Five AI projects we won't sell you

Part of our job is selling artificial intelligence projects. So here is a text our sales department would never have written: the list of projects we will say no to.

Let us be precise about what this text is not. A young company claiming to have "turned down dozens of contracts" would be storytelling — and our editorial principle is that everything we publish must be verifiable. So we do better: we publish our refusal grid in advance. It is dated, public, and you can hold us to it. If one day we sell you one of the projects described below, you can send us this page. It is the opposite of a marketing pitch: it is a commitment that will cost us contracts.

Every criterion comes from real experience — our own, on our own products and in the field. And every refusal comes with its alternative, because a no without a path helps nobody.

1. "Management wants AI" — the project without a problem

The brief starts with the solution: "we need AI", "our competitors are doing it", "the board is asking for our AI strategy". What is missing: the problem. No metric that hurts, no process whose current cost is known, no definition of what "success" would mean.

These projects do not fail loudly — it is worse: they succeed technically and change nothing. An impressive demo, a pilot that runs, and six months later nobody uses it, because it solves nothing anyone experienced as a problem.

What we offer instead: start with an inventory of your operations, not with the technology. Where do the hours go? Which tasks have volume, repetition, a rule? Which metric, if it moved by 20%, would change your quarter? If there is no answer, there is no project — and you have just saved six months.

2. Calculation disguised as conversation

"The chatbot will also be able to calculate the quote, right?" No. Not with us.

A language model is probabilistic by design: same input, possibly different outputs. That is a strength for writing; it is disqualifying for calculating a price, a total, a margin, the application of a rate card. We built an entire platform around this separation — the model interprets the request and writes the document, a deterministic, auditable engine does all the arithmetic. Same input, same result, every amount explainable line by line.

What we offer instead: the right division of roles. The model at the edges of the system — understanding a free-text request, writing, summarising. Classic software at the core — rules, calculations, data. If your need is 90% calculation, the honest answer is sometimes that you need good business software, not an AI project.

3. RAG on a documentary swamp

"Ask your documents questions" is the most-sold promise of the moment, and it works — on one condition the promise leaves out: that the documents deserve answers. Three contradictory versions of the same procedure, binders with no owner, archives nobody has sorted in ten years: an AI-augmented search system (RAG, retrieval-augmented generation) laid on top will produce fluent, sourced answers — and wrong ones one time in three, with nobody knowing which.

There is a second condition, which we detailed in our nFADP playbook: the access rights of the source system must survive in the index. An assistant that answers everyone with everyone's documents is not a product, it is an incident waiting to happen.

What we offer instead: a narrow, governed scope — a single corpus, one owner, one authoritative version, clear access rights — and a measurable pilot on it. Document governance first; it will remain an asset even if the AI project stops.

4. Nobody has hours for ground truth

This is the criterion the field taught us the hardest, and we told the story in detail: AI compresses construction, it does not compress validation. A business system has to be anchored in the company's reality — its prices, its rules, its special cases — and that truth is written down nowhere. It lives in your experts' heads.

If, at signing, the company cannot name who will validate and how many hours they will spend on it, the project will not die — it is more insidious than that: it will stay a pilot forever. Technically successful, operationally orphaned.

What we offer instead: expert hours in the contract, scheduled validation sessions, and a product designed to make validation as cheap as possible. If those hours really do not exist, the project's time has not come — and saying so before signing is a service, not a withdrawal.

5. The silent replacement

Sometimes the real specification is unwritten: automate a high-stakes decision — screening job applications, granting or refusing something to a customer, assessing employees — without human review, because human review "costs". Or replace a team without saying so, and call it "augmentation".

We turn these projects down for an ethical reason that does not need to be disguised as a technical one. But the technical reason exists too: Swiss law requires informing people about automated individual decisions and allowing a human review, and a system with no human checkpoint is fragile exactly where being wrong costs the most. Our architectural position does not change, from a single quote to an entire catalogue: AI prepares, the professional signs.

What we offer instead: openly automating what can be automated — the volume, the preparation, the draft — with human sign-off points at the decisions that commit. And if the goal is a headcount reduction, let it be decided and announced as such, by management, not delegated in silence to a model.

What this page costs us

Let us reread: we have just described five families of contracts we forbid ourselves. It is a real cost, accepted for an economic reason as much as a moral one: French-speaking Switzerland is a small market. A delivered project that serves no purpose gets talked about faster than a successful one — and it gets talked about with our name on it.

So bring us your project, including if it looks like one of the five. The diagnosis is the start of our work, and it is not complacent: if the answer is no, you will know why, and you will leave with the path to what needs to be built first. A reasoned no is worth more than a pilot that will never finish — it is, literally, our first sales page.