AI-Driven Microcopy: How Voxly Used LLMs to Personalize Error Messages

AI · 4 min read

Voxly faced a recurring UX problem: generic error messages frustrated users and generated a high volume of support tickets. The product team proposed using a light-weight LLM to tailor microcopy based on context — for example, suggesting concrete next steps after a failed audio upload or giving empathetic explanations when a session disconnects.

To maintain safety and control, Voxly built a constrained generation pipeline: the LLM produced draft microcopy that passed through a dynamic template layer enforcing tone, length, and content filters before rendering. Critical content (billing, policy infractions) bypassed generation and used pre-approved templates. The team ran offline evaluations comparing baseline copy to LLM drafts using metrics like clarity, actionability, and user sentiment from moderated studies.

A controlled rollout reduced support tickets related to upload and connectivity errors by 28% and increased task completion on retry flows by 21%. Engineers highlighted the maintainability benefits: updating global tone or adding new templates propagated across contexts without manual translation work. Voxly's documentation emphasizes audit logs, human-in-the-loop checks for escalations, and clear fallback strategies if the model returns undesired outputs.