Making your company readable by AI. Outside, and in. Operational AI · Proof from our own house
What we run ourselves: our AI systems in production
The best proof of our integration method is not a promise. It is the systems we operate every day, for our own firm and for the sites we serve. All of it is built in-house: proprietary tools, not rebranded subscriptions. Each is declared for what it is, a system of the house, a proof of concept or a prototype demonstrated for a client, and part of them can be verified from the outside. For anyone who wants to open the bonnet, how they are built is described separately.
An editorial pipeline under human control
The business workflow: start from the questions buyers actually ask, produce an article that answers them, publish it in two languages. The AI drafts under written editorial rules. A dozen automatic checks then run, and their independence is not declarative: most are deterministic (links, markup, structured data, similarity against the archive, structural constraints, language tells) and involve no model; the substantive checks compare the text against references external to the writer, a maintained fact base, a citation library and a voice corpus; and the fact check is itself calibrated against a set of hand-verified answers. A failure blocks. A human then reviews, and a veto right precedes every publication. In production since April 2026. Published articles are dated, signed and visible: the examples and their living counters keep the count.
And you can judge on the evidence, one article per site: for the house, for the school (realfrench.co), for the studio (nicolehenusse.be), for the English practice (english-performance.com, in French). Each is dated, signed, and came out of this pipeline.
A client portal in production
Validation notices and publication history do not live in lost emails: they live on a dedicated portal, in production, where the client exercises their veto right and finds every dated publication. It is a complete business workflow, installed end to end: validation happens where the client works, not inside our own tooling.
A voice assistant on the road
The business workflow: the admin of people who spend their days on the road. For Food Collective, a food-service consultancy in Brussels, we operate an assistant driven by voice, on the move, between two appointments: it reads the professional mailbox, summarises, and prepares reply drafts in the tone and context of each thread. It draws on the team’s shared tracking base, which it keeps up to date: if it sees an open task for a client there, it takes it into account, like a real assistant who knows the files. Nothing is sent without a human decision: the assistant prepares, the human signs. It is connected through the vendors’ official access points, and it is a working proof of concept with them, not a finished engagement. The same engagement includes a company context base and a drafting environment; its co-founder speaks about it.
For a client: four years of invoices turned into dashboards
For a client, a farm in the Côtes-d’Armor, about 1.2 million euros of feed invoices over four years became queryable data and cost dashboards, compared against market reference prices. The full circuit was demonstrated end to end, from the photo of the invoice to validated data: the farmer uploads, corrects what the machine misread, confirms, and nothing enters the base without human validation. It is a demonstrated prototype, not yet a daily tool. The exact period and the as-of date for this amount are not established by published evidence.
A control room for running several projects at once
Our own project-management tool, designed for teams: a user portal with roles, company context, document parsing, and a forge that turns stakeholder interviews into project scopes and structured company knowledge. It is rough, but functional; today, it is what this firm is run with. It is also our test bench: what breaks on us will not break on you.
Running the consultancy itself
Quotes, invoices and follow-up for the consultancy run on a tool we built and use every day. AI holds two precise roles in it. A support assistant helps find or correct an entry. And when a client pays several invoices with a single transfer, often of similar amounts, the reconciliation delegates its reasoning to a model that proposes the best hypothesis; a person confirms before anything is written.
Status: internal tool, in production. The reconciliation model is a mid-range model, chosen because the problem is within its reach and it costs less: putting AI in its right place also applies to model size.
Automated measurement rails
Citation readings with frozen questions and repeated passes, performance captures on the official tools, and the correlation study that validates the instrument against the real application: all of it runs in production, at known dates, and feeds the Barometer and the diagnostics. Publication is signalled to the engines automatically, every change is backed up before it lands, and every reading can be replayed.
Our audience measurement is built in-house, cookie-free, and distinguishes visits coming from AI engines from classic web visits: we measure for ourselves what we sell the measurement of. This count is a floor: some AI-driven visits arrive with no referrer and are classed as “direct”.
What this proves
That the method we sell is the one we practise: start from the business process, keep the human decision, document, stay reversible. These systems have been running for months, in real conditions, with dated publications anyone can verify.
What this does not prove
Only one field system is named, Food Collective’s, and what we publish about it is a testimonial: a result reported by the client, not a measured one. The other field systems remain anonymised. No engagement counts as a measured reference until the result is quantified. The distinction between house system, proof of concept and client system is written on this page, every time.
Three honest questions
Are these client engagements?
Not all of them, and each status is written. The systems of the house are ours. The voice assistant is a working proof of concept at Food Collective, the only field system named here, and what we publish about it is a testimonial, not a measured result. The farm data system is a prototype demonstrated end to end for a client, anonymised. Nothing will count as a measured reference before the result is quantified.
Why show your own systems?
Because an integration provider who runs nothing of their own is selling a plan, not a practice. Ours wears in daily: the guardrails we propose to clients are the ones that already constrain our own publications.
Can I see them up close?
Part of it is public: the dated articles, the Barometer, the correlation study. The rest, portal and pipeline, is shown in a meeting, on the record, in 30 minutes.
Your first business workflow, installed end to end?
The 30-minute call is prepared, on video, with no commitment. Théo Henusse answers you himself.