AI agents estimate a device's carbon footprint in under a minute, within 19% of experts

Checked against iPhone figures in Apple's environmental reports, the system's error was 18.02% (mean absolute percentage error) with no proprietary data. On seven product categories it had not seen during development, including GPUs, motherboards and wearables, estimates stayed within 5 to 19% of reported values.

Embodied emissions from manufacturing can make up 50 to 82% of a device's total life-cycle footprint, and the authors say proprietary or missing supply-chain data makes assessments slow and costly to scale. The authors note that there is no unambiguously true value for a product carbon footprint, and that company-reported figures used as the baseline are not ground truth.

The system uses two AI agents built on GPT-4o-mini. One plays the life-cycle assessment expert and critiques the inventory; the other plays stakeholders such as engineers and product managers and retrieves data using search tools, an FCC database scraper, and computer-vision tools to detect circuit-board components and measure their size. The two exchange queries until no critical data gaps remain.

The researchers also built a weighted nearest-neighbour estimator from a benchmark of 1,446 electronics products parsed from official carbon footprint reports. Its error converged at about 80 training samples, and all company-reported values fell within its 95% confidence interval. It ranked first in five of six conditions against GPT-4o and three standard machine-learning regressions. A within-subject user study found it improved user confidence, perceived accuracy and trust compared with Google Search and AI chatbots.

In the expert benchmark, nine specialists picked the closest emission-factor match for masked entries, with search engines and AI chatbots allowed. They often chose entries with similar names but very different emissions: for example, polystyrene, general purpose (3.75 kg CO2e per kg) instead of polystyrene, extruded (10.5), a nearly threefold gap. Variation across targets stayed high at 46.9%.

The authors say the agent output should complement human experts rather than replace them, and they describe the system as useful for early design decisions and for flagging hotspots, even where accuracy falls short of formal carbon accounting.