ID Interactive agency logo

Local AI

Local AI: the complete guide for businesses that care about their data

Local AI is artificial intelligence that runs on infrastructure you control, on your own premises or those of a trusted partner: the data it processes never leaves your business. This guide explains why we deployed our own and what it means for you.

Why our agency invested in its own artificial intelligence data centre, and what it means for you.

Your strategic data, contracts, competitor analysis and sensitive client information: whenever a consumer AI tool processes this content, it passes through servers you do not control, in countries whose data rules may differ from yours.

IA locale et souveraine en entreprise


For most of the business leaders we meet, this risk has become a real barrier to adopting AI. How can you benefit from its promised productivity without giving up control of what makes your business valuable?

Local AI is artificial intelligence deployed in a private environment, capable of using a company's internal data without sending it to external cloud platforms by default.

Local AI offers a credible answer. At ID Interactive, we chose to build our own artificial intelligence infrastructure, hosted in our Vannes offices, operated by our team and used every day to serve our 100 clients without any of their data leaving our premises.

This guide explains why we made this choice, how we put it into practice and what it means for you.

Why does local AI matter in 2026?

2024 was the year businesses adopted ChatGPT and its competitors on a massive scale. 2025 brought a wake-up call: most organisations that had encouraged the use of consumer AI began to realise what they were sending, often unknowingly, to servers abroad.

Accidentally pasted passwords, contract extracts, client files, strategic presentations and financial data: according to several recent studies, nearly one in three employees uses a public cloud AI tool without approval from their IT department, and almost half of these uses involve data that should never have left the business.

Alongside this operational reality come two regulatory pressures: the already familiar GDPR and the EU AI Act, which has progressively come into force and requires more extensive documentation of processing activities and strict oversight of automated decisions.

Against this backdrop, three main approaches to using AI currently coexist in business:

  • Public cloud (OpenAI, Anthropic, Google, Microsoft): the most powerful models, available through a simple internet connection. Maximum performance, but data travels through, and may be stored in, locations outside the European Union.
  • Private cloud: the same models, but hosted in dedicated environments, generally expensive and reserved for large businesses.
  • Local AI: an open-source model running on hardware you control, on your premises or those of a trusted partner. No outbound calls and no data sent to third parties.
  • Hybrid AI : often the best compromise. The business uses a local foundation for sensitive tasks while retaining controlled access to selected cloud models for tasks that need more power.
At ID Interactive, we favour a hybrid approach: a sovereign local foundation for sensitive data, supplemented where needed by cloud models accessed via APIs and governed through a single interface.

Local AI is no longer confined to the laboratory. Since 2024, open-source models have caught up with leading cloud providers in quality across many use cases, while becoming capable of running on reasonably sized professional hardware. What was impossible two years ago is now within reach of an ambitious SME — or the agency advising it.

What 'local AI' really means

The term 'local AI' is often used imprecisely. It actually covers two distinct concepts that are important to keep separate.

An open-source model is an artificial intelligence model whose 'weights' — the result of its training — are freely published. You can download, inspect and modify it. Mistral, Qwen, Llama and Gemma are well-known examples. But downloading an open-source model does not mean it runs locally: you can just as easily use it through an API hosted in a public cloud.

Locally hosted AI is AI that runs on hardware you physically control. It may use an open-source model — as is almost always the case — or one developed in-house. The deciding factor is that the data it processes never leaves your infrastructure.

Truly local AI therefore combines two characteristics: a model whose operation you understand and control, and infrastructure whose sovereignty you can guarantee. This is the combination we have put in place, and what we refer to as 'local AI' throughout this guide.

In practice, this is what happens when one of our consultants asks our AI a question:

  1. The question and any attached documents are sent from the consultant's workstation
  2. They are processed by our AI server, located in our Vannes offices
  3. The model generates a response on that same server
  4. The response is returned to the consultant
  5. No calls, fragments of data or metadata leave our network

By comparison, doing the same with ChatGPT would involve sending your question, documents and conversation context to OpenAI's servers in the United States, where they would be processed by its models, potentially used to improve them and stored according to its own retention policies.

The 4 reasons we chose local AI

Confidentiality

Our work involves serving businesses that entrust us with sensitive information: competitive positioning analysis, growth strategies, sales performance data and, sometimes, intellectual property or unpublished content. Most of our clients are SMEs in western France, and for many of them this data is a vital part of their competitive advantage.

Sending this information to a third-party cloud service, even with 'do not use for training' options enabled, raises two issues. First, we could not fully uphold our confidentiality commitments to clients, because the contractual chain depends on a provider over which we have no control. Second, the industry's track record has shown that data retention and usage policies can change when terms and conditions are updated.

