In short
Make AI features run when a user asks, not automatically on every page load, message or file. Generate once and reuse shared results, use plain code for decisions a rule can make, and remove AI features that few people use. The W3C Web Sustainability Guidelines ask teams to evaluate the need for AI before implementing it, and that applies to each trigger as well as each feature.
Steps
- List every place your product calls a model and what triggers each call.
- Change automatic triggers to user-initiated ones where the result is often ignored.
- Generate shared content once and cache it instead of per user or per view.
- Use rules or plain code for decisions that don't need a model.
- Track usage of each AI feature and retire those few people use.
The greenest AI call is the one that never runs because nobody needed it. Product decisions about when a model fires often matter more than which model it is.
Why does it matter?
An AI feature switched on by default runs for every user, every time, whether or not the result is read. Multiply a single summary or suggestion by millions of page views and the cost is large. Generative output is the expensive kind: Luccioni, Jernite and Strubell found image generation used about 2.9 kWh per 1,000 requests against 0.002 kWh for classifying text. The W3C Web Sustainability Guidelines ask teams to evaluate the need for AI before implementing it.
Design patterns that help
- User-initiated, not automatic. A "Summarise" button instead of a summary generated on every open.
- Generate once, serve many. Product descriptions, help answers or translations shared by all users should be generated once and cached. See how to use caching.
- Rules first. The Green Software Foundation recommends using plain code for decisions a rule can make and calling the model only when needed.
- Event-driven agents. Trigger background agents on real events, not on a timer.
- Honest defaults. Let users switch AI features off and remember the choice.
Common mistakes
- Measuring adoption by calls made, which rewards waste. Measure by results used.
- Pre-generating "in case" content for every item in a catalogue.
- Never retiring features. Review usage regularly.
This is sufficiency in practice: read efficiency is not sufficiency.
Sources
- Web Sustainability Guidelines, W3C (2025)
- Optimize agent orchestration to reduce unnecessary model calls, Green Software Foundation, Green Software Patterns (2025)
- Power Hungry Processing: Watts Driving the Cost of AI Deployment?, Luccioni, Jernite and Strubell, ACM FAccT (2024)
Related
How do I use caching to cut the energy of LLM calls?
Two kinds of caching that cut LLM work: provider prompt caching for repeated context, and your own cache so identical questions aren't answered twice.
How do I turn off AI features I don't need?
How to find and switch off AI assistants, summaries and suggestions you never use in apps, phones and browsers, and why defaults matter for energy.
How do I switch off unneeded AI features across my organisation?
How IT teams can find and switch off AI features nobody needs across office, collaboration and business software, without blocking useful AI.
Efficiency is not sufficiency: the rebound effect in AI
Why making AI more efficient does not guarantee it uses less energy overall, what the rebound or Jevons effect is, and what helps keep savings real.