handle a trial with no accepted output
If nothing is accepted, cost per accepted result is undefined. Record total spending and zero accepted outputs. Don’t divide by one to make the sheet work, and don’t report a zero unit cost.