With local AI, the promise becomes simple, verifiable and enforceable by contract: no client data leaves ID Interactive's premises.

Sovereignty

The vast majority of cloud AI models are operated by non-European providers: US companies in the case of OpenAI, Anthropic, Google and Microsoft, and Chinese companies for some alternatives. This means data flows through jurisdictions whose rules on disclosure to authorities, retention requirements and the extraterritorial application of laws differ from ours.

For a French SME, this is not an abstract geopolitical issue: it is a very practical guarantee that commercial analysis concerning one of your competitors will never end up on the desk of a foreign economic intelligence service, and that your strategic communications will remain within the scope of French and European law.

Sovereign cloud services from SecNumCloud-certified European providers partly address this concern. Local AI addresses it in full.

Zero marginal cost

Cloud AI usage is billed by volume: for each request, you pay according to the number of words processed. For light use, the cost is negligible. For intensive, recurring use — typically automating repetitive tasks across an entire client portfolio — the bill can quickly reach several thousand euros a month.

Local infrastructure requires a substantial initial investment, but once that cost has been amortised, each additional request has a marginal cost close to zero. The economics shift in favour of local AI as soon as volumes become significant and predictable.

This economy of scale is one reason we can now offer clients deliverables that previously required hours of manual work, without charging for processing on a per-request basis.

Full control

When you depend on a cloud AI provider, you implicitly accept several compromises: your pricing may change overnight, your preferred model may be deprecated, your service may suffer outages entirely beyond your control and the terms of use may be changed unilaterally.

With local infrastructure, these variables are back under our control. We choose when to update a model, adjust its settings or introduce a new version. We guarantee availability in line with our network infrastructure, without depending on a third-party service. And we know exactly what any given use costs.

Our infrastructure: under the bonnet

Our local AI infrastructure is built on four components, all now accessible to an SME or a medium-sized agency.

The hardware at its heart: NVIDIA DGX Spark

The DGX Spark is a compact workstation designed by NVIDIA specifically to run large artificial intelligence models. It has 128 GB of unified memory — memory shared between the processor and graphics hardware — allowing it to handle models that consumer computers cannot accommodate.

In practice, this 128 GB of memory lets us run models with up to 70 billion parameters, comfortably covering the needs of an agency like ours. To put that into perspective, it offers the hardware equivalent of several professional graphics cards in a form factor that fits on the corner of a desk and uses as much electricity as a standard desktop computer.

Schéma organisationnel de l'IA locale à l'agence


The inference engine: Ollama

Ollama is open-source software that runs AI models on the hardware. It loads the model's 'weights' into memory, receives requests and generates responses. It is the local equivalent of what OpenAI does on its own servers when you ask ChatGPT a question.

Ollama's advantage is its simplicity: it runs with a single command, supports dozens of open-source models and exposes a standard API that most tools can query.

The user interface: OpenWebUI

OpenWebUI is the chat interface our consultants use to interact with the AI. Visually, it resembles ChatGPT: an input field, a conversation history, file attachments and 'folders' to organise topics.

The difference is that the entire interface runs on our infrastructure, communicates only with our Ollama instance and stores conversations on our drives. No data is ever sent to a cloud service.

The model: Qwen3-30B-A3B

Qwen3 is a family of open-source AI models released by Alibaba Cloud under the Apache 2.0 licence — a permissive licence allowing unconditional commercial use. The model we use, Qwen3-30B-A3B, is based on a mixture-of-experts (MoE) architecture: 30 billion parameters in total, with around 3 billion activated for each request. It performs particularly well in French, structured reasoning and following complex instructions.

We selected the model following a systematic comparison of several candidates (Mistral Small, Gemma, Llama and DeepSeek). Qwen3-30B-A3B stood out for the quality of its analysis of structured data and the fluency of its writing in French.

Importantly, this choice is not set in stone. The open-source model ecosystem is evolving rapidly, and our architecture is designed to let us switch to a new model within hours if a more suitable alternative emerges.

Local AI vs cloud AI: a comparison

We believe local AI is the right answer for a large proportion of business use cases. But we do not subscribe to the idea that local is always better than cloud. Here is our comparison.

Criterion - Local AI - Public cloud AI

Data confidentiality ★★★★★ - ★★

Sovereignty - ★★★★★ - ★

Cost for intensive use - ★★★★★ - ★★

Initial investment cost - ★- ★★★★★

Raw performance (frontier models) - ★★★ - ★★★★★

Multimodal capabilities (images, voice) - ★★- ★★★★★

Automatic updates - ★★ - ★★★★★

24/7 availability - ★★★★ - ★★★★★

Operational control - ★★★★★ - ★

Cloud AI still leads in several areas. Frontier models — the most powerful models available at any given time — are almost always released in the cloud first. Advanced multimodal capabilities, including detailed image analysis, video processing and conversational voice, are more mature in the cloud. And for occasional, exploratory use, avoiding an infrastructure investment is an obvious advantage.

