Can companies estimate the footprint of employee AI use?

Companies can estimate parts of the environmental footprint of employee AI use when they can connect a defined activity to relevant supplier or technical data. The estimate should identify its period, coverage, allocation method and uncertainty. A useful partial estimate keeps missing services visible instead of presenting the available number as the whole company's footprint.

For a sustainability lead working with IT and procurement, the first deliverable is a coverage record. It should show what employees use, which evidence covers that use and what remains unresolved. Greenhouse gas emissions, electricity and water need separate evidence; a carbon estimate does not answer every environmental question. The tools and workflows guide places resource use alongside the other responsibilities involved in supporting employee AI.

Begin with the services and work you need to cover

In a September 2026 Reddit discussion, a self-described AI adoption project manager wanted evidence about their company's Copilot use. They were estimating activity and borrowing published Gemini information, while explicitly recognizing that the services and assumptions differed. A follow-up described a company of about 140 people. The employer and calculation are unverified, but the difficulty is recognizable: a relevant-looking public number may not describe the service employees actually use.

Suppose Claire Bennett leads sustainability work at a building-surveying firm. Jamie Lee, an IT colleague, helps her examine AI assistance used to draft client visit-summary letters. A responsible surveyor checks each letter against the visit record and agreed next steps before it is accepted.

The firm uses a shared office assistant and a separate document-drafting project. Claire cannot assume that one supplier report covers both. She and Jamie first record:

• The service. Identify the specific purchased product or internally operated system and the account or project involved.
• The activity. Describe the work, relevant usage records and other tasks sharing that service.
• The period. Match the reporting dates with the activity being assessed.
• The coverage. Identify available emissions, energy or water evidence, its exclusions and the person responsible for resolving gaps.

Start with an appropriate aggregate of activity. An emissions estimate does not require collecting employees' confidential letters or publishing individual usage rankings. Ask the existing data and service owners which records can be used for this purpose.

Check that a dashboard includes the AI service

A familiar supplier name is insufficient evidence of coverage. A report may cover some products, omit others or describe an allocation across a wider service bundle.

Microsoft's Emissions Impact Dashboard for Microsoft 365 documentation, checked in October 2026, explicitly says that Copilot for Microsoft 365 is not currently included. The report's coverage of applications such as Word or Teams should therefore not be treated as proof that the Copilot activity is included. Recheck that scope before a reporting decision because product coverage can change.

Google Cloud's covered-services list, stated as of August 2026, includes Gemini API and Vertex AI. That is a different reporting context. It does not establish that an employee's separate subscription, an unrelated application or every product with a similar name appears in the company's cloud report.

For Claire's review, keep the evidence and the claim it can support together:

The absence of the second service from an emissions report is an unresolved gap. It does not make its impact zero, and it does not justify assigning all of the supplier's emissions to Claire's firm.

Keep the scope statement beside the number when the estimate reaches a manager. If the report covers only the drafting project, name that project in the conclusion. “Employee AI footprint” would promise coverage that the evidence has not yet supplied.

Understand what the supplier has allocated

Cloud services share equipment and supporting infrastructure. A customer report can therefore involve allocation rules rather than a direct meter reading for an employee's task.

Google Cloud's reporting methodology describes allocating shared infrastructure emissions to products and customers using service and usage information. It also includes allocations for non-electricity sources and identifies exclusions. The company states that the customer-specific emissions data has not been independently verified or assured. Read the method and the particular report together; access to a report does not remove its assumptions.

For each relevant number, ask:

• Is it measured activity, a supplier allocation, a modelled estimate or a combination?
• Which lifecycle activities and supporting infrastructure are included?
• What unit, period and account does it describe?
• Which electricity-accounting basis, location assumptions and method version apply?
• What is explicitly excluded or unavailable?

A count of licences answers a purchasing question. A count of requests records interactions. Neither automatically gives the resources used by a particular workload. Long documents, retries and different kinds of generation can make apparently similar interactions unlike one another.

The Green Software Foundation's Software Carbon Intensity specification, version 1.1.0, provides a useful methodological distinction. It expresses a rate using operational and allocated embodied emissions per consistent functional unit, meaning the unit of activity the rate describes. It also distinguishes measurement from modelling and requires comparable methods when assessing a change. That is a method to apply with suitable evidence, rather than a ready-made conversion factor for an undisclosed service.

If Claire lacks the data required by a chosen method, record that limitation. Do not label a rough allocation as a certified score or a complete company inventory.

