The gap between AI conversation and AI delegation is costing professionals real work. A UX researcher at The Context Window observed seven people working with AI tools across real tasks and found one consistent failure: users write exploratory, ambiguous prompts shaped by years of Slack and WhatsApp habits, then hand those same prompts to agents that resolve ambiguity silently, by committing to a single interpretation and acting on it. This is not hallucination. It is misinterpretation, and the distinction matters. The researcher's own Claude-based job-scoring script flagged 'new entries' by discovery date, not by score threshold, because 'new to me' was never specified. The system chose the more plausible reading. It was wrong.
The observation surfaced three user archetypes: the collaborative, who treat AI like a colleague and talk through their thinking; the commanding, who are blunt but forget to supply context; and the over-explainers, who bury the actual instruction in surrounding thought. None performed clearly better than the others. What cut across all three groups was deference: most users accepted AI suggestions even when those suggestions had drifted from original intent. The machine guesses, the human ratifies, and the error compounds silently. The researcher's proposed fix is direct and worth examining: before sending any brief, scan it for question marks and modal verbs like 'could', 'would', 'maybe'. Each one marks an unresolved decision. Separate them out. What remains is an actual brief.
The central design argument here is what makes this worth reading in full. Conversational interfaces carry social signals, turn-taking, acknowledgement, rhythm, that create the false impression of a shared mental model. Most products lean into that fluency and leave users to find the seams themselves. The researcher argues the opposite approach: the more agentic the task, the more the interface should resist conversational fluency. That principle has real consequences for anyone building or buying agentic tools in 2025, and the methodology section alone, a deliberate first pass at seven participants framed explicitly as pattern-detection, not proof, is a model of intellectual honesty rarely seen in this space.
[READ ORIGINAL →]