Budget for the completed work, not only the subscription. Add up setup, running costs and the people who review and correct the results, then divide by the outputs you actually accepted.
The trial was cheap, so won’t running it be cheap?
A trial doesn’t show the whole operating picture. Ask what the work costs once people depend on it.
1. Separate setup from ongoing work.
Setup. Preparing information, connections, configuration and testing.
Running. Licences, usage, hosting, integrations and support.
People. Reviewing results, correcting exceptions and approving changes.
A fictional month: count the review too.
- Tools and usage
- CA$80 + CA$20 + CA$15 = CA$115.
- Review
- 100 runs × 3 minutes = 5 hours. At CA$40 an hour, CA$200 of staff time.
- Combined
- CA$315, before setup, maintenance and other costs.
- Accepted outputs
- 80 of the 100 runs. CA$315 ÷ 80 is about CA$3.94 per accepted output.
Practical check: Staff time is an opportunity cost, not necessarily an extra invoice.
2. Measure a usable result, not just a run.
Compare the same job and quality standard with the current process, and investigate the unsuccessful runs as well.
3. Ask what grows when volume grows.
Volume. What happens when usage doubles?
Exceptions. Which failures consume extra usage or people’s time?
Limits. Which spending controls and alerts actually stop or constrain use?
Practical check: Record the assumptions, the exclusions and the owner of each cost.
Read more: Does AI Actually Save Time? How to Measure the Whole Job · Maintaining an AI Agent: What to Do When Things Change
The figures are fictional and are not a benchmark, quote or forecast. This is general guidance, not advice for your situation.