An AI tool demo always works. That is its defining feature.

The dataset has been picked. The case on show is the favourable case. There is no history to migrate, no volume, no exception, no colleague who filled the “comments” field with three lines of indispensable context. None of that is dishonest: a demo that failed would demonstrate nothing, so it gets built to succeed.

It is a show home. It is real, clean, well lit, and it teaches you nothing about living there. The furniture is undersized so the rooms look bigger, there are no neighbours and no clutter, and nobody is lying to you while showing you round. You simply are not buying the show home.

So the question is not whether the demo lies. It is what the demo is structurally unable to show.

The test you can run during the demo#

There is one, it is simple, and you can run it from your phone while the tool is being presented to you.

Take the request the salesperson just typed. Paste it into ChatGPT, into Claude, into any consumer assistant. If you get 80% of the same result, what you are being offered is not a product, it is an interface.

The industry has a word for this: a wrapper, a thin layer laid over a model’s API. And we should be fair, because the word gets used a great deal to dismiss things without thinking. The charge holds against a product that amounts to a prompt and a nice layout. It is lazy against a company that has genuinely built something. The difference is not in calling a model, everybody calls a model. The difference is in what surrounds it.

The test is therefore not a verdict, it is a question to ask: what do you have that I do not already get from my twenty-euro subscription? The answers that hold up talk about your data, the way you work, what the tool knows about your company. The answer that does not hold up is “a better prompt”.

What the demo cannot show is you#

Back to the show home. What it fails to show is not a hidden defect in the building. It is you inside it.

A demo runs on the vendor’s data. It is clean, consistent, complete, and it was chosen. Yours lives elsewhere: in a management system configured eight years ago, in Excel files that carry authority without anyone ever having decided so, in a half-filled CRM, and in the heads of two people.

Your catalogue, your customer records, your internal procedures, your business rules and your regulatory obligations are not in the model. They never will be: it was trained on the internet, not on your company. A tool with no access to them can answer questions. It cannot solve your problems. That is the whole distance between an improved FAQ and a colleague.

This is not an opinion. In its 2026 digital trends report, Adobe finds that 78% of French companies cite data quality and integration as the main barrier to scaling AI agents. And that only 36% consider their data fit to be used at all.

Above all, remember this, because it is what will happen to you. When the tool disappoints, everyone will conclude the model is not good enough, and someone will suggest waiting for the next version. In the vast majority of cases the model has nothing to do with it. It is an architecture problem, not an intelligence problem.

The gap between what works and what you are sold#

We need to be precise here, because the subject invites caricature and these tools do produce real results.

The same Adobe report notes that 74% of French companies see a genuine gain in the volume and speed of content production thanks to generative AI. That is not nothing, and nobody disputes it.

But move up one level, to the agent that carries out tasks inside your processes, and the figures collapse: only 13% of French brands have deployed agentic AI at scale for customer support.

Globally, PwC’s annual survey of 4,454 chief executives across 95 countries gives a result that is hard to wave away: 56% report neither a revenue increase nor a cost reduction attributable to AI after twelve months. Only twelve per cent achieve both. A much-discussed MIT report separately claims that 95% of generative AI pilots produced no measurable effect on the bottom line. That figure comes with caveats, the report being preliminary, not peer reviewed, and its definition of success unusually narrow.

The gap between these two worlds is not a lie. It is a shift in scope. You are shown what works, content generation, and sold what does not yet scale, the agent that works inside your company.

Integration is the entire job#

If you take one thing away: in a corporate AI project, the model is the easy part.

The real work is everything else. Plugging into the existing systems, the ones nobody documented. Formalising business rules that were never written down because everyone knows them. Deciding what the tool may do, what it must propose without executing, and who it asks before acting. Cleaning and structuring the data so it becomes usable. Then measuring, comparing against the situation before, and starting again.

None of this fits into a monthly subscription, because none of it is generic. It is engineering carried out at your premises, on your data, with your constraints.

It is also where the costs the demo does not show are hiding. Beyond the headline price there is training time, usage overages, the disruption to existing workflows during the switch, and the integration debt you take on with every additional tool. The bill can double.

And there is that consequence nobody talks about: 77% of small businesses using AI have no written rule about it. This is not a compliance footnote. It is why, one Tuesday morning, an employee pastes a customer file into an online tool without anyone ever having told them it was a problem.

Three questions before you sign#

I am not telling you to turn down the demo. Watch it, they are often impressive, and the tool behind it is sometimes excellent.

Just ask three questions. What happens when the tool has to work on my data, and who does that work? What do you have that I do not get by pasting my request into a consumer assistant? And how will I know, three months from now, that it worked?

If all three answers are solid, you have found a good supplier. If they stay vague, you have not viewed the property. You have viewed the show home.

Sources#