Local AI, however, comes out ahead once data, recurring costs and control become central considerations. For a business that regularly handles sensitive content, wants to integrate AI into its business processes for the long term, or already has significant usage volumes, local AI becomes the better economic and strategic choice.

Our recommendation for most SMEs is a hybrid approach: local AI handles regular tasks and sensitive data, while cloud AI remains available for occasional use cases that justify it.

Our first live use case: automating SEO reports

The infrastructure described above is not just theoretical. It is now central to our day-to-day delivery, and our first large-scale deployment focused on our SEO ranking reports.

Before local AI

Each month, our SEO team produces a monthly Google ranking report for our 100 clients. The report analyses the keywords each client ranks for, identifies gains and losses and provides commentary ready to send.

This task took around 30 minutes per client, amounting to more than 240 hours a year spent on repetitive, time-consuming work — time our consultants could not devote to higher-value strategy or optimisation.

The reliability challenge

Automating this work with AI may seem obvious. In reality, it is more complex than it appears: a client report cannot tolerate incorrect figures. If the AI invents a ranking or gets a keyword count wrong, the agency's reputation suffers.

By default, however, language models are poor at arithmetic. They excel at writing, rephrasing and structuring information, but are notoriously unreliable at counting.

Our solution

The key to our workflow can be summed up in one sentence: we never ask the model to calculate. An in-house Python script extracts the key metrics from the Semrush report, checks every figure and produces a structured data file. This file then serves as the model's non-negotiable 'source of truth': its only job is to write commentary based on figures that have already been validated.

The model cannot invent figures, because it no longer calculates any.

The results

After several weeks of refinement, the workflow went live across our entire client portfolio. The figures speak for themselves:

Metric - Before - After

  • Writing time per report: 30 minutes - 3 minutes
  • Monthly time for 100 clients: 50 hours - 5 hours
  • Accuracy of figures: Variable (human error) - 100% (Python calculations)
  • Data confidentiality: Complete (manual) - Complete (100% local)

Above all, the 240 hours freed up each year are reinvested where they really matter to our clients: in-depth competitor analysis, strategic recommendations and technical optimisation of their websites.

For readers interested in the technical details, we have published a full account of this project: model selection, configurations, problems encountered and the solutions adopted (read the article)

Other use cases we are deploying

Automating SEO reports is only a starting point. We are putting the same infrastructure to work on other use cases, both internally and for our clients.

Google Ads campaign analysis. Using performance exports, local AI identifies changes from the previous month, spots underperforming ad groups and suggests potential optimisations. The consultant retains control over decisions, but the initial diagnosis is produced in minutes.

Support for writing client content. When creating SEO-optimised content (product pages, blog articles and category pages), local AI helps writers with outlines, rewording and alternative versions — always based on briefs and data that remain within the agency.

Meeting summaries. Working from a transcribed recording, the AI produces a structured summary covering decisions made, actions required and the people responsible. Once again, both the transcript and the summary remain entirely within our infrastructure.

Competitor monitoring. Local AI is used to quickly summarise large volumes of public content (competitor websites, news articles and industry publications) and extract the information relevant to our clients.

Handling recurring client enquiries. For the routine technical questions our project managers receive, an assistant powered by our internal knowledge base suggests an initial response within seconds, for the consultant to validate and refine.

This range of use cases justifies the initial investment: rather than being dedicated to a single task, the infrastructure is gradually becoming the technical foundation for several distinct productivity gains.

What this means for our clients

Everything above describes how we work behind the scenes. Here is what it delivers for clients who entrust us with their digital challenges.

A confidentiality guarantee backed by contract. When we commit to keeping your data away from third-party services, it is not a marketing promise: it is a verifiable technical reality that we can set out in a contractual appendix and, if you wish, demonstrate with a tour of our infrastructure.

Better quality and turnaround for recurring deliverables. The hours our consultants no longer spend on repetitive tasks are hours they can devote to your strategy, competitor analysis and website improvements. You are not paying for AI: you are paying for a more comprehensive, faster and more consistent service.

Real control over costs. The billing model used by cloud AI tools makes integration costs very difficult to forecast. Our local infrastructure, by contrast, has a fixed, amortised cost. We can include automation in our services without charging extra for every request.

