SmartForms: when an LLM-powered onboarding slowed signups — a UX postmortem

AI · 6 min read

SmartForms, a B2B forms startup, introduced a natural-language prompt that populated profile fields using an LLM trained on optional user inputs. The idea was to reduce typing and speed onboarding, but product metrics showed a 14% rise in drop-off on the second step of the flow. Heatmaps and session replays revealed users hesitated at AI-populated fields and frequently erased content rather than accept it.

Qualitative interviews uncovered two root causes: users didn’t understand where the suggested content came from and were worried about privacy and accuracy, and the autogenerated text often used fuzzy phrasing that didn’t match the user's domain terminology. The UX decision to present the AI suggestions inline without explicit labeling produced cognitive friction.

The remediation combined three moves: explicit provenance labels (“Suggested by AI — edit freely”), a lightweight confidence score and the option to accept all or edit individual fields, and a toggle to disable AI suggestions. After implementing these changes and running another experiment, abandonment returned to baseline and Net Promoter Score on onboarding rose by 6 points.

The case highlights a broader lesson for designers shipping LLM features: prioritize transparency, control, and domain-appropriate language. AI can accelerate tasks, but without clear signaling and easy reversibility it can feel intrusive rather than helpful.