Outgrace
OutgraceThe Leadership Magazine, home
Technology & AIQ&A primer

Before an AI tool monitors your emissions data

Twelve questions on data quality, accuracy, audit trails and reporting duties, to settle before a pilot, and the public records to test the answers against.

Communication, workplace, imac, desktop, creative, computer
Photo: ptra / Pixabay

Most companies that want an AI system to watch their emissions or energy use will not be asked whether the model is clever. They will be asked whether the numbers it produces can be defended. This primer is for founders, finance leads and directors who must approve or refuse such a proposal, and who need to know which questions separate a useful pilot from an expensive way of producing figures nobody can check. According to the US Environmental Protection Agency, approximately 8,000 facilities already report greenhouse gas data every year under its programme, so the reporting bar is not theoretical. Remember one thing: a monitoring tool is only as good as the record it is checked against, and that record is usually older and less tidy than the sales pitch suggests.

Approximate number of facilities reporting annually to the US Greenhouse Gas Reporting Pro
8,000
Surveys through which the US Energy Information Administration collects energy data
60+
Share of global primary energy use that electricity represents, roughly (Our World in Data
One-fifth
Source categories covered by EPA's September 12, 2025 proposal to remove reporting obligat
46
What exactly will the tool produce, and which decision will it change?

Start with the output, not the model. Write down the figure the tool will produce, the person who will rely on it, and the decision that figure will change. If the result is a dashboard that nobody acts on, the project has no business case, however accurate it is. In our experience the most useful proposals name a single figure, such as monthly generation at one plant or annual emissions at one site, and a date when someone will use it. Anything broader tends to become a reporting exercise with no owner. Ask the vendor to show that figure on your own data before discussing features.

Which existing number will it replace, and does the new one agree with it?

Every monitoring proposal competes with a number that already exists, so find it. For a US facility, the record filed with the EPA is a public benchmark. The agency says reported data is made public each October, and that its 2023 programme data is now available. Compare the tool's output with the filed figure for the same facility and year. A difference is not automatically an error, since methods differ, but it must be explained before anyone relies on the new number. If the model cannot be reconciled with the filed figure, the board should hear that before it hears the accuracy claim.

Is the input data complete enough to trust?

Test completeness before accuracy. The US Energy Information Administration says that in theory supply and demand balance, but in practice small differences remain between the totals. It explains the gap by stock changes and other uses that it tracks. A monitoring model fed with partial meter readings will show the same kind of gap, only larger and harder to explain. Ask what share of readings is missing in a typical month, how the gaps are filled, and whether the output flags every filled value. A model that hides its filling is not monitoring. It is estimating with better formatting.

Are we measuring what we think we are measuring?

Electricity is the usual trap. Our World in Data says its electricity dataset measures generation, not consumption. The gap between the two is small for most countries but can be substantial for heavy importers or exporters. A system that reports generation and labels it demand will look accurate and still mislead a buyer or a regulator. Ask which quantity each output label refers to, and whether the model can switch between them. If your business imports or exports power on any scale, make this the first test you run.

Which energy measure is the model using?

Primary energy is the measure most international energy data uses. It counts energy entering the system before transformation. Our World in Data notes that it also counts energy wasted along the chain, so it overstates end-use consumption. A model that reports primary energy as if it were the energy your site consumes will show a figure that is too high. Ask the vendor to name the boundary it uses: primary, final or useful energy. Our World in Data also says no international open database exists for final or useful energy consumption, so ask how the vendor built its baseline, and ask to see it.

What accuracy do we need, and who decided it?

Set the tolerance before the pilot, in writing, and let the business set it rather than the vendor. A single headline accuracy percentage tells you little unless you know what it was measured against, on which sites and in which months. Ask for the test data, the period it covered, and the error at the level of the figure you will actually report. A tolerance that is acceptable for an internal trend line may be unacceptable for a filed number. Keep the two standards separate, and do not let success on one use case quietly set the standard for the other.

