Claude Dispatch and the Power of Interfaces

  • Author/Source: Ethan Mollick (Wharton School, University of Pennsylvania), via One Useful Thing Substack (2026-03-31)

  • Original: https://www.oneusefulthing.org/p/claude-dispatch-and-the-power-of

  • Key Ideas

  • The chatbot interface itself creates cognitive costs that offset AI productivity gains --- research shows financial professionals were overwhelmed by walls of text, sprawling discussions, and AI mirroring disorganized user structure.
  • The "capability overhang" in AI comes largely from bad interfaces, not model limitations. Most people access AI through free chatbot tiers with less capable models.
  • Specialized interfaces for coding (Claude Code, Codex, Antigravity) are mature; Google is experimenting with specialized interfaces for other domains (Stitch for design, Pomelli for marketing, NotebookLM for research).
  • OpenClaw (open-source personal agent) became the fastest-growing open source project by solving the interface problem: let users talk to AI through familiar messaging apps (WhatsApp, Telegram, Slack) instead of chatbots or command lines.
  • Claude Cowork with Dispatch: Anthropic's answer gives Claude access to local files and applications via a desktop workspace; Dispatch adds phone-based remote control (scan a QR code, message Claude from your phone while it works on your desktop).
  • AI can generate interfaces on demand --- Claude now creates interactive visualizations directly in conversation, building the right interface for each moment rather than forcing a fixed format.

  • Much "AI disappointment" likely comes from interfaces being wrong, not AI being bad.

  • Summary

Mollick argues that the biggest barrier to AI productivity is not model capability but the interface through which people access it. He cites research showing that chatbot interactions create measurable cognitive overload --- especially for less experienced workers who could benefit most from AI. The chatbot format (walls of text, topic sprawl, mirrored disorganization) actively undermines the AI's intelligence.

The article surveys three approaches to solving this problem. First, specialized interfaces built for specific professions: coding agents are the only fully realized example, while Google's experimental tools (Stitch, Pomelli, NotebookLM) hint at what domain-specific AI interfaces could look like. Second, personal agents that use existing communication interfaces: OpenClaw's explosive growth proved that people want to message their AI the way they message people, not through a chatbot window. Anthropic's Claude Cowork with Dispatch follows this insight, letting users control an AI agent from their phone while it works on their desktop --- Mollick demonstrates it updating a PowerPoint presentation by finding newer data, downloading a PDF, extracting a graph, and replacing the old one.

Third, and most forward-looking, AI can generate its own interface on demand. Claude's new in-conversation interactive visualizations represent a shift from adapting to the AI's interface to the AI adapting its interface to you. Mollick predicts the future is not one interface but AI that generates the right interface for each task --- and that every improvement in interface will feel like a leap in AI capability, even when models haven't changed.

  • Relevance to Economics Research

The interface problem is acutely relevant to economists. Most researchers access AI through chatbots, which this article argues systematically undermine productive use. The Cowork/Dispatch model --- where an AI agent works on your local files (datasets, Stata do-files, LaTeX documents, PowerPoint presentations) while you direct it from your phone --- closely maps to how economists might want to interact with AI during a workday. The cognitive load research cited here also has implications for how we think about AI adoption barriers in the profession: the issue may be less about AI quality and more about the tools we use to access it.