An agency in control of its tools. For us, AI is not a black box bought from a supplier: we deploy it, operate it and adapt it to our needs. This enables us to offer bespoke automation where other agencies might simply recommend an off-the-shelf tool.

A partner that treats AI as a business project, not a gimmick. For an SME, AI is not a question of tools: it is about processes, data governance and upskilling teams. Our local approach requires us to address these issues in depth, and that is precisely the perspective we bring to clients embarking on their own AI transformation.

Let's discuss your AI project

If local AI is relevant to you — whether you are looking for a digital agency that can work with your data without exposing it, or considering deploying AI infrastructure within your own business — we would be delighted to talk.

We offer two starting points:

A free, 30-minute AI and data sovereignty audit. By the end of the conversation, you will have a clear view of the AI use cases relevant to your business, the sovereignty requirements that apply to your sector and the next three or four actions to take. No obligation and no commitment.

The full case study of our SEO automation project. A detailed document covering every stage of implementation, mistakes made, fixes applied and the final architecture — useful if you want to understand exactly what a live local AI project involves.



ID Interactive is a digital agency based in Vannes, Brittany. It supports more than 300 businesses and local authorities, mainly in western France, with digital visibility, websites and SEO, and, since 2025, with sovereign artificial intelligence projects.

Frequently asked questions

ID Interactive, agence digitale installée près de Vannes (Morbihan) depuis 2006, opère sa propre IA locale : une station NVIDIA DGX Spark hébergée dans ses locaux, un modèle open source Qwen paramétré par l'équipe, Ollama et OpenWebUI. Aucune donnée client ne quitte l'agence. Cette infrastructure produit déjà en production les rapports SEO mensuels de ses clients. Pour les entreprises qui envisagent la démarche, ID Interactive propose un audit IA et souveraineté gratuit de 30 minutes : cas d'usage pertinents, exigences de confidentialité et prochaines étapes.

Cela dépend des tâches. Sur les usages les plus avancés (raisonnement complexe à plusieurs étapes, traitement multimodal, génération créative haut de gamme), les modèles cloud les plus récents gardent une avance. Sur la grande majorité des tâches d'entreprise (synthèse de documents, rédaction structurée, analyse de données tabulaires, classification, extraction d'informations), un bon modèle open source de 30 à 70 milliards de paramètres suffit largement. C'est ce qu'ID Interactive constate en production sur les rapports SEO de ses clients.

Oui, et l'agence l'assume. Pour des usages exploratoires sur des sujets non sensibles, ou pour bénéficier ponctuellement de capacités qu'aucun modèle open source ne couvre encore, ID Interactive utilise des outils cloud. Ces usages sont strictement isolés des données clients : un consultant qui utilise un outil cloud pour un client le fait sur des données anonymisées, ou pas du tout.

Trois mécanismes sont en place chez ID Interactive : l'isolation réseau entre le serveur d'IA et internet (les seules connexions sortantes sont contrôlées et journalisées), l'absence de toute clé d'API vers un service tiers dans les outils qui manipulent des données clients, et un audit interne mensuel des journaux d'accès. L'agence peut documenter cette chaîne dans un contrat client si nécessaire.

L'investissement initial pour une station de travail IA professionnelle se situe entre 3 000 € et 6 000 € selon la configuration. S'y ajoutent le temps de mise en œuvre (paramétrage, choix du modèle, intégration aux outils existants) et la maintenance opérationnelle. Pour une PME qui souhaite déployer une infrastructure équivalente à celle d'ID Interactive, l'enveloppe complète de la première année se situe généralement entre 15 000 € et 40 000 €, selon l'ampleur des cas d'usage. ID Interactive accompagne cette démarche pour les entreprises qui le souhaitent.

Plus aujourd'hui. Les évolutions matérielles des deux dernières années ont rendu accessibles des configurations hors de portée d'une PME auparavant : une entreprise de 30 collaborateurs peut s'équiper d'une infrastructure IA locale proche de ce dont disposait une grande entreprise il y a deux ans. Selon ID Interactive, ce qui distingue les projets qui réussissent n'est plus la taille de l'entreprise, mais la clarté du cas d'usage et la qualité de l'accompagnement.

Une station IA locale consomme l'équivalent énergétique d'un poste de travail puissant : c'est significatif, mais sans commune mesure avec l'usage cumulé des grands centres de données de l'IA cloud. Ramenée au volume de tâches automatisées, l'empreinte par requête est favorable. Pour ID Interactive, le local n'est pas la solution miracle au coût environnemental de l'IA, mais il est plutôt mieux placé que les alternatives.

Last updated: 19 November 2026 at 10:53