Industrial AI
AI energy optimisation: measure the result
Evaluate AI energy optimisation with clear measurement boundaries. Separate consumption, output, normalised performance and future saving scenarios.
An AI energy-optimisation claim needs a measured energy result and an explanation of what changed in production. A smaller electricity bill alone cannot separate lower consumption from a tariff change, less output or different operating conditions.
The IEA's Energy and AI report explores a Widespread Adoption Case for 2035. It is a scenario built around assumptions about adoption and barriers. Its estimated potential is not a saving a particular factory has already achieved.
Define the boundary before comparing periods
Choose the equipment, process or facility being measured. Keep the energy meter, production count and observation period aligned.
US Department of Energy guidance defines simple energy intensity as energy consumption divided by productive output. It also requires the energy and production measures to cover the same time frame.
An energy measure with a stated denominator
Electricity intensity = electricity consumed within the boundary ÷ stated productive output
Example unit: kWh per unit produced
If you choose good units as the output, state the acceptance definition and keep rejected and reworked production visible. That is a defined metric choice, not a reason to drop their energy consumption from the numerator.
Read consumption and intensity together
| Measure | What it describes | What to investigate beside it |
|---|---|---|
| Total energy | Energy used within the boundary | Production, operating hours and relevant conditions |
| Energy per output unit | A ratio of energy to stated output | Product mix, fixed loads, rejects and the chosen denominator |
| Energy cost | Expenditure at the applicable tariffs | Price changes, demand charges and consumption |
| Normalised performance | A comparison accounting for included relevant variables | Model validity, baseline choice and unexplained changes |
DOE's explanation of normalisation distinguishes simple intensity from measures that account for other relevant factors, including weather and fixed facility consumption. Dividing by output does not automatically control for all of them.
Identify what the AI actually changes
Record whether the application recommends a setting, predicts demand, detects abnormal consumption or controls equipment. Those are different interventions with different evidence needs.
Retain the recommendations, approved changes and operating outcomes. Check quality, throughput and equipment constraints alongside energy. The benefits-realisation guide explains why a reported technical improvement still needs a link to the claimed business result.
Make the pilot comparison reviewable
Agree the baseline, relevant operating variables and acceptable production outcomes before the trial. Identify other changes made during the comparison period and retain them in the record.
At the review, show consumption, the declared output, operating conditions and energy cost separately. Ask the energy and production owners to verify those records before attributing a saving to the AI application.