In short
Make the case in cost, speed and risk, because the things that cut AI's energy use (smaller models, fewer tokens, fewer wasted calls) also cut bills and waiting. Show one measured example from your own work, add the reporting and regulatory reasons, and keep the environmental benefit as a real but secondary point if that suits your audience.
Steps
- Pick one high-volume AI task and measure its current cost, speed and quality.
- Test a leaner option, such as a smaller model or shorter outputs, against the same quality bar.
- Present the saving in money and response time, with energy alongside.
- Add the reporting, supplier and regulatory reasons that apply to your organisation.
- Propose a small pilot with a clear measure of success and a date to decide.
Lead with cost, speed and risk; energy follows. In AI, less computation means lower bills and quicker answers, so a green AI proposal rarely asks a business to give anything up.
A 2026 UK green AI sandpit of mathematicians, computer scientists and industry reached the same view: buyers choose on cost, speed and quality, so green AI has to arrive with those, not compete with them.
What evidence can I use?
- Smaller models. UNESCO and University College London found task-specific small models could cut energy use by up to 90% on suitable tasks, and shorter prompts and responses by more than half.
- Routing. The RouteLLM study reported cost reductions of more than half in some cases, without lowering response quality, by sending easy requests to cheaper models.
- Runaway costs. Gartner predicted in 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs among the reasons. Controlling tokens is controlling cost.
Your own measured example beats any published figure. Run a small test on a real task.
What other reasons count?
- Reporting: AI is part of your energy and emissions, and larger customers increasingly ask for it.
- Regulation: EU rules are adding energy documentation duties for AI providers and data centres; see EU AI Act energy rules.
- Reputation: a credible, evidenced story is safer than vague claims.
Common mistakes
- Promising savings without a baseline.
- Comparing cost per call, which hides retries; compare cost per successful task.
- Leading with guilt rather than value.
See our guide for business.
Sources
- Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027, Gartner (2025)
- Smarter, smaller, stronger: resource-efficient generative AI and the future of digital transformation, UNESCO and University College London (2025)
- RouteLLM: learning to route LLMs with preference data, Ong et al., ICLR 2025 (2024)
Related
Green AI for business
A practical guide for businesses on cutting the energy, cost and carbon of AI: model choice, suppliers, measurement, procurement and honest claims.
How do I make my company's AI use more sustainable?
A practical starting plan for making a company's AI use greener: map use, right-size models, set defaults, ask suppliers and measure per task.
How do I route requests to smaller models?
How to route routine requests to a small model and only hard ones to a large model, and how to tune the router so retries don't eat the saving.
How do I check an AI project is worth its environmental cost?
A simple way to weigh an AI project's own energy, carbon and hardware against the benefit it delivers, before you build it and once it runs.