GreenPT Advisory
How we calculate this
We sell measurement. That means it has to be traceable how our own estimates are arrived at. This page shows, for each outcome of the calculator, which figure we use, where it comes from, what exactly was measured, and how certain it is. We link to the original report, not to news coverage of it.
The calculator deliberately gives ranges, not single numbers. That is not carelessness. The differences between organisations are so large that any average misleads. The calculator makes that spread visible, and in doing so shows why only a real measurement gives certainty.
Some sources below are companies that sell software or AI. We use them because their datasets are by far the largest publicly available, and because the findings corroborate each other across independent sources. That interest exists, and you should know about it.
Employees using AI
What the calculator shows: an estimate of how many of your employees use AI for work, based on a median of 15% of employees.
Source: Netskope, Cloud and Threat Report: 2026, based on Telemetry from millions of users worldwide, Oct 2024 - Oct 2025.
What was measured: the share of users in the median organisation who interact with AI apps monthly, established through network telemetry. This is measured behaviour, not a survey. Surveys ask people whether they use AI and come out higher; telemetry sees what actually happens. We use the telemetry figure because it measures established use rather than perception.
Confidence: high for the figure itself. The band around it is a chosen margin, not an interval published by the source. That is why the calculator calls the outcome an estimate.
Number of AI tools
What the calculator shows: an estimate of the number of distinct AI tools running in your organisation, scaled to your size from a measured average of 6 to 8 per organisation.
Source: Netskope, Cloud and Threat Report: 2026.
What was measured: the average number of distinct AI apps an organisation sends traffic to. The measure has broadened sharply in a year: more than 1,600 AI apps are now tracked, against 317 a year earlier.
Why the tail matters more than the average: at the top one percent of organisations the number runs to 89. These are organisations that thought they had control and did not. The average reassures; the tail is the warning.
Confidence: high. Note: this counts AI apps, not every application with an AI feature. Under a broader definition the number is higher.
Prompt volume
What the calculator shows: an estimate of the number of prompts your organisation sends per year, scaled to your size from a measured median of 18,000 prompts per organisation per month.
Source: Netskope, Cloud and Threat Report: 2026.
What was measured: the median number of prompts to AI apps per organisation per month, up from 3,000 a year earlier. At the top 25 percent of organisations this exceeds 70,000 per month, at the top one percent 1,400,000.
A deliberate choice: we calculate from this per-organisation figure, not from an assumption about prompts per employee per day. Such a per-person figure does circulate, but cannot be verified against a primary source. So we do not use it.
Confidence: high for the per-organisation metric.
Scaling to your size
What the calculator does: the tool count and the prompt volume are measured per organisation, not per employee, so both have to be scaled to your headcount. We do that with a damped size factor: (your employees ÷ 1,000) to the power 0.35. At 1,000 employees the factor is exactly 1, so the measured figures are shown unchanged. At the calculator's upper limit of 10,000 employees the factor is 2.2, not 10.
Why damped and not linear: a company of 10,000 people does not run 10 times as many AI tools as a company of 1,000. Tools get shared, and sprawl grows more slowly than headcount. Zylo measured spend growing faster than proportionally at very large organisations, but not the counts.
Confidence: both numbers in this formula are ours. The exponent 0.35 is our judgement of how strongly the curve should flatten, and 1,000 employees is the size we chose to anchor it to. No source publishes either. The anchor matters as much as the exponent: below 1,000 employees the factor drops under 1, so a smaller organisation is shown fewer tools and fewer prompts than the measured average, and a larger one more.
The other parameters we chose, not measured: the 10% to 25% band around the share of AI users, and the usage-profile factors in the CO2 outcome. Both are flagged in their own sections above. We list them together here so that the chosen parts of the model can be counted in one place. Everything else on this page is a figure a source published. The share of AI users and the spend per employee are deliberately not scaled by size at all: no source supports doing so.
Spend
What the calculator shows: the distribution of what organisations spend per employee per month on AI tools, with a median of €10.47, a top ten percent of €562.12, and a top one percent of €6,853.08.
Source: Ramp AI Index, June 2026, based on Card and bill-pay data from more than 70,000 US companies.
What was measured: card and bill payments by companies to AI tools, subscriptions and tokens. These are US companies, and the amounts are converted from dollars to euro at the rate of 30 July 2026. The measurement excludes infrastructure, in-house development and salaries. European figures may differ.
Per employee means per employee on the payroll, not per employee who uses AI. That is how the source divides it, so the annual total in the calculator multiplies the monthly amount by your full headcount, not by the smaller number in the first outcome. Expressed per AI user the amount per person would be several times higher and the annual total exactly the same.
