On screen, the agent understands a request, consults documents and prepares a response in seconds. Impressive? Certainly. Distinctive? Much less so when neighboring stands promise the same thing. For a startup preparing for RAISE Summit, the real challenge is no longer to show that AI works. It is to convince a buyer that it deserves a budget, a place in their IT systems and a measure of their trust.
This feature looks ahead to September 2026: it draws on developments already observed in generative AI and offers a preparation framework, without claiming to report on announcements or results from a 2026 edition. RAISE Summit, the Paris gathering dedicated to artificial intelligence, provides a revealing setting: technical experts, investors and businesses cross paths there, but they are not looking for the same evidence.
The promise of automation has become a starting point
Since ChatGPT reached the general public in late 2022, conversational interfaces, augmented document search and coding assistants have become widespread. Major software providers have integrated these features into their offerings. The commercial consequence: “we use a large language model” now describes a technical choice, rarely a reason to buy.
The rise of open models and services accessible through APIs has also lowered some barriers to prototyping. A small team can quickly build a convincing demonstration. But this ease of development makes similarities more apparent. The competitor is not just the startup at the next stand: it is also the feature the customer already has in their software subscription.
To escape this unfavorable comparison, the conversation needs to shift. Rather than selling “an agent that automates operations,” it is better to explain which operation, for which business function, under what constraints and with what verifiable outcome. Precision does not necessarily limit ambition; it makes the first purchase possible.
Start with the problem that has a budget
Before scheduling meetings, the team should be able to answer four questions: who suffers from the problem, who pays to solve it, who authorizes deployment and who can block it? Within a company, these roles are often split between a business manager, finance, IT and security. An enthusiastic user is therefore not yet a customer.
Take a startup specializing in processing supplier files. “Automating administrative tasks” remains abstract. “Helping procurement teams identify missing documents before approval” describes a clearly defined intervention. The proposition becomes stronger when it specifies the existing tool it connects to, the exceptions it passes to a human and the method used to evaluate the outcome.
A commercial proposition in four lines
- The target customer: a specific business function and environment, rather than “all companies.”
- The problem: a delay, error or cost observable in day-to-day work.
- The evidence: a comparative test, a documented pilot or customer feedback cleared for use.
- The commitment: a scope, a price, responsibilities and success criteria.
Without a customer reference, there is no point dressing up a prototype as a proven product. User interviews, a testing protocol and assumptions still to be validated can all be presented. This transparency makes it possible to discuss a pilot without passing off laboratory results as performance already achieved in a customer’s environment.
Prepare a demonstration that can withstand difficult questions
A trade-show demonstration should be short, but not misleading by omission. A perfectly scripted walkthrough shows the intended experience; it proves neither reliability nor robustness. Ideally, a typical case should be presented, followed by an ambiguous one: an incomplete document, a contradictory request or information that cannot be found. What does the system do then?
The right answer is not always an automatic response. A credible product can ask for clarification, cite its source, refuse an action or request approval. For an agent capable of modifying data, permissions, traceability and the ability to undo an operation matter as much as the smoothness of the interface.
The technical conversation behind the spectacle also needs preparation: the origin of test data, measurement methods, latency, dependence on a model provider and the handling of sensitive information. European regulatory developments, notably the AI Act adopted in 2024 and its phased implementation, make these questions even more relevant. Their implications nevertheless depend on the use case and the company’s exact role: blanket compliance cannot simply be declared on a slide.
Sell an outcome without concealing its real cost
Time saved is a useful argument, but an incomplete one. A task completed faster may shift work toward human verification. An inexpensive response can become costly if it triggers corrections. To establish value, it is better to compare an entire process before and after, including checks, errors and rework.
The economic calculation must also include less visible expenses: integration, data preparation, oversight, model usage and support. For investors, these items shed light on future margins. For buyers, they determine the total cost. A strong commercial promise explains where value is created and who bears the costs when a case falls outside the expected scenario.
A pilot should therefore have a defined duration, an owner on the customer side and a decision scheduled at its conclusion. Its price and objectives may vary, but the principle remains the same: test a purchasing hypothesis, rather than keep a free experiment running indefinitely.
Two audiences, two conversations, one reality
Investors look for an accessible market, repeatable growth and a defensible advantage. Customers want to know whether the product integrates with their systems, whether it delivers on its promises and who is accountable if something goes wrong. Presenting the same pitch to both audiences risks failing twice: too operational for one, too speculative for the other.
Preparing for the summit therefore begins with a shortlist of priority meetings. For each contact: a presumed problem, relevant evidence and a realistic next step. An investor discussion may lead to a review of revenue and costs. A meeting with a business team may lead to a workshop bringing together the user, IT and the budget decision-maker.
After the event, follow-up should revisit the stated need, the concerns raised and the agreed action. The number of badges scanned measures stand activity, not commercial traction. Fewer conversations, if properly qualified and followed up, can produce more than a collection of contacts with no intention of buying.
What next? Looking ahead to September 2026, the likely commoditization of technical capabilities could make domain expertise, integration quality and trust even more valuable. This is not a prediction of who will win the market, but a sensible direction for preparation. At RAISE Summit, a startup would therefore do well to arrive with fewer superlatives and more evidence: a specific problem, an honest demonstration and a proposal someone can actually sign.
