Catch users before they give up, not after they've already left

Create an AI-derived attribute 'recent_friction_signal' on the User object, refreshed in real-time. Patterns detected: (a) repeated failed actions on the same UI element (3+ attempts at same form / button within a session); (b) abandoned setup step (started a multi-step flow, didn't complete, didn't return for 24+hrs); (c) error-encountered events without subsequent retry success; (d) help-content visits without subsequent action; (e) rage-click or rapid-back-button patterns. Output: structured object with friction_type, friction_location (where in product), severity (mild / moderate / severe based on time-stuck and repetition), and recommended_intervention.

Build a segment 'Active friction signal' capturing users with recent_friction_signal in the last 7 days where severity is moderate or severe. Real-time refresh. Excludes users already in active intervention from this journey (no double-poking). Partitions by friction_location so the journey routes contextually.

Generate in-app contextual help variants per friction_location. For setup-abandonment: contextual tooltip at the abandoned step with the answer to the common blocker. For repeated-failed-action: floating help bubble with 'Looks like you're stuck — here's what most users do' + GIF or short video. For error-encountered: helpful error explainer rendered where the error appeared. For decision-paralysis (long dwell time on a critical step): comparison guide or recommendation card. Each renders mid-session when the friction is freshest. Tone: helpful colleague, not aggressive sales.

Generate email fallback content sent if the user doesn't engage with the in-app help OR has left the session before help was shown. Variants per friction_location. Format: subject directly addresses the issue ('Stuck on [specific step]? Here's the fix.'), body is a 60-second tutorial (video link or step-by-step screenshots), with a one-click resume-where-you-left-off deep link. Tone: not generic 'we noticed you got stuck' — specific to the exact friction.

Configure an AI agent scenario 'Friction rescue' that engages users with severe friction signal who haven't responded to in-app help OR email. Scenario: agent opens with awareness of the specific friction ('I see you've been working on [step] — happy to help walk through it'), offers to: (a) co-pilot the action with screen-share or step-by-step in chat, (b) escalate to human support if the problem is technical/account-level, (c) flag the friction as a product bug for the engineering queue. Agent NEVER pretends to be human; handoff to human is offered when agent can't resolve in 3 turns.

Build a graduated-escalation journey triggered when recent_friction_signal becomes moderate or severe. Touch 1 (mid-session, immediate): in-app contextual help renders at the friction location. Touch 2 (4 hours later, if user left without resolving): email fallback with tutorial + resume-link. Touch 3 (24 hours later, if friction signal still active OR user hasn't returned): trigger agent scenario or, for severe-tier high-value users, create a CSM task for direct outreach. Exit on: friction_signal resolves (user completed the step — success), user explicitly dismisses help 3 times (respect — don't pester), unsubscribe.

Compose a friction intervention dashboard: top 10 friction locations by frequency (a product-team-priority signal — these are the UX bugs the data is telling you about), in-app-help dismiss rate (high = irrelevant help, low = useful), friction-to-resolution rate by intervention touch (which level of escalation actually resolves), severe-tier-to-churn correlation (do users stuck severely actually churn — proves the intervention's necessity), and journey-attributed activation lift (does this play meaningfully move activation rate). The product team gets a free UX-research stream as a side effect.