LadderTeam automates UX interviews with dual-agent LLM, hits 99.1% convergence

Eliciting detailed, actionable software requirements from end-users is a bottleneck for product teams. Traditional laddering interviews—where an interviewer systematically probes for deeper reasoning—are effective but manual, expensive, and hard to scale. Enter LadderTeam, a framework from researchers Manjushree Aithal, Alexander Kotz, and James Mitchell, accepted at ACM AI Summit 2026. It automates UX wireframe interviews using a dual-agent LLM architecture. An active interviewer agent executes one of three probing strategies—ACV (attribute-consequence-value), 5-Whys, or JTBD (jobs-to-be-done)—to extract actionable requirements from usability feedback. Meanwhile, a concurrent background judge agent evaluates probe-response pairs in real time and triggers guardrails to prevent topic drift, keeping conversations focused and productive.

rigorous evaluation, the team introduced a controlled simulation methodology using scripted ground-truth transcripts, isolating probe quality as the sole experimental variable. Across 216 simulated interviews, LadderTeam achieved a 99.1% chain convergence rate and an 81.0% ground-truth actionable response match. It handled personality variance effectively: 86.1% accuracy with reluctant personalities and 75.9% with terse ones—all with zero drift across every run. These results suggest LadderTeam can replace expensive human interviewers for structured requirements gathering, delivering consistent depth while scaling across thousands of users. The authors promise to open-source all evaluation code, transcripts, inputs, and a live demonstration platform upon acceptance, making it immediately adoptable for UX researchers and product managers.