Build1 publisher3 min readPublished
A bigger platform shipped the feature in three of five AI closures with a stated cause
Relay stopped operating on Monday after OpenAI and Google added the same workflow-agent capability to their own products. Two of the other four closures in the same account went the same way, and the 42% failure figure quoted alongside them counts something else.
The Engineer · Build desk

What happened
- Relay, a workflow automation platform pitched as a Zapier alternative that ran AI agents over email and everyday tasks, stopped operating entirely on Monday.
- mezha.net, citing TechCrunch, said Relay could not compete once OpenAI, Google and other technology giants added analogous capabilities to their own products.
- Notion Mail, launched in April 2025, shuts on 22 September 2026 after Notion concluded that users increasingly handle inbox work with separate AI agents.
- Huxe, built by former Google NotebookLM developers to turn text into podcast-style audio, closed in May 2026 after Spotify and other large platforms offered similar tools to far wider audiences.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision Continuity for an agent-tooling purchase turns on a model vendor's roadmap, so the exit path has to be priced at signing.
- contradiction Using the 42% to explain Relay mixes internal enterprise programmes with a vendor's commercial failure, and the two populations stop for different reasons.
- constraint With no share per cause, the figure cannot rank competition against scaling in a portfolio review; it supports only the statement that a large minority stop.
Count the closures in the account and the pattern is narrow. Three of the five products described as stopping with a reason attached (Relay, Huxe and Notion Mail) lost the work to a larger platform or to general-purpose agents [14]. The other two failed on their own economics. Sora stopped in March 2026, with high operating costs and difficulty holding an audience among the reasons cited [8]. Humane ended the AI Pin in February 2025 after serious performance problems and a warning about fire risk from its charging case. HP bought most of the company's assets for $116 million, roughly half the $230 million Humane had raised [10][15].
The 42% belongs to a different population. S&P Global Market Intelligence puts the share of corporate AI initiatives that eventually lose funding or stop at roughly 42% [3], and the causes given are technical limits, shortage of funds, scaling difficulty, hard competition and weak demand [4]. Those are programmes run inside companies. Relay was a product sold to them [17]. The five causes are listed without a share for each [18], so the figure cannot tell a buyer whether competition or scaling is the likelier way a given tool dies. I have sat in reviews where a number of that shape decided a budget line, and nobody asked what population it counted.
What a buyer can act on is narrower. If the tool's core function is a plausible feature of the assistant the company already pays for, the vendor's roadmap is the continuity risk. mezha.net, citing TechCrunch, gives exactly that as the reason Relay could not compete [2].
Standardising on the incumbent leaves that gap open. OpenAI wound down separate products and moved their functions into ChatGPT, including the Atlas browser, the Operator agent and part of DALL-E's image generation [7]. In July it merged the Chat, Codex and Work modes into a single app and renamed the standard version ChatGPT Classic. Users criticised the structure as confusing and overloaded, and the company restored the previous format [9].
Two other cases in the account failed on review. Microsoft delayed Recall by nearly a year and hardened its protections after security specialists warned that the screenshot database could contain passwords, private correspondence and financial information, and the concerns did not go away [11]. Apple's reworked Siri, shown with Apple Intelligence in 2024, reached the iOS 27 beta in July and began rolling out to English-speaking users in September 2026. The delays were one factor in a dispute settled for $250 million over the marketing of iPhone 16 capabilities [12].
The Rabbit R1 is the one device here that kept shipping through bad reviews. It was shown at CES in January 2024 and sold 100,000 units. Reviewers called it unfinished, with unstable operation and few useful integrations, and the company continued updating it as a controller for agentic tasks while announcing a portable Cyberdeck project for programming in natural language [13].
What to watch
- Whether S&P Global Market Intelligence publishes a per-cause breakdown of the 42%. That split is what would make the figure usable in vendor selection.
- Whether Notion ships an export or migration path for Notion Mail customers before the 22 September 2026 shutdown.
- Whether OpenAI attempts the combined Chat, Codex and Work app again after reverting to the previous format.