This article summarises analysis from One Useful Thing by Ethan Mollick. The assessments described are Mollick’s own.
Ethan Mollick, writing in One Useful Thing, argues that the chatbot interface — the default way most people access AI — imposes cognitive costs that partially offset the productivity gains from AI capability, and surveys alternatives he sees emerging across the field.
The chatbot interface as obstacle
Mollick cites a paper in which a small group of financial professionals performed a complex valuation task using GPT-4o1 while researchers measured cognitive load from conversation transcripts. According to Mollick’s description, participants saw a productivity gain from AI but that gain was partly offset by the AI presenting information in ways that overwhelmed them: “giant walls of text, offers to pursue new topics, and sprawling discussions.” The chatbot interface, Mollick writes, appeared to be the obstacle rather than the work itself. He notes that less experienced workers — “exactly the people who could benefit the most from AI” — were most affected.
Specialized and consumer interfaces
Mollick surveys existing alternatives. He describes Claude Code (Anthropic), OpenAI’s Codex, and Google’s Antigravity as the most complete specialized interfaces, built primarily for programmers and assuming knowledge of Python and Git. He describes their interfaces as resembling “a 1980s computer lab.”
On the consumer side, Mollick discusses Google’s Stitch (described as an AI-native design canvas where users describe an app and receive multiple interconnected screens), Pomelli (which generates on-brand social media campaigns from a website URL), and NotebookLM (for working with diverse information sources). He calls these “a bit rough around the edges” but indicative of where specialized interfaces for non-developer knowledge workers may be heading.
Mollick also describes OpenClaw, which he characterises as “the fastest-growing open source project in history.” He describes OpenClaw as a personal agent that integrates with WhatsApp, Telegram, or Slack to take actions on a user’s computer. He calls it “a security nightmare” while acknowledging its growth.
Claude Cowork with Dispatch
Mollick describes Anthropic’s Claude Cowork as a version of Claude Code for knowledge workers, giving Claude access to local files and applications through a desktop workspace with connectors to outside apps and a computer-control fallback. Dispatch, which he says arrived in recent weeks, adds a mobile component: users scan a QR code and can then message Claude from their phone while it operates on their desktop.
Mollick describes asking Claude from his phone to prepare a morning briefing from his calendars, emails, and online channels, and to check whether a graph in a recent presentation was up to date. He reports that the agent opened the PowerPoint, searched his computer for more recent data, downloaded a PDF when he provided a link to an updated paper, extracted the relevant graph, and updated the file — with one point of failure where a site blocked a download.
He describes Cowork as sandboxed and safer than OpenClaw but more limited, with a growing but incomplete connector ecosystem and computer control that is “error-prone in practice.” He says the core insight is that “people don’t want a chatbot. They want an agent that works on their actual files, with their actual tools, accessible the way they talk to people.”
On-demand interfaces
Mollick closes by noting that Claude recently gained the ability to generate interactive visualisations directly in conversation — not static images, but adjustable outputs Claude can modify in response to follow-up questions. He describes this as a different approach: rather than companies building specialized interfaces in advance, the AI generates the right interface for the moment.
This piece is based solely on Mollick’s account in One Useful Thing.