Define deflection as a user who intended to contact support but successfully resolved the issue through community content. Use interstitial prompts like “Did this answer your question?” with randomized exposure, log intent signals, and apply holdouts. Triangulate with agent-tagged references to threads and natural language matching between cases and discussions.
Quantify the cost of assisted channels versus community self-service. Calculate fully loaded cost per contact by channel, including staffing, tooling, and overhead. Multiply reliable deflection counts by avoided cost, then subtract community platform, moderation, and content program expenses. Present ranges with sensitivity tables to reflect uncertainty and drive pragmatic decisions.
Strengthen credibility through tests. Run geography or user-segment holdouts where forum search suggestions are disabled, compare case rates, and compute confidence intervals. Complement with pre/post experiments around feature launches. Document assumptions, seasonality effects, and edge cases, and invite support leaders to validate results before celebrating savings publicly.