Responsible AI: our memo

Stage 1 of our engagements ends with a position memo, sourced and dated, in the client's name. This is ours. Seven principles we apply to every project, what AI really costs, and the framework that applies to everyone.

Version of 11 September 2026. Revised with every regulatory or methodological change.

Seven principles, in project order

Numbered because they follow one another: value is decided before the tool, the tool before the model, training before scale.
  1. 01

    Value before the project

    Before writing a line, we establish what the project must deliver: time given back, an error avoided, a shorter lead time. We give an order of magnitude, never a figure we have not measured. If we cannot say why the project deserves to exist, we do not start it, and the diagnostic says so.
  2. 02

    AI only where it is needed

    Many problems are solved by a rule, a search or a script. They are leaner, more predictable and cheaper to maintain. We only use a model where a script is not enough, and we write that decision into the project file.
  3. 03

    The right model for the right task, no more, no less

    When a model is needed, we take the smallest one that does the job properly. A large model on a narrow task wastes energy and money; a model that is too small for a complex task gets it wrong. The choice is made on trials with your data, not on the supplier's reputation.
  4. 04

    Security by design

    Whatever commits the company, a price, an availability, a commitment, is computed and checked server side, never produced by a model. The agent answers from a closed corpus and cites its sources. Your data is never used to train a model. When it must not leave, we run open models on local infrastructure. And the model call stays abstract: no project depends on a single supplier.
  5. 05

    People validate, operate, and do better

    The tool takes the repetitive actions; people keep validation, judgement and relationships. They do not endure the tool, they operate it: they review what commits the company, correct what goes off track, and win back the time to do properly what they used to rush. An agent always introduces itself as an agent and says who to turn to.
  6. 06

    Measure every workflow, every agent, every operation

    Every workflow and every agent has its measurement: tokens used, cost, response time, carbon estimate with the factor used and its date. Every technical operation too, a scheduled job as much as a model call. What we do not measure, we do not claim.
  7. 07

    Train before scaling

    A tool rolled out widely to people who do not know what it does creates more errors than it prevents. We train first on good practice: framing a use, writing instructions that hold, knowing what you never hand to a model. We scale afterwards, with the people who learned.

What AI really costs

Every figure has a primary, dated source, listed at the bottom of the page.

485TWh

used by data centres in 2025

Worldwide

Data centres used around 485 TWh of electricity in 2025. The International Energy Agency projects around 950 TWh in 2030, close to 3% of global demand, with AI the main driver of the increase.

4.4%

of the national carbon footprint

In France

Digital accounts for 4.4% of the national carbon footprint, 29.5 Mt CO₂e (ADEME and Arcep, 2022 data). Half comes from devices, 46% from data centres, 4% from networks. The rise over earlier estimates comes mostly from data centres located abroad but used from France.

1.14g CO₂e

for a 400 token response

Per response

The first complete lifecycle analysis of a large model, published by Mistral AI with Carbone 4 and ADEME and peer reviewed, gives 1.14 g CO₂e and 45 mL of water for a 400 token response, excluding users' devices. A small model counts in fractions of that.

?

what nobody measures well yet

What is still hard to measure

Local water use of data centres, the manufacturing footprint of specialised chips, and the real share of training in each response. Any precise claim on these points is an estimate, and we present it as one.

The regulatory framework

GDPR remains the foundation for any personal data. The EU AI Act adds to it, on this timeline.
  1. August 2024The EU AI Act enters into force.In force
  2. February 2025Prohibited practices apply.In force
  3. August 2025Obligations for general purpose AI models apply.In force
  4. July 2026The digital omnibus enters into force and postpones high risk obligations.In force
  5. August 2026Transparency obligations apply: telling people they are talking to an AI.In force
  6. Today
  7. December 2027Obligations for Annex III high risk systems.Upcoming
  8. August 2028Obligations for high risk systems embedded in products (Annex I).Upcoming

Questions we are asked

Doesn't AI use an enormous amount of energy? Worldwide, yes, and it is rising fast. For a properly sized tool, the impact is modest and measurable. We measure it, and we publish the source of every factor.
Where does our data go? Where the contract says, and it is never used to train a model. When it must not leave, the model runs on local infrastructure.
Will it replace our teams? No. The tool takes what wastes everyone's time, and your teams keep validation, judgement and relationships.
What if the tool gets it wrong? It answers from your content, cites its source, and says when it does not know. Whatever commits the company is computed server side, not written by the model.
Why not put AI everywhere? Because a rule or a script often does better, more leanly and more reliably. We only use a model where it brings something a script cannot.

Sources

  1. European Commission, regulatory framework for AI (AI Act and digital omnibus), accessed 11 September 2026. digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
  2. Regulation (EU) 2024/1689 on artificial intelligence. eur-lex.europa.eu/eli/reg/2024/1689/oj
  3. International Energy Agency, Key Questions on Energy and AI. www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary
  4. Mistral AI, Our contribution to a global environmental standard for AI, 22 July 2025. mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai
  5. ADEME and Arcep, updated assessment of the environmental impact of digital in France, January 2025. ecoresponsable.numerique.gouv.fr/actualites/actualisation-ademe-impact/

Want one of your own?

The position memo is the deliverable of our first stage: three sessions, a document in your name, and the written list of what you will not do.

Get in touch See the stages