Can someone outside the company check how each figure was reached?

This is the auditability question, and it is where most pilots are weakest. Ask for a record of inputs, model version, adjustments and the person who approved each reported figure, kept for as long as a regulator or auditor could reasonably ask about it. The EPA's reporting system offers a useful benchmark for access control: according to the agency, users of its e-GGRT system have needed multi-factor authentication since August 16, 2024. Your AI vendor should meet at least that standard for access to data about your operations. Ask who can change the model and the historical records, and how each change is logged.

Are the rules we will report under still the same?

Do not assume they are. The EPA has issued a proposal, dated September 12, 2025, to permanently remove reporting obligations for 46 source categories under the greenhouse gas reporting programme. The source for this article does not confirm whether that proposal has been finalised, so check its current status with your regulatory counsel before you design a reporting system around it. A monitoring system built for a duty that later disappears is a sunk cost. A system that produces auditable records of your own choosing survives the change. If the reporting duty affects your legal position, that question belongs with a lawyer, not a monitoring vendor.

Who owns the third-party data the model learns from or ingests?

Read the licence before the contract. Our World in Data says third-party data carries the original authors' licence terms, which should be checked before reuse or redistribution. The same discipline applies to public energy datasets, regulator databases and vendor-supplied benchmarks. Ask the vendor for a list of every external source in the pipeline, the licence attached to each, and whether your reports would redistribute any of them. A model that cannot say where its inputs came from is one you cannot defend in front of a board or a regulator.

How should we cite the figures the model produces?

Cite the underlying data, not a topic page or a vendor summary. Our World in Data's own practice is to cite the underlying data sources, not only the topic page. Apply the same rule internally. Each figure in a board paper should point to the dataset, survey or filing it came from, and to the version used. The US Energy Information Administration publishes its survey forms with descriptions and links to instructions, which is the kind of traceability to ask of your vendor. If the output cites only the vendor, treat the figure as an opinion until someone has reconciled it.

Can we test it on our own data before we sign?

Yes, and you should insist on it. A serious vendor will accept a test on a defined set of your sites over a defined period, with the results compared against figures you already hold. The US Energy Information Administration invites users to test products in development and to give feedback, which is a sensible model for a buyer to copy. A pilot that runs only on the vendor's demonstration data proves that the vendor's data works, and nothing more. Agree in advance what would count as a pass and what would make you stop, and write both down before the test begins.

What should we pay for, and what should we refuse to pay for?

Pay for the parts that make a number defensible: data cleaning, reconciliation to filed records, access controls, a documented model version and a way to explain each figure. Be sceptical of anything that sells an impressive dashboard without these. Money spent on output that cannot be audited is money spent on presentation. This is a judgement about where to spend effort, not a financial or tax recommendation. For the accounting treatment of any such spend, take advice from your accountant.

The mistake behind most of these

Most of these questions fail for the same reason. The proposal treats monitoring as a question of model quality, when the real question is whether the underlying record can be trusted and explained. Electricity generation is not consumption. Primary energy is not end-use energy. A balance that does not close is not a flaw the model can average away. Each gap is a fact about the data, and a sophisticated model can make a wrong input look convincing. In our experience, the boards that get into trouble are the ones that approved the model before agreeing what the number was for. Reverse the order. Settle the figure, the tolerance and the audit trail first, then choose the tool.

This week, take one proposal already on your desk and run it through these questions. Write the answers on a single page, and leave a blank wherever the vendor could not give one. The blanks are your conditions before any pilot begins. Ask your finance lead to pull the filed figure for one site from the public record and compare it with what the tool would have said. Then send the page to the people who will sign the number. If the reporting duty behind the figure may change, bring your regulatory counsel in before the pilot starts, not after. A pilot that starts with a clear answer to each question is a test. One that starts without them is a purchase.

Newsletter

The Briefing

One email a week: the leaders we profiled, the ideas worth stealing, and nothing else.

Weekly, free, one click to leave. Read our privacy policy.