What is the carbon footprint of an AI sales training session?
A 15-minute sales training session with Marvin.ai emits around 7 grams of CO2e, less than charging a smartphone (8 to 15 g) and far less than one kilometre by car (120 to 200 g). That figure comes from measuring a real production session, not from a theoretical estimate. Marvin.ai trains no model: the heavy carbon cost of AI, training a foundation model, does not enter this usage footprint.
First things first: "AI training" does not mean what you think
When a sustainability team hears "AI training", it pictures the headlines: hundreds of tonnes of CO2 to train a foundation model. That is true, and it is a different subject.
At Marvin.ai, "training" means sales training: a spoken role-play between a rep and a simulated customer. Technically this is inference, the use of an already-trained model.
Marvin.ai trains no model. There is therefore no training carbon cost to charge to usage, only a few minutes of compute per session. This is the first thing to say in a tender that carries an environmental criterion, because it is also the first confusion to clear up.
What we measured, on a real session
Rather than reusing a market average, we instrumented a real production session in July 2026: 14.5 minutes of conversation, around thirty model calls, then generation of the analysis report.
The impact estimate uses EcoLogits (GenAI Impact), the open-source reference library for assessing the footprint of an inference. It combines the number of tokens produced, model size, datacentre efficiency and the carbon intensity of the electricity grid.
| Measured quantity | Value |
|---|---|
| Session length | 14.5 min |
| Estimated energy | around 18 Wh |
| Estimated CO2e (US electricity mix) | around 7 gCO2eq |
The energy consumed does not depend on the country. Only the grid's carbon factor changes, and the gap is wide.
| Electricity mix | CO2e per session |
|---|---|
| France (heavily nuclear) | around 1 g |
| United States (model provider's infrastructure) | around 7 g |
| World average | around 8 g |
One unexpected lesson along the way: it is the spoken conversation, not the report analysis, that weighs most in a session's consumption. That makes sense, real-time audio keeps the model busy continuously, while analysis is a one-off task.
Seven grams, what does that look like?
An isolated figure says nothing. Here are the same orders of magnitude, mapped onto uses everyone knows.
| Use, per person | Estimated footprint | Reference point |
|---|---|---|
| Marvin.ai sales training session, 15 min | 7 to 15 g CO2e | Less than charging a smartphone |
| Online self-assessment or questionnaire, 15 min | 5 to 20 g CO2e | Equivalent, light text usage |
| Social media, 10 min | 25 to 200 g CO2e | Up to 1.5 km by car |
| Video conferencing, 1 h | around 530 g CO2e | Around 3 km by car |
| One classroom training day, consultant travel | 10 to 16 kg CO2e | A 300 km round trip, plus taxi and a meal |
Put differently: ten minutes of scrolling on social media weighs more than a full sales training session. One hour of video conferencing weighs a good fifty of them.
For a group of ten people, a 15-minute Marvin.ai session represents 70 to 150 g of CO2e in total. That is the order of magnitude of a one-kilometre round trip by car, for ten reps trained.
And against a classroom day?
The contrast is sharp, but it has to be read correctly.
The dominant item in a training day is neither the room nor the materials: it is travel. For a consultant who travels (300 km round trip by train, 30 km by taxi, one meal), the total lands around 10 to 16 kg of CO2e, before counting participants' journeys. That is roughly one hundred times a Marvin.ai session for the same group of ten.
This is not a contest, and we do not present it as one. A day in a room with a senior consultant does things no tool does: reading a group, handling resistance, settling a disagreement between a manager and their team. At b-flower, which trains around 6,000 salespeople a year and has done so for 25 years, that is the core of the job and it stays that way. The difference between training and coaching is not settled by software.
What AI training changes is the volume of repetition between sessions. Historically, a rep who wanted to practise fifteen times between two training days needed a trainer, a room and a slot. Today they do it from the car, between two meetings, for a few grams of CO2e per session.
The environmental gain does not come from replacing the classroom. It comes from the repetition which, on its own, did not only cost carbon: it cost so much in logistics that it simply never happened.
Why this is an order of magnitude, not a certified measurement
We would rather show the limits than sell a precision nobody in the industry can offer today.
- The calculation model is driven by tokens produced. It counts what the AI generates, not what it listens to. For a spoken conversation, processing of the incoming audio is therefore underestimated. Our 7 g figure is a floor.
- No public tool models real-time voice models yet. We use a close text model as an approximation, which adds uncertainty.
- Providers publish neither the exact size of their models nor the real energy per request. Every estimate, whatever its source, inherits that grey area. It is structural.
- The actual execution region is unknown. Depending on the electricity mix, the result varies by a factor of ten, as the table above shows.
- We publish no water consumption estimate, for lack of data reliable enough to defend.
Even after generously correcting the underestimation of voice, we stay in the order of a few tens of grams of CO2e per session at most. The conclusion does not move, only its precision is at stake.
What this changes for a sales leader
Three concrete uses for this figure.
Answering a tender with a sustainability criterion
More and more large-account tenders include an environmental clause on training services. Having a sourced order of magnitude, with its method and its limits, beats a statement of intent. It is the same reflex as on the GDPR and AI Act side: show the mechanics rather than promise compliance.
Arbitrating a skills-building programme
The dominant carbon item in a training plan is travel. Any serious thinking about a programme's footprint starts there, not with the choice of digital tool. The reasoning matches that of the training budget and its KPIs: the visible item is almost never the dominant one.
Avoiding declarative greenwashing
Marvin.ai records the real AI consumption of every session. The estimate is therefore reproducible on request, on your own scope, with your own volumes. It is not a marketing argument, it is operational data.
Frequently asked questions
Does Marvin.ai train its own AI models?
No. Marvin.ai uses already-trained language models and runs inference at usage time. The heavy carbon cost of AI, training a foundation model, therefore does not appear in our usage footprint.
How much CO2e does one Marvin.ai sales training session emit?
Around 7 grams of CO2e for a 15-minute session, based on a real production session measured with the EcoLogits method. Depending on the datacentre's electricity mix, the range runs from 1 g (France) to 8 g (world average).
Does conversational AI consume more than a video call?
No, the opposite. One hour of video conferencing emits around 530 g of CO2e according to the French ADEME Base Empreinte, against around 7 g for a 15-minute training session with Marvin.ai. Video weighs far more than text and voice.
Does Marvin.ai replace classroom training?
No. Marvin.ai enables the repetition between classroom sessions, where it was logistically impossible. Human support remains essential for group dynamics, handling resistance and managerial coaching.
How do we get an estimate for our own deployment?
AI consumption is tracked per session. We can produce an estimate on your real training volume, with the assumptions and sources shown. It is a deliverable available on request as part of a pilot.
Sources
- Mistral AI x ADEME environmental impact study (2025), around 1.14 g CO2e per request
- EcoLogits, GenAI Impact, methodology for estimating the impact of LLM inference
- Vert.eco, ecological footprint of generative AI
- Alliance Green IT, impact of a one-hour video call
- ADEME Base Empreinte / Negaoctet, video conferencing
- Statista, carbon footprint of social networks
- CO2e emissions per passenger-kilometre, data.gouv.fr
- Impact CO2, carbon footprint of a meal, ADEME
Need to quantify the impact of a training programme?
A 30-minute demo, including a carbon estimate on your own volumes.
Book my demo