AI inside software that has to run, not inside a demo

The request arrives as a chatbot, the problem is elsewhere

When a company asks for a chatbot, what it almost always wants is for an employee or a customer to find something in ten seconds instead of ten minutes. That is a search problem. Decent search, with filters, dated results and a link to the source, answers better, costs less to maintain and does not make things up.

The other half of these requests is a process problem. A form retyped by hand, a document passing through five inboxes, a business rule applied out of habit. There too, the answer does not start with a language model. It starts with looking at the flow and deleting the steps that no longer have a reason to exist.

There are real cases where conversation is the right interface: when the question is open ended, when the user does not know what to ask for, when answering means combining several sources. Those are worth building. For the rest, you hear it before the quote rather than after.

A retrieval system is only as good as the documents behind it

Retrieval augmented generation has one property that demos hide: it answers with whatever you hand it. If your knowledge base holds two contradictory procedures, one of them three years out of date, the model will repeat the contradiction with exactly the same confidence as everything else.

So the expensive part of this kind of project is neither the model nor the vector store. It is deciding who owns each document, which source wins when two disagree, and by what mechanism a stale version leaves the index. If nobody inside the business will take that on, the project ships a confident and wrong assistant, which is worse than shipping nothing.

In practice we ask for the real corpus before pricing anything. Not a curated sample: the whole thing, with its duplicates, its 2019 revisions and its crooked scanned PDFs. Its condition decides the rest of the project.

Generative output facing customers, and what it requires

Anything generative and customer-facing will eventually produce a sentence the business has to answer for. That is not a risk to eliminate, it is a risk to contain, and it gets decided before launch rather than after the first complaint.

You need a path out to a human, one click away, not buried behind six attempts. You need an evaluation set, meaning a list of real questions with the answers you expect, replayed on every change of prompt, model or corpus. You need to be able to roll back and switch the feature off within minutes, without redeploying half the product.

You also need a record of what was asked, what was answered and from which sources. The day a customer disputes a reply, that record is the only way to find out what actually happened.

Where your data goes

Sending company documents to a third-party model provider is a decision, with contractual and regulatory consequences. It can be the right one. It should not be made by default, in a configuration file, because that was the fastest thing to wire up.

Where it matters, the work stays with you. Our engineers can work inside your tenant, on your infrastructure, under your access controls, with the model your organisation has already approved. The data stays in your environment, but our engineers reach it remotely from Morocco, which is not on the European Commission adequacy list. That access is defined before work starts, not after. The contract runs through JADEV GROUP SRL in Belgium, in euros, under Belgian law.

On security, stated precisely: our practices follow ISO/IEC 27001 principles and are mapped against CyberFundamentals from the Belgian Centre for Cybersecurity, with CyberFundamentals certification targeted for 2026 and the ISO 27001 Stage 2 audit targeted for 2027 with a BELAC-accredited body. We are not certified today, and you will never read otherwise on this site.

Where we are the wrong answer

We are not a research lab. We do not train foundation models and we are not about to start. If what you need is a domain model trained from scratch, or a paper, this is not the place.

We are not a reseller and we have no partner status to put on a slide. If what reassures you in a tender response is a partnership badge with a model provider, other suppliers have one and we do not.

If your need fits an off-the-shelf product, buy the product. A generic support assistant, meeting transcription, writing help: these exist for a few euros per user per month, and commissioning the equivalent costs far more for a worse result. We are for the work that has to plug into your systems, your data and your rules, which off-the-shelf software does not do.

And if your procurement process requires a valid ISO 27001 certificate at signature, we do not clear that filter today. That is a legitimate reason to rule us out.

How it gets built

The technical base is the one we use everywhere else: Azure, .NET, Node, Next.js, React, TypeScript, PostgreSQL. An AI feature is still software, with tests, CI, environments and someone to call when it falls over. What makes it different is that it can fail by producing something plausible, which is why evaluation and logging matter more here than error monitoring alone.

We run an AI-assisted development cycle internally and we write about it publicly. That means people who have seen where these tools hold and where they give way, on their own code, before putting them anywhere near yours.