A part of the platform, not an add-on
Your AI workloads have a carbon footprint. Carbonex is designed to account for them like any other source.
Not something bolted to the side, and not a separate spreadsheet. The same methods, the same evidence, the same approvals, the same reporting, and the same security and quality controls as every other emission source in the platform.
A growing source with nowhere to live.
Organisations are running more AI workloads every quarter – training, fine-tuning, inference at scale, and a long tail of embedded model calls inside products that nobody centrally tracks. The associated energy consumption and emissions are real and growing.
They are also unusually hard to account for. The computing is often somewhere else, run by someone else, in a region with a different grid intensity, and at a level of use the customer cannot see. Dividing a shared facility fairly between the organisations using it is genuinely difficult. Vendor disclosure varies from detailed to non-existent.
Faced with that difficulty, most ESG platforms have done nothing. AI emissions arrive, if they arrive at all, as a loose line item with no stated method, nothing behind it and no route through approval. It is the part of the inventory least likely to survive a question.
The same eight controls, applied without exception.
| Control | What it means for AI emissions |
|---|---|
| Identity | An AI workload is an emission source in its own right – not a note attached to something else. It can be set up, given an owner, and asked for like any other source. |
| Method | The calculation method is explicit, configured and versioned. Where a method rests on an assumption – a grid intensity, a utilisation figure, a vendor-published factor – that assumption is part of the method rather than part of somebody’s reasoning. |
| Evidence | Whatever supports the figure – vendor reporting, metering, usage exports – is attached at submission and carried with the value. |
| Approval | AI emissions pass through the same review and approval path as every other source. There is no fast lane and no exception. |
| Closed periods | They are fixed into the closed period with everything else. A report’s AI emissions are as repeatable as its fuel combustion figures. |
| Reporting | They appear in the same outputs, traceable to the same depth, with the same calculation annex available behind them. |
| Security | The same separation between organisations and the same access rules apply. AI usage data can be commercially sensitive and is treated accordingly. |
| Quality | They carry data quality and completeness like every other value, and those attributes aggregate into the totals rather than being lost on the way up. |
The difference this makes to a real inventory.
When an auditor asks how you arrived at the emissions attributed to your AI workloads, the answer should be the same kind of answer you give for a fleet of vehicles: here is the activity data, here is where it came from, here is the method, here is the factor and its version, here is who approved it, and here is the closed period it was locked into.
Today, for most organisations, that answer does not exist – not because the work was done badly, but because the systems have no place to put it.
Where the data genuinely is not available, we say so.
This is the part that matters most, and it is where a lesser approach would be tempting.
For a meaningful share of AI workloads, the emissions data an organisation would need simply is not obtainable today. A vendor may not publish it. Shared-infrastructure attribution may not be determinable at the granularity required. Regional grid intensity for a specific facility may not be disclosed.
Carbonex is designed to record those situations as what they are – unavailable, with the reason attached – rather than producing a plausible-looking number to fill the cell. Where an organisation decides an estimate is appropriate, that estimate is designed to require a stated method and a named approver, and it is designed to be visibly an approved estimate in every report it appears in, never indistinguishable from a measured value.
An AI emissions figure that quietly guesses is worse than no figure, because it will be quoted.
Being built, not shipped.
AI Emissions is part of what we are building, not something we have delivered. It is described here in the detail it is being designed to, because it is the clearest expression of how the whole platform is meant to work – and because we would rather you evaluate the thinking now than be surprised by it later.