Standalone AI apps are losing users, and buyers want AI inside the tools they already use

Dan Maccarone argues that the standalone AI app is dying while AI keeps growing, because intelligence is moving into the tools people already use. RevenueCat's analysis of more than 115,000 apps found that AI apps convert free users to paid faster, then lose them roughly a third faster, and only about one in five subscribers is still paying a year in. An a16z study with OpenRouter found that stickiness depends on the underlying model, and fewer than one in ten weekly ChatGPT users ever tries a competing assistant. He cites Character.AI, Jasper and Yupp as examples, and Brian Balfour's term "product-market fit collapse". Buyers want the AI "Invisible to the user, totally visible to the owner", and he says almost nobody builds both. Why it matters: Maccarone says the apps that keep users fit a daily workflow so closely that leaving costs something. Snyk found companies stitching AI into their existing stack, and he cites MCP at roughly 97 million monthly downloads and production use in about 40 percent of technical organizations. Owners still ask for a dashboard, an audit trail, a kill switch, proof of ROI and a guarantee that the model isn't learning from their data. His line is "Invisible is fine. Ungoverned is a non-starter." Kong found that fewer than 1 percent of organizations had AI-specific governance controls, and Verizon's 2026 DBIR puts unapproved AI use at 45 percent of the workforce. He argues that design work moves to system behavior, and the scarce skill is judging what the system should do, when it should speak and what it may touch. Many statistics come through blogs and aggregators, some numbers differ between the body and the source list, and the enterprise quotes come from his own sales calls in a market he sells into. ChatGPT itself is a sticky standalone app, so the evidence shows thin wrappers dying, not every app.

Try this: When you ship an AI feature inside an existing product, design the owner view first: a log of every run, what it touched, a kill switch and an ROI measure.