AI in SMEs: 32 straight questions, complete answers.

A private AI is a language model installed inside the company's own infrastructure: its servers, or a sovereign European cloud.

It answers from your own documents, and no question and no document travels to a third-party service.

With a public assistant, every question leaves with its context, and often the document it concerns, to servers the company does not control.

With a private AI, the model is isolated on your side: it reads your documents without anything leaving your walls.

Four moves: capturing every document that comes in (quotes, purchase orders, invoices, proof of delivery), reading it, reconciling its data across documents, and alerting when something does not add up.

The rule fits in one sentence: the machine reads, reconciles and alerts; decisions stay human.

An AI connected to your ERP that prepares your management actions (quotes, orders, entries) on your real data of customers, stock and orders.

Every action goes through human validation before it leaves: the system never runs in full autonomy.

No: our offers are designed for SMEs of 10 to 250 employees, with or without an internal IT department.

You can start with training alone, with no installation and no access to your data.

No. Your teams use a simple interface, the technical complexity is handled in the background, and every deployment includes training.

Yes, almost always as a deployer: the SME uses AI systems, it does not build them.

Its real obligations come in three blocks: demonstrable AI-literacy measures since 2 February 2025, transparency on 2 August 2026 for the minority that exposes a conversational agent or AI-generated content to the public, and prohibited practices that rarely concern day-to-day management.

Article 4 of the EU AI Act (Regulation (EU) 2024/1689) requires AI-literacy measures suited to your uses, and the ability to demonstrate them. It mentions neither mandatory training nor any diploma.

A documented training session, a signed usage policy and an AI-usage register answer it simply.

Companies that expose a conversational agent or AI-generated content to the public: the agent will have to identify itself as such.

A minority of SMEs is affected, and an inventory of your tools establishes whether you are among them.

No text specifically requires a "policy". It is, however, the simplest demonstrable document to produce: a signed page that frames accounts, data and verification.

The full template is published openly on this site, to copy and have signed.

The document that lists who uses what in the company: which tools, on which accounts, with which data.

It is the first deliverable of our four-week audit, and the basis of any compliance work.

Yes, provided the architecture and the legal bases account for it: it is a question of design, not of prohibition.

Personal data requires a framing, and health data a reinforced regime, hosting included.

On French infrastructure (OVH, Scaleway) or directly on your servers, as you choose.

The location is stated in the contract, along with the models used and the exit conditions.

Never: data entrusted to us on an engagement is used to train, tune or evaluate no model, neither for us nor for a third party.

No client data goes to a US API, including during demonstrations.

The engine is demonstrated on public data before it touches yours.

The reversibility clause is written into the contract: models used, location, exit conditions, data in open, exportable formats.

Nothing is locked in: you take the installation back or hand it to whoever you like.

European models by default, of the Mistral type, self-hosted or on a sovereign cloud.

The choice of components follows your situation, not a catalogue: we are paid for the project, never for reselling licences.

The question is settled in a pilot, on your real cases: the scope is chosen narrow and measurable, demonstrated on public data first.

Any extension is decided only on figures measured at your premises, never on a promise.

Hands-on training in generative AI built on your real business cases: prompting, automations, best practices, change management.

No data is touched at this stage, and every session leaves a written operating procedure.

A private AI deployed in your infrastructure: indexing of your documents, augmented search, an interface in your brand.

The model is isolated, nothing leaves your walls, and a maintenance contract keeps the whole thing up to date.

All of Bastion, connected to your ERP: the copilot prepares quotes, orders and entries on your operational data.

Systematic human validation, reinforced governance on sensitive processing, an audit trail.

The ranges for our offers are published, as "from" amounts, on the pricing page of this site.

The exact scope is framed in the free thirty-minute diagnostic, and the audit is quoted individually.

An AI system is not frozen software: models evolve, your data changes, an unmaintained infrastructure degrades.

The three managed-operations packs (Guet, Garde, Rempart) cover supervision, updates and support, under a written service commitment.

Yes: thirty minutes, no commitment. And the online test takes two minutes, with no email.

If the scoping shows the project is not worth it, we tell you, and the file stays yours.

In five stages: Diagnostic, Scoping & compliance, Pilot, Deployment, Managed operations.

Between each, a decision point: the next stage begins only on the proof produced by the previous one, and you decide on figures measured at your premises.

Four weeks, with one deliverable per week: the AI-usage register, an inventory of flows, a measurement of search costs, then a drafted usage policy and a costed action plan.

Everything belongs to you, whether or not you continue.

Yes, at every stage: it is a planned outcome, not a failure. The deliverables of each stage stay yours, usable without us.

Yes, and the roles are set in writing from the first week: they keep the machines, we take the uses.

A dedicated point of contact, from end to end, with a weekly progress update and a reply within 24 business hours.

These service commitments appear in our presentation dossier, and hold for every engagement.

DocAgora, our e-health platform in production in Portugal, ten internal vertical SaaS products, and more than fifty sites and platforms delivered since 2024.

Detailed case studies are available on request, during the diagnostic: we publish no client figure without a baseline measurement and a written agreement.

Our method is aligned with ISO/IEC 42001 and the EU AI Act, and we are a member of the IAPP.

The certifying programmes (PECB ISO/IEC 42001 Lead Implementer, AIGP) are in progress: nothing will be shown as certified before it is.

If your company is established in France, yes: from 1 September 2026, every company established in France must be able to receive electronic invoices, under the reform led by the French tax administration (DGFiP). The obligation to issue follows in September 2027 for SMEs.

A reconciliation circuit installed before the deadline turns the constraint into a production gain.