AI is rapidly gaining ground in the corporate world and placing the IT department at the heart of new trade-offs. Which models should be chosen? How can we avoid an proliferation of tools? How can we manage costs, infrastructure and the resources deployed? From AI governance to eco-design and Green IT, discover the tools available to the IT department to continue innovating whilst reducing the IT system’s environmental footprint.
Two in three companies are already using generative AI
The adoption of artificial intelligence has reached a new milestone.
According to a study by the Banque de France published on 3 September 2026, 67 % of French companies with at least 20 employees report using generative AI. The survey, carried out amongst some 7,000 businesses, shows that the phenomenon extends far beyond the major groups
The usage rate has already reached 64 % in companies with 20 to 49 employees and climbs to 83 % in companies with 1,000 or more employees. For the IT department, this widespread adoption is changing the very nature of the subject. AI is becoming part of our everyday tools and, gradually, part of the information system itself
AI is taking hold first in areas where it is easy to implement
However, this rapid uptake does not mean that all applications are already fully developed. Applications remain concentrated on certain tasks.
Among the user companies surveyed by the Banque de France, 85 % use text-generation tools. In 74 %: in many cases, AI is mainly used in support functions : marketing, administration, finance and accountancy, or human resources.
Production activities still account for a minority: only 14 per cent of companies report that these are their main areas of use.
This snapshot is of interest to the IT department. It shows that AI is first adopted in areas where it is straightforward to implement, before being applied to more sensitive or critical processes.
These applications are easy to roll out. Their industrial-scale implementation is less so.
This is precisely where the IT department comes in.
A marketing department is looking for an editorial assistant. The developers are calling for a co-pilot. The customer service team wants to query its document database using natural language. The finance department is trialling automated document analysis. Each of these projects may well be worthwhile. But each also involves additional models, licences, APIs, data and processing.
An IT department’s capacity for innovation is therefore measured as much by what it deploys as by what it chooses not to over-engineer.
In addition to these traditional concerns, there is now the issue of resource consumption.
Behind the prompt lies a very real infrastructure
The often very simple interface of a generative AI easily masks what is happening behind the scenes. Every query draws on computing power, memory, storage, networks and data centres. On a large scale, the impact becomes significant.
Their energy consumption is in addition to that of an information system for which many companies are already seeking to reduce the environmental footprint.
In its opinion published on 22 July 2026, ADEME points out that the digital sector already accounted for 4.4 % of France’s carbon footprint in 2022. It also estimates that data centres consumed 485 TWh of electricity worldwide in 2025 and that this consumption could double by 2030, particularly with the rise of AI.
The location of treatment facilities also matters. According to ADEME, approximately Two-thirds of digital usage in France relies on data centres located abroad, sometimes in countries where electricity generation is much more carbon-intensive than in France.
For an IT department, the criteria for selecting a cloud provider or infrastructure are therefore expanding. Price, availability, security and sovereignty remain essential, but resource consumption and the location of processing are gradually becoming increasingly important.
Energy efficiency is therefore a factor long before the data centre’s electricity bill arrives. It begins at the point of choosing which technology to use for which task.
For an IT department, this has very practical implications: two solutions offering a comparable service may have very different resource requirements.
The most powerful model isn’t necessarily the best choice
Simplicity is also a matter of choosing the right designs.
The research carried out by PEReN on behalf of Arcep, published in May 2026, offers particularly useful insights for CIOs: A model that uses more energy is not necessarily more efficient.
Among the models tested, relatively energy-efficient solutions achieve performance levels comparable to those of models that consume more energy. Architecture also plays a significant role. Models of the Mixture of Experts consumed on average 45 % less than dense models of equivalent size during the tests. 8-bit or 4-bit quantisation, on the other hand, resulted in an average reduction of 39 % of consumption in inference.
The subject therefore extends well beyond the model’s displayed dimensions.
For an IT department, comparing two AI solutions should involve comparing at least four factors: the quality achieved in the real-world use case, the cost, the response time and the resources deployed.
Which level of AI for which need?
Simplicity is, first and foremost, about not using more technology than is necessary.
| Business requirement | Technology to be tested first | Resources | A concrete example | DSI Arbitration |
|---|---|---|---|---|
| Apply a specific rule | Script, SQL, rules engine | ●○○○○ | Check whether a file contains all the required documents | Generative AI offers little value if the rule is deterministic |
| Predicting or detecting an anomaly | Predictive machine learning | ●●○○○ | Anticipating a breakdown, detecting fraud, forecasting sales | Check the quality of the data before increasing the complexity of the model |
| Extract or summarise documents | Small model or specialised model | ●●○○○ | Search for a technical procedure or summarise contracts | Testing a compact model before moving on to a large general-purpose LLM |
| Creating a versatile business assistant | General LLM | ●●●○○ | HR assistant or internal support staff with access to the company’s documentation | Monitor data volume, data sent and cost per query |
| Solving a complex problem in several stages | LLM with reasoning | ●●●●● | Complex technical diagnostics, advanced code generation | Only use logical reasoning when its benefits have been demonstrated |
| Combining text, images, sound or video | Multimodal model | ●●●●○ | Analyse a photograph of a manufacturing defect, accompanied by a comment | Do not send an image or video if text alone is sufficient |
These figures are for guidance only: actual consumption depends on the model, the infrastructure, the volume processed and the number of tokens. But the principle is simple: first assess the level of technology required to achieve the expected quality.
