Roughly 42 percent of corporate AI initiatives never make it past the pilot stage, according to research from S&P Global Market Intelligence cited in a running tracker TechCrunch maintains of shuttered AI products. The causes vary by company, but the tracker’s newest entries point to a single mechanism showing up again and again: a feature that once justified a standalone product gets absorbed into a bigger platform, and the standalone product loses its reason to exist.
Relay, a five-year-old automation startup pitched as a Zapier rival, closed this week after OpenAI, Google, and other large platforms wired similar automation directly into tools their users already had open. Relay’s collapse is not a story about bad execution. It is a story about a category getting absorbed by the infrastructure layer above it, the same fate that has now caught up with several products built on top of frontier labs’ own platforms.
OpenAI supplies the clearest pattern inside a single company. In July, it tried to fold Chat, Codex, and Work modes into one “super app” interface and renamed the familiar version ChatGPT Classic; user backlash over a cluttered redesign forced a rollback within weeks. Separately, OpenAI discontinued its standalone Atlas browser in August after less than a year, folding the useful parts back into ChatGPT, and did the same earlier with Operator, its web-browsing agent. DALL-E’s role has shrunk the same way as image generation moved natively into the chat product. Sora, OpenAI’s video-sharing app, shut down in March under the weight of operating costs and weak retention, a different failure mode: too expensive to run, not too easy to absorb.
Notion Mail is the newest case of the absorption pattern. Launched in April 2025 to automate inbox triage with AI, it is scheduled to shut down September 22 after Notion found its own users routing email through separate AI agents instead of the built-in one. The lesson mirrors Relay’s: when the orchestration layer (the agent picking which tool to use) sits outside your product, your product becomes a commodity the orchestrator can route around.
Two failures broke for reasons that have nothing to do with platform absorption. The Humane AI Pin, a $230 million wearable bet, struggled with basic performance and a battery fire warning before shutting down its Pin business in February 2025, with assets sold to HP for $116 million. Yupp, a side-by-side AI model comparison playground that let users test more than 800 models, shut down in March 2026 after its own founders said it never found product-market fit, regardless of platform competition.
The pattern matters for anyone building a point solution on top of a frontier lab’s API right now. If the feature you ship is one a foundation model provider could plausibly bundle into its own chat interface within twelve months, the addressable window is not a roadmap decision, it is a countdown. Startups building agent orchestration, memory layers, or workflow glue should treat “does this get absorbed by GPT, Gemini, or Claude’s native app” as a pre-seed diligence question, not a Series B one.
Reporting per TechCrunch’s Lauren Forristal, in a running list published September 15, 2026.
