A robot folds a garment, an artificial intelligence system books a trip, a sensor promises to prevent a breakdown before it happens. In the aisles of CES, VivaTech or the RAISE Summit, everything seems to work. But what remains when the discreet operator, dedicated network and rehearsed scenario disappear? As the industry returns from the summer break in September 2026, here is a method for assessing announcements without confusing spectacle with industrial maturity. It draws on previously documented cases; any potential developments discussed are forward-looking analysis, not a verified assessment of the 2026 events.
Three events, three attention-generating machines
CES, in Las Vegas, puts the emphasis on tangible products: screens, cars, connected devices and household robots. It is easy to judge a design or an interface there, but much harder to assess maintenance. VivaTech, in Paris, brings together startups, large corporations and public policymakers. The risk lies in mistaking a trial with a prestigious partner for a contract implemented at scale. Focused on artificial intelligence, the RAISE Summit places greater emphasis on models, infrastructure and business applications: an announced partnership is no more proof of recurring use there than anywhere else.
These differences change the questions to ask, not the standard of scrutiny. A demonstration establishes that something can work under certain conditions. A deployment shows that it works often enough, at an acceptable cost, for someone who bears the consequences. Between the two lie the least photogenic stages: integration, security, training, support and contract renewal.
First filter: establish what actually exists
Before assessing an innovation, its status must be established. Is it a concept, a working prototype, a product available for preorder, a delivered product or a service used every day? Press-release language often cultivates ambiguity: “available” may mean accessible to a handful of partners; “deployed” may mean installed at a single site. Asking for a delivery date, documentation and commercial terms already clears up much of the confusion.
AI-powered hardware assistants offer an instructive precedent. The rabbit r1, unveiled at CES 2024, did indeed go on sale. That was not enough to validate its promise of acting as a universal intermediary between users and their digital services: early independent reviews found limited functionality and inconsistent reliability. The product existed; its value proposition compared with a smartphone remained unproven.
Another warning, unrelated to the trade-show calendar, is the Humane AI Pin. After it went on sale in 2024, its consumer services were shut down in February 2025 as HP acquired some of the company’s assets. The lesson extends beyond this particular device: a product that has been delivered may depend on a remote service whose disappearance disables its essential functions. The supplier’s long-term viability is therefore part of technical maturity.
Second filter: look for the customer beyond the pilot
The best counterpoint to a trade-show booth is often an operations manager. How many people actually use the tool? Since when? For what proportion of cases? Which process has been eliminated, rather than merely duplicated? A named customer is an initial piece of evidence, not a conclusion. They may receive exceptional support or pricing that bears no relation to future commercial terms.
In business AI, coding assistants provide a more substantial example than many general-purpose agents. GitHub Copilot is a commercial service integrated into development tools, with documented use cases. But the findings require careful interpretation: a controlled experiment published in 2023 showed faster completion of a narrowly defined task. It does not prove that all software projects automatically become faster, cheaper or more secure.
By contrast, Klarna heavily promoted its customer service assistant in 2024, claiming significant gains. In 2025, its chief executive publicly stressed the need to preserve quality and access to human agents. That does not negate the value of automation. It is a reminder that a metric tracking the volume of cases handled measures neither lasting satisfaction nor the cost of poorly resolved cases.
Third filter: recalculate the economics
On stage, innovation produces a result. In a business, it also produces a bill. For AI, the cost goes beyond the subscription: data preparation, software integration, computing, monitoring and human correction all add up. For a robot, installation, premises modifications, consumables, downtime and technical servicing must be factored in. The right comparison is the cost of a task completed correctly, not the cost of an attempt.
The questions that strip away the stage set
- What is the failure rate under normal conditions? Require a definition of failure and a clear account of the test’s scope.
- Who steps in when things get stuck? A remote operator may be necessary, but their work must be included in the calculation.
- What result is achieved without the solution? Compare it with an existing process, not an artificially degraded situation.
- Does the customer renew? A paid expansion reveals more than a subsidized pilot.
These questions sometimes favor less spectacular technologies. Machine vision on a clearly defined production line, energy optimization in a building or assistance with processing standardized documents may offer a clearer commercial path than a supposedly multipurpose robot. A stable environment reduces exceptions; it also makes the benefits easier to measure.
Fourth filter: test the constraints, not just the features
A commercially viable innovation must withstand more than an unexpected question. Can it work with confidential data? Does it respect access permissions? Does it produce usable logs? Can suppliers be changed without rebuilding the entire system? In sensitive sectors, traceability and accountability sometimes matter more than a few extra points on a benchmark.
The European Union’s AI Act, adopted in 2024, phases in its obligations over time. For a buyer, the right question is not simply “Are you compliant?” but “Which obligations apply to your use case, on what timetable, and with what supporting evidence?” A broad claim at a trade-show booth is no substitute for either use-case analysis or contractual documentation.
Cross-check the evidence rather than pile up announcements
The editorial method boils down to three columns: the supplier’s promise, independent observation and evidence from actual operations. A video sheds light on the scenario; an external review, on the limitations; detailed customer feedback, on performance over time. When these sources diverge, the gap must be explained. And when they are missing, the project should explicitly be described as experimental, without dismissing it: some innovations legitimately need several years.
What next? The most credible outlook for upcoming events is a tougher distinction between demonstrations of autonomy and tools that can genuinely be supervised. The winners may be less those promising to do everything than those documenting their limitations, controlling their costs and finding customers willing to renew. To make sense of VivaTech, CES or the RAISE Summit, it is therefore better to follow an innovation through to its second contract than to the final round of applause.
