AI integration that users actually finish using

Integrated language models into an existing product where they reduced real friction—not where they looked impressive in demos.

Problem

The team wanted AI features. Early experiments felt clever but didn't change completion rates for core workflows.

Approach

Mapped user jobs-to-be-done, identified three high-friction moments, designed AI as assistive scaffolding with clear human control, and instrumented outcomes.

Outcome

Task completion improved on the targeted flows. Support load for those paths decreased. Features stayed explainable to customers.

Lessons learned

AI is a material, not a product. The product is still the job the user is trying to finish.