Should AI features integrate into existing workflows or stand alone?

Friedman warns that "many products are obsessed with being AI-first" and argues teams "might be way better off by being AI-second instead." By his definition, AI-second means the team "focus[es] on user needs and sprinkle[s] a bit of AI across customer journeys where it actually adds value," rather than building a feature whose primary interface is the AI itself.

His example is interface-level: good AI features map "into the mental models that people have adopted over the years" — he names browser autofill as the template — "rather than leaving users in front of a frightening and omnipresent text box." The same logic appears in his closing argument that "many AI interactions are locked within a specific product, but good AI experiences happen where the actual work happens" — citing Slack, Teams, Jira, and GitHub integrations (Dia Browser, Dovetail) as the AI-second pattern in practice, against a dedicated AI tab or panel as the AI-first pattern.

This cuts against treating a single general-purpose chat assistant as the default AI product shape, in the same direction as Does the personal assistant model actually serve most users?, which argues the assistant for everyone framing over-generalizes from a narrow segment; Friedman's complaint is about interface placement rather than use-case fit, but both converge on not centering the product around the AI itself. It also extends Is the AI capability gap really an interface problem? by naming a specific failure mode — AI-first framing, the "omnipresent text box" — as a cause of the interface problem Mollick attributes the capability-accessibility gap to. The companion note on this source, Can task-oriented UI patterns reduce the friction of AI chat interfaces?, gives the five concrete design levers; this note states the higher-level placement principle those levers serve.

Friedman offers no data on adoption, retention, or user preference comparing AI-first and AI-second products — the claim rests on his own design judgment and named examples (Dia Browser, Dovetail, Adobe Firefly, Flora AI), not a measured outcome. The implication, held to that evidence, is a design heuristic for where to place an AI feature in a product, not a demonstrated effect on usage or abandonment, even though the essay opens by asking how to stop a new AI feature from being "abandoned by users after a few runs."