40 Questions Energy and Utilities Should Ask Before Trusting AI
Your grid operators and traders make decisions that affect millions of people based on AI outputs they cannot fully explain. A single bad question or missing interrogation costs money, stability, and trust in your organisation.
These are suggestions. Use the ones that fit your situation.
1When Palantir recommends load shedding in a frequency event, can your control room operator manually override it in under 30 seconds without checking documentation?
2Does your Azure AI grid forecasting model have a documented failure mode for the specific weather patterns that caused your last three emergency responses?
3If your AI predicts a voltage collapse in zone 7, what is the human operator supposed to do in the first 90 seconds before the recommendation becomes obsolete?
4Can you identify which training data was used to teach your grid stability model to distinguish between a harmless transient and a real threat?
5When your SCADA system disagrees with your AI's recommended reactive power settings, which one are your operators trained to trust?
6Has your grid AI been tested on the exact sequence of events that would occur during a loss of your largest generator?
7If your renewable integration AI fails at 2am on a high wind night, what manual procedures exist to hold the grid stable?
8Does your AI alert your operators when it is operating in conditions that fall outside its training data, or does it output recommendations anyway?
9Can a new grid operator understand why the AI recommended that specific action, or do they have to accept it on authority?
10If the AI makes a decision that appears correct but causes a cascading failure six minutes later, how will you know the AI was the cause?
Energy Trading and Market Decisions
11When your ChatGPT or similar tool gives a trader a market recommendation, what specific data sources does that recommendation actually draw from?
12Has anyone in your trading desk verified that the AI's confidence score means what it claims to mean at the speed your traders need to act?
13If your AI recommends selling forward at a specific price, can your head trader articulate the market logic in their own words?
14When market conditions shift in the last five minutes before a trade window closes, can your traders ignore the AI recommendation or do they lack confidence?
15Does your AI trading tool have a separate alert when its recommendation contradicts historical patterns, or does it treat all outputs equally?
16If an AI recommendation loses money, do you have a trader who understands why well enough to decide whether to follow similar recommendations in future?
17Can your compliance team identify the exact reasoning a trading AI used for a position that regulators later question?
18When your AI recommends hedging based on correlation analysis, do your traders understand what correlations it is using and why?
19If the AI model is retrained weekly, how do your traders know whether today's recommendation comes from yesterday's model or a new one?
20Does your trading organisation still maintain expertise in the market fundamentals that your AI is supposed to predict, or has that knowledge atrophied?
Predictive Maintenance and Asset Management
21When IBM Maximo recommends replacing a transformer in 90 days, what sensor readings is that prediction actually based on?
22Has your maintenance team verified that the AI's failure predictions match what actually fails when you later inspect those assets?
23If your maintenance AI fails to predict a failure that then occurs, can you determine whether the training data was insufficient or the model has a blind spot?
24Do your field technicians still know how to assess transformer oil condition by practical inspection, or do they only trust the AI result?
25When the AI recommends deferring maintenance on a critical asset, what is the risk tolerance that recommendation is based on?
26Can you identify the age and condition of the assets used to train your predictive maintenance model, and whether your fleet matches that profile?
27If your AI recommends increased inspection frequency for a specific asset type, do you understand whether that is based on physics or on a pattern in historical data?
28Does your maintenance scheduling system allow your teams to override an AI recommendation with a documented reason that gets logged and reviewed?
29When you schedule a major maintenance outage based on AI prediction, what happens to grid reliability during that outage if the failure does not actually occur?
30Has anyone calculated the cost difference between following every AI maintenance recommendation versus ignoring them all and using experienced judgment?
Sustainability Reporting and Data Integrity
31When your sustainability report contains AI-processed emissions data, who independently verified that the AI's calculation method matches your regulatory definition?
32If your carbon accounting AI processes data from Aurora Solar or similar sources, what happens when the source data is incomplete or delayed?
33Can your compliance officer explain the calculation your AI uses for scope 2 emissions, or do they have to trust the output without understanding the method?
34Does your AI sustainability model handle changes in your energy mix, or does it assume your current portfolio is stable throughout the reporting period?
35When your renewable energy AI calculates avoided emissions from solar or wind, what baseline grid carbon intensity is it assuming?
36If a regulator requests the raw data your AI processed to calculate sustainability figures, can you provide it with documentation of any filtering or transformation?
37Has your organisation tested what happens to your sustainability metrics if the AI model changes or gets retrained on new data?
38When your AI estimates renewable generation on days with missing weather data, what method does it use to fill the gaps?
39Do your sustainability reports disclose the limitations of the AI models used, or are the figures presented as directly measured?
40If a sustainability claim in your public reporting is later found to be inaccurate due to AI error, can you demonstrate that your controls were reasonable?
How to use these questions
Ask your AI vendor for the specific training dataset age and composition. If your fleet includes assets built after the training data was collected, you have a coverage gap.
Require that every significant AI recommendation includes a confidence level with a clear definition of what that number means at your decision speed.
Schedule quarterly exercises where your control room operates without AI recommendations for one hour. If operators cannot function, you have a judgement dependency problem.
Document every instance where an operator overrode an AI recommendation and why. This becomes your evidence of where human judgement still adds value.
For compliance and trading decisions, maintain a small team of people with deep domain expertise who do not primarily rely on AI. They become your reality check and your recovery option in a crisis.