In short
Pick a functional unit that matches the value the AI delivers, such as per document summarised or per successful task, and divide energy by it. Track energy per unit, quality per unit and the total side by side. Per-unit figures show whether the system is getting more efficient; the total shows whether growth is outrunning it.
Steps
- Choose a functional unit, preferring per successful task over per call.
- Tag every model call with the task it belongs to, including retries and failed attempts.
- Divide energy, measured or estimated, by the number of units delivered.
- Track quality or success rate for the same units.
- Report per-unit and total figures together every month.
Totals rise with growth and hide efficiency; per-task figures show it. Tracking both tells you whether you are getting better and whether that is enough.
Which unit should I choose?
| Unit | Good for | Watch out for |
|---|---|---|
| Per call | Simple logging | Hides retries and agent loops |
| Per request or answer | Chat and search features | Doesn't show whether the answer was useful |
| Per document, ticket or image | Batch processing | Documents vary in size |
| Per successful task | Agents and workflows | Needs a clear success check |
Per successful task is usually best, because one user request can involve many calls. Anthropic reported that its agents typically used about four times more tokens than chat interactions, and multi-agent systems about fifteen times more.
Why keep the total too?
Luccioni, Strubell and Crawford warn that efficiency gains in AI can be cancelled out by growth in use, the rebound effect. If energy per task halves while the number of tasks triples, the total still goes up. The Green Software Foundation's SCI specification is a rate per functional unit for this reason, but you should still report the absolute figure beside it.
Common mistakes
- Counting only successful calls, so failed attempts vanish from the figures.
- Switching units between reports.
- Reporting efficiency without quality. A cheaper task that fails more often isn't an improvement.
See stopping efficiency gains being cancelled out and how to measure AI energy.
Sources
- Software Carbon Intensity (SCI) specification, Green Software Foundation (2024)
- How we built our multi-agent research system, Anthropic (2025)
- From efficiency gains to rebound effects: the problem of Jevons' paradox in AI's polarized environmental debate, Luccioni, Strubell and Crawford, ACM FAccT (2025)
Related
How do I calculate a Software Carbon Intensity score for an AI feature?
Work out an SCI score for an AI feature: energy times grid intensity, plus embodied emissions, divided by a functional unit such as per request.
How do I estimate the energy of my AI API calls?
You can't meter a hosted AI API, but you can estimate energy per call from token counts and model assumptions with EcoLogits, and log it next to cost.
How do I stop AI efficiency gains being cancelled out by more use?
How to keep AI efficiency savings from being swallowed by growing use: track totals, cap or budget use, retire freed capacity and ask if it is needed.
How to measure the energy and carbon of AI
A practical guide to metering or estimating the energy and carbon of AI work: tools, system boundaries, per-task metrics and honest uncertainty.