What is a RAG and what does it actually do with a corpus of published content?

A RAG does not train any model: the corpus is chunked, vectorised and indexed. When a visitor asks a question, the system retrieves the relevant passages from what has been published, and the language model formulates its answer from them rather than from its general memory. The index therefore traces the perimeter of what the tool can assert.

Why is a coherent corpus considered the most difficult proof of expertise to falsify?

The article argues that coherence across positions taken, trade-offs defended and vocabulary held from one topic to the next cannot be manufactured overnight. A year of consistent publications is a heritage that visitors arriving today would otherwise never perceive as a whole. Making it queryable in a single request turns that coherence into an observable property.

What risks arise when an AI assistant is deployed without a bounded reference corpus?

An assistant that improvises on security or on the AI Act is a risk, whereas one grounded in a reference corpus written by an identifiable and accountable human is an interlocutor of choice. The article also warns that deploying this kind of interface without the right competences mostly manufactures technical, legal and reputational debt. A bot that converses with prospects commits your word.

What distinguishes a truly intelligent interface from a customer-service chatbot reciting a FAQ?

A chatbot reciting a FAQ on a loop has nothing behind it, which is a purely mechanical flaw. An interface becomes intelligent when proven human expertise feeds the index and when a capacity to act extends the answer: qualifying a need, guiding, preparing the next step. Without the first you get fluent noise, without the second a chatty search engine.

How does the demonstrator translate into an offer for organisations sitting on their own internal corpus?

What a visitor handles while talking with the interface is exactly what 199A Consulting proposes to build for a business: internal procedures, product documentation, lessons learned, training offers, compliance frameworks and business knowledge that today sleeps in shared folders and in the heads of two or three key people. The demonstration runs on a real case, so the prospect judges on the evidence. The same logic can apply to a sales team, a support desk or a compliance department.

How can 199A Consulting help an organisation frame and execute on this topic?

199A Consulting acts as guarantor, vector and activator: guarantor because the expertise injected into the tool rests on real practice in software architecture, security and compliance, from ANSSI to the GDPR by way of the AI Act; vector because human competence must reach the interface without loss, since a good corpus that is poorly chunked or poorly framed produces a bad bot; activator because the goal is to give you control of your own tool, using open-source and independent tooling wherever possible. With more than 20 years of experience, the firm covers strategic framing (audit, governance, risks) and execution (architecture, build, integration, training) in a logic of digital sovereignty. Contact: business@199a.agency

IA .md 17 views Published on 07/10/2026 By JAS 7 min read

FR EN ZH ES DE

Listen

Demonstrating expertise

When a year of published articles becomes an interface that answers, qualifies and moves things forward.

A showcase website, as a communicating asset, is a brochure waiting to be read. Let us be honest: nobody reads it to the end, if anyone reads any of it at all. It asserts, it demonstrates nothing, and the busy prospect has no reason to take it at its word. Attempts to express technical depth through a profusion of graphic effects, of more or less happy animations or, on the contrary, through a hollow minimalism, will only ever reach an audience confined to the cultural or commercial sphere it occupies. The falls are dizzying: the audience does not exist, because a mosaic only holds together when the message carries cultural consensus. But that culture, in the European perimeter, lost its way along the road. Back to the point of my demonstration.

For more than a year, I deliberately built a mechanism of analysis and production, with rigour and tenacity, on 199A.agency: publishing, article after article, expert content. Digital sovereignty, AI governance, cybersecurity, compliance, the dynamics of the telecoms market, but also the place of humans and of intellect inside a managerial and technological machinery that now outruns them in speed and capacity. Texts grounded in what I practise in the field, consistent. What I had not seen coming is what this corpus would naturally become: raw material for designing its own formal extinction.

A corpus must be an asset, not a cold archive

A year of publications is a heritage that nobody steers. Each article is read once, sometimes shared, then it slips into the flow of time. The visitor who arrives today will never perceive the coherence of the whole: the positions taken, the trade-offs defended, the vocabulary held from one topic to the next. Yet this coherence is the most difficult proof of expertise to falsify, because it cannot be manufactured overnight. So I decided to make it queryable and traversable in a single request, able to synthesise a long and evolving body of thought.