To give you a clearer idea, here are a few examples of well-known models:
Costs and environmental impact of different AI models
| Model | What’s it for? | Indicative cost* | Resources mobilised** | DSI Reflex |
|---|---|---|---|---|
| GPT-5.6 Luna | Classification, literature review, large-scale automation | ≈ 0.0014 $ × 0,31 | 🟢 Low | A good choice when volume matters more than complexity |
| Gemini 3.1 Flash-Lite | Translation, simple data processing, high-volume automation | ≈ 0.0018 $ × 0,40 | 🟢 Low | Worth considering for the automation of frequent, standardised tasks |
| Gemini 3.8 Flash | Business assistant, agents, code, complex workflows | ≈ 0.0045 $ × 1.00 — reference | 🟡 Moderate | Adapt the level of reasoning to the task |
| Claude Haiku 4.5 | Responsive assistants, rapid analysis, generation and automation | ≈ 0.0060 $ × 1,33 | 🟡 Moderate | The best choice when a balance between speed and quality is required |
| Claude Sonnet 5 | Analysis, complex writing, coding, agents and research work | ≈ 0.0120 $ × 2,67 | 🟠 High | To be used only for tasks where the extra quality is beneficial |
| GPT-5.6 Terra | Advanced specialist work, coding, reasoning and complex automation | ≈ 0.0140 $ × 3,11 | 🟠 High | A good intermediate level before moving on to a Frontier model |
| Gemini 3.1 Pro | Advanced reasoning, multimodal analysis and complex agent-based tasks | ≈ 0.0140 $ × 3,11 | 🟠 High | To be used when the depth of analysis justifies the additional cost |
| GPT-5.6 Ground | Advanced reasoning, complex problems, sophisticated agents | ≈ 0.0240 $ × 5,33 | 🔴 Very high | Avoid using it by default for simple tasks |
| Claude Opus 5 | Advanced coding, autonomous agents and highly demanding specialist tasks | ≈ 0.0300 $ × 6,67 | 🔴 Very high | Reserve it for high-value tasks where its superior capabilities are utilised |
| Claude Fable 5 | Long-running processes and workflows requiring a high degree of autonomy | ≈ 0.0600 $ × 13,33 | 🔴 Very high | Best suited to complex processes where autonomy justifies the resources required |
* Example of API costs for 1,000 tokens in + 1,000 tokens out. The multiplier compares each model with the indicative cost of Gemini 3.8 Flash, set to × 1.00.
** The colour code does not represent a CO₂ measurement for each model. It represents a relative level of resources to be allocated, determined on the basis of the model’s positioning, complexity and intended use.
The «reasoning» mode clearly illustrates the risk of oversizing
Models capable of multi-step reasoning, as well as AI agents, are useful for difficult tasks. However, their energy consumption can be significantly higher than that of a model without reasoning capabilities.
In the PEReN tests, the activation of reasoning leads to up to 92 % of additional fuel consumption on average for the models concerned. In a code-generation task, the measured deviation even reaches +849 %. And the improvement in performance varies significantly depending on the task.
It is easy to put this difference into practice within the company.
An engineer who asks an AI to analyse the possible causes of a complex incident may need this capability. An employee who wants to rephrase an email probably does not need it.
When implemented on a large scale, this simple configuration choice becomes a means of achieving energy efficiency.
Text, images, video: the format makes all the difference
Not all queries submitted to an AI are the same. For the IT department, choosing the right model is important; ; Choosing the method is just as important.
The differences are considerable. In its May 2026 report, Arcep states that the available studies estimate that generating an image takes on average, around 60 times more energy than text generation. When it comes to video, the figures are even higher: Arcep cites an estimate from the International Energy Agency of approximately 115 Wh to generate a six-second video.
In a business, this difference can quickly alter the cost-effectiveness of a use case. A marketing department that occasionally produces a few illustrations does not face the same challenge as an e-commerce platform that automatically generates multiple visual variants for tens of thousands of products.
Resolution also matters. The computational costs of image models generally increase with the required size and quality; experimental studies also show that computational costs can vary significantly depending on the resolution and architecture of the model. A study published in 2025, which examined 17 image-generation models, found that up to a factor of 46 between the models tested.
Multimodal agents receive multiple pay rates
The multimodal AI agents introduce another, less visible form of consumption: they often do not make a single inference, but a series of operations.
Let’s take an IT support agent as an example. They receive a ticket, analyse a screenshot, consult the knowledge base, query the ITSM tool, may run a diagnostic command, analyse the results, and then draft a response.
What the user perceives as a single request can therefore trigger multiple calls to the model, image processing, searches and calls to external tools.
