AI and data

We use it every day. That is why we know how to make it answer correctly.

An assistant plugged into your data always answers. The question is not whether it will answer, but whether the answer will be right, and what, underneath, decides that. We have the answer, and it is not in the language model.

Internal portals where we use it ourselves: 4
Condition that decides correctness: 1
This page is kept current: 2026
Expert consultants: 11

1
Condition decides correctness: the semantic model: 4
Internal portals where we use it ourselves: 11
Expert consultants: 2026
This assessment is dated and kept current

The mechanism

What makes an assistant answer correctly

Ask “what was last year’s revenue?” of an assistant plugged straight into your ERP tables. It has to choose: a table of invoice lines or a table of accounting entries, an invoice date or a posting date, with or without credit notes, with or without the divested subsidiaries.

It will make a choice, and that choice will be plausible. But you are the one who knows which is right, and that knowledge has a place to live.

The same assistant, plugged into a semantic model where the revenue measure is defined once and for all, no longer has a choice to make. It reads the definition, applies it, and returns the same figure as the official report, because it is literally the same calculation.

AI does not replace data modelling. It makes it the only thing that matters.
Our position, in one line

The deciding layer

An assistant is only as good as the layer beneath it

The architecture we have deployed for years (raw, reconciled, modelled) was not designed for AI. It happens to be exactly what AI needs. An assistant plugged into the gold layer inherits every decision already taken and validated; plugged into the bronze layer, it takes them all again from scratch, alone, and without telling you.

BronzeIngestion: Raw data, exactly as the systems produce it. ERP, CRM, Dataverse, SQL databases, files.
SilverCleaning: Cleaning rules and business alignment. This is where the charts of accounts of six subsidiaries are reconciled.
GoldSemantic model: Aggregated and optimised data. One source of truth, shared by every report and every department.

Current assessment

What we recommend, what we watch, what we set aside

This assessment is dated: it holds for 2026 and it will move. We keep it current rather than let a page quietly age.

Deploy now

Copilot for producing code

Writing a DAX measure, a Power Query transformation, a SQL statement: the assistant proposes, the consultant checks, and the gain is real because a mistake is immediately visible. A wrong DAX measure does not compile, or returns a figure that can be reconciled in ten seconds. It is the most mature use, and we rely on it daily.

Conditionally

Asking questions of the model

Querying a semantic model in plain language works, provided the model is clean, the measures are named the way the business names them, and the scope is narrow. We have not yet put it into production for a client, and we will say so the day we do rather than before.

We prepare the ground first

Pointing an assistant at raw data

“No need for a model, the AI will figure out the database” is the most-sold and most expensive promise. It produces plausible, wrong answers in a setting where nobody checks any more. It is a project we decline, and we explain why rather than quote it.

On ourselves first

We use it on our own tools before yours

We built our own internal portals (project management, HR, invoicing, quality) on the same application framework we offer clients. They serve as our proving ground: what does not hold up here does not go out to you.

A proving ground that costs us

When an assistant gets our own invoicing wrong, we pay for the correction. That concentrates attention on what actually works.

Public, free applications

We also publish applications open to everyone, with nothing asked in return. They are honest test beds: nobody tolerates an approximate answer just because it is free.

A team that trains continuously

The team holds Microsoft Fabric certifications and keeps training, a habit from our years as independents, when learning was how we kept working.

A dated opinion, not a stance

We publish what we think of each new release, with its date, and we revisit it when we were wrong. An opinion that never changes is a posture, not expertise.

Going further

What everything above rests on

Frequently asked

What people ask us about AI

Can you really ask questions of your data in plain language?

Yes, on one condition that changes everything: that a semantic model underneath defines revenue once and for all. An assistant plugged into a clean model answers correctly. An assistant plugged into raw tables invents a plausible join and returns a wrong figure with the same confidence. The difference does not come from the AI, it comes from what sits underneath.

Does AI replace the modelling work?

It makes it more decisive, not less. As long as reports were read by humans, an ambiguous definition got caught in a meeting. The moment a machine answers instead of a human, the ambiguity becomes a wrong answer served without reservation. The semantic model stops being good practice and becomes the condition for it to work at all.

Will our data be used to train a model?

No, for the Microsoft services we implement: Copilot in Power BI and in Fabric processes your data without using it to train the foundation models, and processing stays within your tenant's scope depending on the regions enabled. It is a point we verify on each project rather than assume, because the terms evolve and differ between services.

Do you use AI on your own projects?

Every day, and on our own tools before our clients': our internal portals for project management, HR, invoicing and quality all use it. That is what lets us speak from experience rather than from documentation, and spares you the trials we have already run.

Where should we start?

On a scope where a mistake is visible and harmless. An assistant answering on one branch's revenue, where three people know the figure by heart, corrects itself within a week. The same assistant let loose on consolidated financial steering produces errors nobody catches before the board meeting.

The mechanism

What a shared layer changes

With no platform, every report pulls its own extract and recodes its rules, hence three figures for the same question. With a shared layer, the rule is written once and every report agrees.

The same report, two architectures underneath

On the left, every report redoes the work. On the right, it is done once and for all.

The move changes nothing of what the user sees: it changes what sits behind, and therefore whether two departments get the same figure.

Before the assistant, let us look at what would sit underneath.

Eleven expert consultantsSaint-Priest, near Lyon, France

Two hours with a consultant, free of charge. We look at your data model as it stands today and tell you plainly whether it would support an assistant, and what it would take. Often the answer is less work than expected.

Request a scoping workshop

Scoping workshop · 2 hours · free · near Lyon or remote