How the annual total is calculated: the median amount per employee per month, times twelve, times your headcount. Nothing else enters it. The amount is shown to the cent for that reason: so you can reproduce the annual figure with a calculator and get the same answer we do. The calculator shows one figure at the median rather than a band up to the top ten percent, because a band spanning more than fifty times tells a reader nothing they can act on. The spread is still there, one screen up, per employee per month, which is where it can be compared against something.
Why one source and no mix: the distribution, from median to top one percent, comes entirely from this single study. There is a second, often-cited source, Federal Reserve Bank of Atlanta, Survey of Business Uncertainty, May 2026, which arrives at €1,903 per employee per year. But it measures something different, namely total AI investment including infrastructure, and may therefore not be placed on the same distribution. We mention it here as context, not in the chart.
Why the spread is the point: the gap between the median and the top one percent is a factor of hundreds. That is not measurement error, but because the distribution is extremely skewed. A small number of organisations spend enormously, most very little. An average therefore says little about your situation. Only a measurement can tell you where you sit.
Confidence: high for the individual figures. Note the geographic and scope limitation: US companies, tool and token spend only.
CO2 emissions
What the calculator shows: a range for the annual CO2 emissions of your AI use, calculated from your estimated prompt volume times an emission factor that sits between 0.03 and 1.14 grams per prompt, adjusted for your chosen usage profile.
Sources: the lower bound comes from Google, Measuring the environmental impact of delivering AI at Google Scale, Aug 2025, the upper bound from Mistral, life-cycle assessment of Mistral Large 2, July 2025.
Why a range and not a number: this is the most important explanation on this page. Published emissions per prompt differ by a factor of tens between the lowest and highest measurement. That difference comes partly from one model being much larger than another, and partly from each party measuring differently and counting different things. Putting a single number on that suggests a precision that does not exist. So we show a margin, and nothing more.
Why the usage profile matters: an organisation that mostly rewrites text consumes far less per prompt than one that generates code or analyses documents. Token-heavy and reasoning tasks consume a multiple. The profile you choose shifts the range with it. It is an approximation, not a measurement, which is precisely why a real measurement makes the difference.
What the sources do and do not include: the Mistral analysis is a peer-reviewed life-cycle assessment to recognised standards, but excludes the user's device and calls itself a first approximation. The Google figure is a production median, but is not yet independently audited and benefits from Google's own purchasing of clean energy. Both are the best available primary figures, with those caveats.
Why this is here at all: this volume and these emissions currently appear in no CSRD or accountant report. That is the gap. Not which figure it is, but that you cannot produce it today.
Sector
What the calculator does: if you choose a sector, only the share of AI users adjusts. The other outcomes do not.
Status as of 30 July 2026: the sector selector is present in the calculator but does not adjust the outcome yet. The translation from "share of companies using AI" to "share of employees using AI" is a substantive assumption that has not been validated, and we do not compute a prominent outcome on an unchecked ratio. Until that translation has been thought through separately, or is carried by our own sector measurements, the calculator uses the cross-sector median. The paragraph above describes the behaviour once it is.
Source: Eurostat, AI use in enterprises (isoc_eb_ain2), 2025 / CBS AI monitor.
Why only this one aspect: official statistics measure, per sector, how many companies use AI. They do not measure, per sector, how many tools run, what it costs, or how many prompts are sent. For those axes no defensible sector source exists. So we apply the sector only where we can substantiate it, and leave the rest untouched. A sector choice that changed everything would suggest a precision the public figures do not support.
Confidence: the adoption figures are official statistics. Note: they measure the share of companies, not the share of employees, and the method was revised for the most recent year. So we apply the ratio between sectors, not the absolute figure.
Limitations that apply throughout
Telemetry and surveys differ. The usage and tool figures come from telemetry and measure established behaviour. Surveys come out higher. Both have their place, but they are not interchangeable.
US data where European data is missing. The spend figures were measured at US companies. For Europe no comparable public measurement exists. That is one of the reasons we started measuring ourselves.
No figure is your figure. Everything on this page is an average or a median of other organisations. Your own position in that distribution is exactly what the gap analysis establishes in four weeks.
What we measure ourselves
We run gap analyses at organisations and public bodies across Europe and North America. Those measurements belong to our clients and we do not publish them without consent. Once we have enough measurements to report on them responsibly, we will replace the public assumptions here with our own figures, broken out by sector once each category holds at least five measurements. Until then the estimates rest on public research, as substantiated above.
Corrections
This page was last checked on 30 July 2026. If something is wrong, a report has been updated, or we have misrepresented a figure, email hello@greenpt.ai. We will correct it or remove it, and note the change on this page.