Startups IA : trois écueils à éviter pour bâtir une entreprise durable

Since the end of 2022 and the emergence of ChatGPT, the technology ecosystem has witnessed the creation of numerous startups focusing on generative artificial intelligence. Despite this growth, many encounter similar structural vulnerabilities. Insights from the AI Ascent event, shared by Sequoia Capital partners, highlight common pitfalls observed within their investment portfolio. Key concerns include misleading revenue figures, deceptive profit margins, and ineffective data flywheels.

Revenue Figures: Misleading Indicators of Traction

A revenue line alone does not necessarily indicate market adoption. Pat Grady, a partner at Sequoia, warns of what he terms "vibe revenue" , which relies on trends or technological curiosity without sustainable usage. Founders often struggle to differentiate between temporary trials and genuine adoption. Analyzing engagement, recurrence, and product usage evolution is crucial for accurate assessment, preventing overestimation of traction that may lead to misjudged team size or market positioning.

Current Gross Margins: Not the Ultimate Goal

AI's current paradox lies in its cost structure. While inference costs decrease as models optimize, with a 99% drop in cost per token over 18 months, margins remain unstable. This is due to reliance on proprietary APIs, real-time usage costs, and challenges in enhancing value propositions to deliver high perceived value outcomes. The focus should be on achieving credible "pricing power" rather than judging current margins. Transitioning from a tool to a solution, and then to a result, allows some players to move beyond feature-centric logic to capture more strategic budgets, often in operational rather than IT cost lines.

The argument often presented is that increased user interaction enhances the product. This positive feedback loop, or data flywheel , is frequently cited as a differentiation lever. However, the collected data must be effectively leveraged to impact a business metric. In many instances, the loop remains theoretical: data is stored and analyzed but does not alter model performance or user experience. If this loop fails to modify strategic data (conversion rate, acquisition cost, execution time), it holds no economic value.

D’après FrenchWeb.