A RAG learns nothing: it retrieves, augments itself and filters

Let us step down one technical level, because the word "training" circulates widely and misleads just as much. My corpus trains no model. It is chunked, vectorised, indexed. When a visitor asks a question, the system retrieves the relevant passages from what I have published, and the language model formulates its answer from them rather than from its general memory. A RAG does not make the bot more talkative, it makes it more bounded, in the noble sense of the term. The index traces the perimeter of what the tool can assert: it does not speak about everything, it speaks about what I have already argued, signed and published. The filtering, the securing and the objectivation of the tool remain in the designer's hands and become, in fact, the activation engine and, in a way, the revealer of the quality of the whole production.

For a prospect, this constraint is worth all commercial promises. An assistant that improvises on security or on the AI Act is a risk. An assistant grounded in a reference corpus, written by an identifiable and accountable human, is an interlocutor of choice.

The medium becomes (finally) the message

Here is the shift that interests me and that, I suppose, may speak to you. On a classic website, I write to you that I know how to connect artificial intelligence, architectural logic, regulatory rigour and business stakes. You have to take my word for it. With this interface, I no longer have to write it: you observe it. You set out your situation, the tool frames your topic, places it back in its technical and regulatory context, asks the questions a consultant would ask, and returns a reading that holds together because a year of coherent positions supports it.

The tool becomes a demonstration the moment it takes the website's place: the visitor is not offered a promise, they arrive on a working system. Content has become action, and action is the proof. Coherence stops being a sales argument and becomes an observable property of the exchange. Each answer is a sample of how I reason, available at any hour, with no intermediary between you and the expertise.

What makes the interface truly "intelligent"

We have all encountered those customer-service chatbots reciting a FAQ on a loop. Their flaw is purely mechanical: there is nothing behind them. An interface becomes intelligent when two conditions meet. First, proven human expertise feeding the index. Second, a capacity to act that extends the answer: qualifying a need, guiding, preparing the next step of the exchange. Without the first, you get fluent noise. Without the second, you get a chatty search engine. The alloy of the two produces what the prospect feels immediately: they are listened to, framed, moved forward.

From demonstrator to offer

Here is the part that concerns decision-makers. What you handle while talking with my interface is exactly what I propose to build for your business. You too are sitting on a corpus: internal procedures, product documentation, lessons learned, training offers, compliance frameworks, and years of business knowledge that nobody can find when it is needed. Today, that capital sleeps in shared folders and in the heads of two or three key people.

The demonstration requires no leap of imagination: the tool runs, on a real case, mine. You judge on the evidence. The commercial reasoning fits in one sentence: a prospect who has tested the tool on my expertise no longer needs anyone to explain what they could do with their own. If the expertise of an independent consultant can become an interface that answers, qualifies and moves prospects forward, then so can that of your sales team, your support desk or your compliance department.

On one condition: that someone holds technique, logic and business together. An expert, in short.

Guarantor, vector, activator

That is the place of 199A Consulting.

Guarantor, because the expertise injected into the tool rests on real practice in software architecture, security and compliance, from ANSSI to the GDPR by way of the AI Act.
Vector, because human competence must reach the interface without loss: a good corpus that is poorly chunked or poorly framed produces a bad bot, and the quality of an answer plays out in technical choices that only a fine understanding of the business can arbitrate.
Activator, finally, because the goal is not to make you dependent on a black box, but to give you control of your own tool, using open-source and independent tooling wherever possible.

Those who deploy this kind of interface without that triple competence mostly manufacture debt: technical, legal and reputational. A bot that converses with your prospects commits your word. It had better be built by someone who measures what that means.

Making people want to try

I am not asking you to believe me; that is the whole point of the approach. Start the conversation, ask the question that really occupies you, the one you are stuck on right now, and watch how the tool handles it. Then imagine the same interface, fed with everything your company knows, made available to your own clients at any hour.

The showcase of discourse has had its day. Expertise, through an augmented medium, can finally express itself. It is up to you to appreciate its quality before wanting to make it your own.

The interface is live: 199a.agency.