Match the numerator to the work being counted

A number becomes easier to misuse when divided by a convenient denominator. Before calculating emissions per letter, Claire needs to establish what the reported total includes and what counts as a finished letter.

Suppose the supplier allocates 12 kg CO₂e to the identified drafting project for September. Suppose further that the project was used only for this letter workflow, all relevant attempts fall within that allocation and the team accepted 60 letters in the same month. The illustrative calculation is:

The allocation covers the identified drafting project for September. It leaves out the office assistant whose coverage remains unresolved, along with activities excluded by the supplier. It also does not predict the additional physical emissions from generating the next letter.

Retain discarded attempts in the activity behind the allocation. Counting only the final generated version can make the work look smaller than it was. In an October 2026 LinkedIn account, Julia Charles-Wiginton described generated clips, voice tests and unused versions behind a finished creative video. She was discussing work with a carbon-tracking business to develop a feature, not reporting a validated footprint. The process detail is useful: the published result is not the complete record of attempts needed to produce it.

The accepted-letter denominator serves a different purpose. It connects the allocation to useful work, provided the activity and period genuinely match. The guide to measuring net time saved explains how to define an accepted result and keep preparation, checking and failed attempts in view. Human minutes and emissions remain different measures; one cannot be converted into the other without appropriate evidence.

If the project also handles unrelated marketing drafts, Claire cannot assign the whole total to visit-summary letters. Obtain a justified separation or report the project total without claiming a letter-specific rate. If the intended comparison uses different quality standards or coverage, fix that mismatch before interpreting the numbers.

Preserve uncertainty and changes in the method

An estimate should be reproducible by someone who did not make it. Keep a compact record containing the source report, calculation, reporting period, allocation choice, exclusions and outstanding questions. Give each material gap an owner.

Where a meaningful input is uncertain, work with the responsible analyst to show how plausible alternatives affect the result. Do not invent narrow upper and lower bounds merely to make a chart. A scenario range describes the assumptions tested; it is not automatically a statistical confidence interval.

Keep these interpretation checks visible:

• Missing coverage. List unresolved services separately from the available total.
• Overlapping coverage. Check whether a supplier allocation already includes hardware or training before adding another estimate.
• Changed methods. Establish whether a movement reflects different activity, a revised factor or a restated supplier calculation.
• Different impacts. Keep carbon, electricity and water separate when the evidence cannot connect them appropriately.

Google's methodology warns that changes can adjust current and historical customer data. Preserve the report version and compare periods on a consistent basis where possible. A lower newly reported number does not, on its own, establish that Claire's team physically reduced resource use.

The GHG Protocol's supplier-data guidance recommends seeking relevant activity or emissions data and documentation of methods, assumptions and exclusions. A focused request is more useful than asking a supplier to provide an unexplained total for “our AI”. Coordinate it with procurement and the person responsible for the organization's environmental assessment.

For the next review, bring one service's coverage record and the decision its unresolved gap affects. Claire might be able to improve the drafting project's estimate while keeping the office assistant's footprint unresolved. That is progress in evidence quality, provided the conclusion remains as narrow as the work supports.

Questions about estimating employee AI footprints

Does a cloud emissions dashboard cover every employee AI tool?

No. Coverage depends on the specific report, purchased service, account and period. Microsoft's documentation checked in October 2026 excludes Copilot for Microsoft 365 from its Microsoft 365 emissions report, while Google Cloud lists particular AI services in a different reporting context. Match each employee service to its current coverage before describing a report as the company's AI footprint.

Can we allocate emissions using employee headcount?

Headcount may be an explicitly chosen approximation in a suitable method, but it does not establish how much computing each group used. A team with occasional short requests and a team running repeated document processing can differ substantially. Ask whether relevant service activity or supplier allocations are available, and document the reason and uncertainty if headcount is used.

Is a request count enough to estimate emissions?

A request count needs a factor that fits the service, workload, period and calculation boundary. It must also account appropriately for retries and other relevant activity. Borrowing another service's median prompt statistic does not establish those conditions. Keep an unmatched public figure as context while seeking the missing data.

Should we add model training to the supplier's estimate?

First check the supplier's included activities and allocation method. Adding training or equipment impact that is already included would double-count it. If the estimate excludes a material activity, identify a compatible method for addressing that gap with the responsible analyst. Leave it explicit when the evidence cannot support a defensible allocation.

What if the supplier revises last year's figures?

Preserve the original and revised report versions and find out what changed. Recalculate comparisons on a consistent basis where possible, and explain which differences arise from activity and which from methodology. A restated estimate is not, by itself, evidence of an operational emissions reduction.