Where should a company start with AI?

With a specific task that costs hours today and whose result can be checked: answering technical questions, sorting documents, drafting quotes. Not with “adding AI”. Measure what it costs today, build, and measure again.

What happens when the AI gets it wrong?

It's designed to be wrong rarely and, above all, to say when it doesn't know. Answers cite their source, so they can be checked, and actions with consequences —sending, paying, deleting— go through a person. Accuracy is measured against control questions before launch.

What does the EU AI Act require of a company that uses AI?

Since 2 August 2026, telling people they're talking to an AI and marking generated content (Article 50). Since February 2025, training the staff who use it. High-risk obligations were postponed to December 2027. A support assistant usually isn't high-risk, but you need to be able to show it.

We use whichever model the task needs: OpenAI, Anthropic, Google, Mistral, or an open model running in your own infrastructure. The system isn't tied to one provider, because prices and models change every few months.

How much does an AI agent or assistant cost?

Per project, with the scope fixed in writing before we start. The figure comes out of the diagnosis, together with the running cost per query, which is measured and stated: for Micblau it's $0.0018 per answer.

Service — AI for businessAgents, assistants and evaluation

Artificial intelligence you can measure

Agents that carry out tasks in your tools, assistants that answer from your own documentation and say where every answer comes from, and the evaluation that proves whether they're right. In production and measured — not a pilot that never ends.

Terms

Pace: Weekly deliveries, tested on your data from the first one.
Price: Per project, with the scope fixed in writing. Never by the hour.
Data: Yours. If it can't leave the company, it doesn't: we use a local model.

Tell us the use case

In 2026 the usual situation isn't having no AI. It's two or three half-finished pilots, nobody measuring whether they're right and a monthly bill nobody can explain. The technology works; what fails is everything around it: which task exactly, with which data, how you know it's right and what happens when it isn't.

So here an AI system starts with its control questions —the ones it has to get right to be useful— and isn't finished until it gets them right. When it doesn't know an answer, it says so. An assistant that confidently makes things up is worse than none at all.

Who it's for

Technical teams with manuals, data sheets and regulations nobody finds in time.
Customer support answering the same question a hundred times a month — and answering it well.
Back-office teams copying data from PDFs, emails and delivery notes into the ERP.
Companies already using AI that now need to explain what it does, as the EU AI Act requires.

What we build

Agents

That act, not chat. They read an email, check the ERP, draft the quote and leave it for review. With narrow permissions and a log of everything they do.

Assistants

On your own documentation, with retrieval over your documents: they answer with what's written in your data sheets and manuals, and cite the document and year each answer comes from.

Evaluation

Knowing whether it's right, in numbers. A set of control questions, an accuracy rate before launch and monitoring afterwards: AI systems degrade silently when the data or the provider changes.

Cost

The model the task needs, not the priciest one. Almost no task needs the most powerful model on the market. Models are chosen by cost per query, with caching and limits, and measured.

Integration

Where you already work. ERP, CRM, email, your website or intranet. AI earns its place when it shows up in the usual tool, not in one more tab.

Compliance

The EU AI Act, solved by design. Since 2 August 2026, a business using an assistant with customers must disclose that it's AI and mark what it generates (Article 50). We build it that way from the start.

What you get

The system in production, integrated into your tools.
The control questions, with the accuracy rate measured before launch.
The cost per query, measured, and limits so it can't run away.
The documentation the EU AI Act asks for, written while it's built.

Where it shows

Micblau Piscines

Micblau Piscines: a technical assistant over 67 documents and 1,148 indexed chunks that got 10 of 11 control questions right, at $0.0018 per query. Read the case →

What people ask us

Where should a company start with AI?: With a specific task that costs hours today and whose result can be checked: answering technical questions, sorting documents, drafting quotes. Not with “adding AI”. Measure what it costs today, build, and measure again.
What happens when the AI gets it wrong?: It's designed to be wrong rarely and, above all, to say when it doesn't know. Answers cite their source, so they can be checked, and actions with consequences —sending, paying, deleting— go through a person. Accuracy is measured against control questions before launch.
What does the EU AI Act require of a company that uses AI?: Since 2 August 2026, telling people they're talking to an AI and marking generated content (Article 50). Since February 2025, training the staff who use it. High-risk obligations were postponed to December 2027. A support assistant usually isn't high-risk, but you need to be able to show it.
Do you use ChatGPT?: We use whichever model the task needs: OpenAI, Anthropic, Google, Mistral, or an open model running in your own infrastructure. The system isn't tied to one provider, because prices and models change every few months.
How much does an AI agent or assistant cost?: Per project, with the scope fixed in writing before we start. The figure comes out of the diagnosis, together with the running cost per query, which is measured and stated: for Micblau it's $0.0018 per answer.

How it's bought

Every service follows the same path: a short diagnosis, the build with weekly deliveries and, if you want it, ongoing maintenance.

Related services

Let's talk

Tell us what you're building.