Organisations

A complete solution for the low-carbon transition of AI

CarboneWatch estimates the electricity and CO₂e of your fleet's use of AI services, without reading any content, and helps you reduce it without working any less well. Totals are read per department, never per person: no per-machine view exists, and no setting can create one.

Talk about a pilot The method

Measure to understand. Understand to act. Act to reduce.

1 · Measure

Measure to understand

Each workstation estimates the electricity and CO₂e of its use of AI services: browser, desktop applications, coding assistants, local models.
The application sees the service and the volume exchanged. Never the content, never the addresses visited.
What can be measured on the machine is measured; the rest is estimated, with published bounds. The result is an order of magnitude the organisation can defend.
On the provider side: Oreus, a French AI host, sends us its consumption readings, hour by hour.

2 · Understand

Understand to act

Fleet totals, per day, per service and per kind of use, on an instance of your organisation's own.
A breakdown by team, site or role, from three machines carrying the same label.
The fleet's dispersion shows where to put the effort: the share of the total carried by the most intensive machines, without ever pointing them out.
Exports and a reporting note designed for the carbon inventory, with the file that lets the figure go into it.

3 · Act

Act to reduce

Use less without working any less well: the right tools, the right settings, the right providers.
Support in three steps: an audit of uses, actions chosen with your teams, measurement before and after.
The effect of the actions is separated from that of activity, as the French BEGES method requires: a reduction is demonstrated, not declared.
For management, the audit also looks at where AI spending goes.

What each of them finds

CSR

What can the CSR team get from it?

An estimate of AI use across the whole fleet, in electricity and CO₂e, per day, per service and per kind of use.
A breakdown by team, site or role, from three machines carrying the same label.
CSV exports, one of them in GHG Protocol format, and a printable reporting note.
A proposed inventory line — scope 3, category 1; item 4.5 of the French BEGES — for the person who keeps the inventory to confirm. If it replaces a spend-based estimate, it does not add to it.

How does IT deploy it?

As a supported pilot on a few machines, with us.
Two packages, in this order: the application, then the organisation's configuration.
macOS: application and organisation package signed and notarised by Apple.
No administrator rights needed to run, no open port, signed updates.

DPO

What should the DPO review?

What is sent, per hour and per service: energy, CO₂e, estimated tokens, kind of service; with each upload: electricity grid, version, fleet key, declared segment, how fine the measurement is, the days the application ran, failure codes from a closed list, and a token renewed every month.
What is never sent: IP address, machine name, process names, visited addresses, content.
Legal basis: the organisation's legitimate interest, with a balancing test it writes and keeps. The data are pseudonymised, not anonymous.
One instance isolated per organisation, as you choose: hosted by us in France, or installed on your own servers. The right to object is exercised with your DPO.

Works council and staff

What do the works council and staff see?

No per-machine or per-person view. A segment is only named from three machines.
In this mode, sharing is decided by the organisation: the person equipped is told at first launch, with the detail of what is sent.
Consulting the works council comes before deployment.
What is sent stays readable, word for word, in the machine's settings.

What we do not promise

Precision to the gram

It is an estimate, as an order of magnitude, with published bounds. Not a meter reading.

Automatic compliance

We propose an inventory line; the person who keeps your inventory confirms it. The tool does not replace their judgement.

Every AI service

The application counts neither the AI built into Microsoft 365, nor Gemini, nor Canva's: their addresses are those of the rest of the tool. The browser extension adds Gemini and Copilot in Microsoft 365 on their web pages, on the machines that install it. The list of what is missing is public.

What is not counted →

Start with a pilot

A few machines, supported, with your IT department, your DPO and your works council. The deployment kit and the information notes are handed over before the first machine.

Talk about a pilot