Building a Carbon Impact Agent

AI is reshaping industries from healthcare to energy, security to agriculture, but this progress comes with a significant environmental cost. Since 2012, the computational power required for cutting-edge AI models has doubled every 3.4 months, fundamentally altering global electricity demand. Training large models like GPT-3 alone has been estimated to produce more than 550 tons of CO₂ - the equivalent of 300 round-trip flights between New York and San Francisco. While these upstream impacts of sustainable model training are well-documented (e.g. model compression techniques, neuromorphic chips and renewable-energy powered infrastructure), the downstream footprint of everyday AI applications remains largely overlooked.

Today, most AI designs emphasise SOPs, data quality, and evaluation, but sustainability is rarely considered. When it is, it tends to be incidental, the byproduct of choosing a cheaper or smaller model. Once deployed, very few applications measure their ongoing consumption footprint. Yet with the right monitoring in place, clear and powerful strategies emerge: caching repeated requests, nudging users toward greener prompts, scheduling tasks off-peak, or routing workloads to cleaner-energy regions.

In this session, we’ll design a Carbon Impact Agent layered on top of a support agent. This multi-agent system tracks, analyses, and recommends approaches to actively reduce the environmental footprint of AI applications. We will explore:
• What to measure: compute usage, energy mix, task-level telemetry, carbon intensity by region.
• How to design the system: agent orchestration for monitoring, calculation, analytics, and recommendations.
• What insights emerge: per-task carbon costs, workflow-level emissions, and team-level benchmarks.
• How to act on them: UI nudges, greener deployment choices, prompt optimisation, and workload scheduling.

By the end, you’ll see how we can build AI systems that are not only impactful, but also sustainable by design, embedding environmental accountability into the fabric of intelligent applications.