A synthetic voice answers without hesitation, a robot picks up the right object, a dashboard promises to save hours of work. In the aisles of VivaTech, everything seems to work. That is precisely the point of a demonstration: to show the best-case scenario. For buyers, journalists and executives, the work begins after the applause. As business resumes this September 2026, here is a framework for assessing trade show innovations, based on developments already documented, without making assumptions about the performance of recently announced products.
A pitch sells a possibility; a purchase commits an organisation
VivaTech brings together everything that makes major technology trade shows useful: quick meetings, accessible demonstrations and the opportunity to compare several approaches in a single day. But this format also encourages shortcuts. A few minutes are enough to understand a promise; they rarely allow time to examine its technical, contractual and human requirements.
The rise of generative AI since ChatGPT’s public launch in late 2022 has widened this gap. Building a convincing interface around an existing model has become easier. Turning that interface into a reliable service remains another matter. Connections to internal data, error handling, access permissions and human oversight rarely make for the best video clips. Yet they determine everyday usefulness.
The right question, then, is not simply “Is it new?” but “Under what conditions does it work?” Four filters can help turn a sales narrative into an informed decision.
1. Availability: can you actually use it?
A prototype, a beta version and a commercially available product can look similar on a stand. They do not entail the same commitments from the supplier. The first check is to ask what is accessible today, in which countries, to which users and with what limitations. A waiting list is not availability; a partnership announcement is not an operational integration.
The simplest test is to request a trial using a case the exhibitor has not prepared. For a document assistant, you could suggest a public document containing a contradiction or missing information. For an industrial tool, ask about operating conditions: lighting, temperature, network quality and variability in the objects handled. The aim is not to catch anyone out, but to uncover the limitations.
You also need to look behind the screen. Is an operator quietly correcting responses? Does processing take place locally or through a remote service? Do certain features depend on another supplier? These choices do not disqualify the product. They affect its autonomy, latency and ability to remain available when an outage occurs.
2. Customer references: a logo is no substitute for a track record
Walls of logos quickly inspire confidence. Yet they can represent very different situations: a paying customer, a free trial, a technical partner or simply participation in a business support programme. Asking about the exact nature of the relationship is a basic test. A reputable supplier should be able to distinguish a proof of concept from a deployment used in production.
The most relevant reference is not necessarily the most famous brand. It is an organisation that resembles yours in its constraints, volumes and resources. A tool adopted by a specialist team in a large corporation does not prove that a small or medium-sized business can maintain it without assistance. Conversely, success on a modest scale can provide excellent evidence if the results are observable.
With the supplier’s agreement, requesting a conversation with an actual user often offers more than another presentation. Three topics are critical: the time taken to achieve the first useful application, the difficulties encountered and the tasks that remain manual. Claimed gains should specify the baseline, the measurement period and any additional verification work. Time saved on drafting can be lost again on corrections.
3. Security: follow the data, not the slogans
“Secure”, “sovereign”, “compliant”: these words alone do not describe an architecture. You need to trace the data’s journey. Where is it stored? Who can access it? Which subcontractors are involved? How long is it retained? Is it used to improve a model? The answers must appear in technical and contractual documents, not just in conversation.
The GDPR has long provided a framework for personal data. The EU AI Act, which entered into force in August 2024 with phased implementation, adds requirements depending on uses and responsibilities. It is not a universal quality seal. A marketing claim of compliance therefore does not remove the need to classify the use case or check which obligations actually apply.
For an AI solution connected to internal tools, one risk deserves particular attention: excessively broad access to information or actions. An assistant that summarises a case file does not necessarily need permission to modify every file. Stronger authentication, restricted permissions, logging and human approval of sensitive operations are more concrete safeguards than a general promise.
Certifications and audits can provide useful evidence, provided their scope and date are examined. They do not guarantee the absence of vulnerabilities. You should also ask how the supplier reports incidents, restores service and handles data deletion.
4. Cost: account for deployment, then for exit
The advertised price rarely tells the whole story. Connectors, data preparation, integration, training and support may come on top of the licence fee. With usage-based AI services, spending also varies according to volumes, document length or the number of calls needed to complete a task.
To compare two offers, it is better to define a meaningful unit: a case processed and verified, a request resolved, a piece of equipment monitored. The cost of a single query tells you little if several attempts and human intervention are needed. A limited pilot should measure quality, the time actually saved and the workload transferred to IT or business teams.
Leaving matters as much as getting started. Can you export the data in a usable format? Retrieve the configurations? Switch providers without rebuilding the entire process? A solution that is cheap to test can become expensive to leave. These questions must come before a wider rollout.
A simple method: demand evidence proportionate to the risk
The aim is not to subject every start-up to a banking-level audit. The level of evidence should match the risk: a creative tool that handles no sensitive data should not be assessed in the same way as software involved in a medical decision. After the trade show, a short checklist is enough to get started:
- Available: a testable version, known limitations and an identified source of support.
- Proven: a comparable reference and results placed in context.
- Controllable: documented data handling, access and responsibilities.
- Sustainable: an estimated total cost and a planned exit.
What now? As demonstrations become more spectacular, their ability to set products apart may diminish. Value will probably shift towards reliability, integration and accountability. For VivaTech visitors and buyers alike, the best approach remains simple: leave with fewer promises, but more verifiable evidence.