Pricing structures already reflect this reality. Google states, for example, that for its agent-based systems, charges are applied not only for input and output tokens but also for the intermediate tokens generated during reasoning loops. As for OpenAI, the tools and delegated workers can also add their own costs to the main model’s tokens.
At present, there is no universal environmental metric for a «multimodal agent request»: two agents may take two or twenty steps to complete the same task. The number of loops therefore becomes, in itself, a measure of simplicity.
From text to video: vastly different orders of magnitude
| AI applications | Example from the workplace | Indicative API cost* | Resources | DSI Reflex |
|---|---|---|---|---|
| 💬 Text | Summarising a document, writing an email | ≈ 0,0014 $ ×1 — reference | 🟢 Low |
Opt for lightweight models for everyday use |
| 🖼️ Image | Marketing visuals, product illustrations | ≈ 0.034 $ per image ≈ ×24 | 🟠 High |
Limit the number of variants and resolution to what is actually required |
| 🎬 Video | Demonstration, presentation, tutorial | ≈ 1 $ / 10 s ≈ ×714 | 🔴 Very high |
Generate only if the video offers real value |
| 🤖 Multimodal agent | Read a ticket, analyse a screenshot, investigate and then take action | Variable × varies depending on the call | 🟠→🔴 Variables ranging from very high to extremely high |
Setting a limit on the loops, tools and models used |
This hierarchy introduces a new way of thinking in AI projects.
A product catalogue may not need to generate five new images every time it is viewed. A technical support team can start by analysing the text of an incident report before asking for a screenshot. An internal training course can use a static illustration when a video does not provide any additional educational benefit.
For staff, a low-key approach can involve rules that are just as practical: limit the maximum number of iterations, reserve reasoning models for complex stages, avoid analysing an attachment when its metadata is sufficient, and cache reusable results.
The question of format is therefore linked to that of choosing a model. The more modalities an AI processes and the more steps it carries out, the more the IT department has a stake in ensuring that this sophistication delivers a genuine business benefit.
Frugal AI does not mean «small AI»
Frugal AI is sometimes reduced to the use of small models. Its principle is broader than that: it seeks to achieve the best balance between the value generated and the resources required.
A larger model may be the right choice if it offers a genuinely greater benefit.
Conversely, a compact model may itself be unnecessary when a search engine, a business rule or a conventional processing method already meets the requirement.
Frugality can also be achieved through less obvious choices: reducing the number of tokens sent, limiting the context size, caching certain results, avoiding repeated identical processing, or filtering data before passing it to the model.
For the IT department, this is ultimately a familiar principle: scale the architecture in line with the expected service level.
Shadow AI also complicates matters
Simplicity also depends on the number of solutions that coexist within the organisation.
One department uses a general-purpose assistant. Another team subscribes to a specialised tool. Developers have several co-pilots. The marketing team is testing image generators. Some staff members are still using personal accounts.
For the IT department, this shadow AI poses a well-known problem in terms of security and governance. It also leads to a dispersal of resources: under-utilised licences, redundant APIs, duplicated data and processing carried out across multiple platforms.
Mapping AI tools can therefore yield a threefold benefit: to minimise risks, streamline expenditure and limit unnecessary use.
Digital frugality is directly linked to good IT governance in this context.
Innovating without multiplying short-lived projects
The findings published by the Banque de France also highlight a key point: the adoption of technology does not automatically guarantee the creation of value.
By early 2026, generative AI was already being used by two out of three French companies with 20 or more employees, but its use often remained at an experimental stage or was limited. The Banque de France believes that the economic focus is now shifting from adoption to AI’s ability to genuinely transform organisational structure and performance.
This development adds a new dimension to simplicity.
Even a proof of concept (POC) that is never rolled out to production has still consumed time, data, computing resources and budget. A responsible IT department can therefore incorporate exit criteria right from the pilot stage: minimum expected quality, number of potential users, cost per transaction, measurable business benefits and the resources required.
Projects that do not meet these thresholds may be terminated early.
Four indicators to monitor from the pilot stage onwards
To put this approach into practice, an AI project can be assessed from its very first weeks across four dimensions:
1. Value. Time actually saved, improved quality, automation rates or user satisfaction.
2. The cost. Cost per query, per case processed or per active user, rather than just the subscription fee.
3. Technical efficiency. Number of tokens, computation time, amount of data transferred, or the rate at which the most powerful models are used.
4. Actual use. Number of active users, frequency of queries and proportion of features actually used.
These indicators make it possible to compare several architectures on the basis of more than just a one-off demonstration.
Making the right power a performance criterion
AI adds a new dimension to architectural decisions. The IT department must continue to ensure security, availability, cost control and service quality. It must now take a closer look at the resources deployed to deliver this service.
Balancing innovation with restraint does not, therefore, mean arbitrarily restricting the uses of AI. It requires selecting use cases that create demonstrable value, comparing models in real-world situations, and reserving the most resource-intensive architectures for the problems that require them.
An IT department can therefore continue to innovate rapidly whilst ensuring that not every new requirement automatically results in a need for more computing power, data and infrastructure.In an IT system where AI is becoming commonplace, achieving the right result with appropriately scaled computing power is itself becoming a performance indicator.
