AI Workflow Recommendations

Instrument workflow events in PostHog: workflow_started, workflow_completed with {workflow_type, duration_seconds}, feature_used with {feature_name}, and at least 1 friction signal (action_undone, error_encountered, or same_action_repeated). Wait 3-5 days for data.

Identify 10-20 test users in PostHog: active 14+ days, login 3x/week, not power users, use at least 1 core workflow. Create cohort 'workflow-suggestion-smoke-test'.

Build the power user benchmark: create a PostHog cohort of top 10% users by usage volume. Extract their feature usage, workflow completion times, and shortcut adoption rates.

For each test user, run path analysis via PostHog HogQL: query their last 50 events, compare workflow times and feature usage against the power user benchmark, identify repeated manual patterns.

Call the Claude API (Sonnet 4.6) for each user with their behavior data, undiscovered features, and power user benchmark. Generate 1-3 specific, quantified suggestions per user.

Human review: verify each suggestion references correct features and makes sense for the user's workflow. Edit or reject generic suggestions.

Create 1 Intercom in-app post per suggestion targeted to the specific user by user ID. Format: benefit headline, 1-2 sentence body with quantified gain, 'Try it now' CTA linking to the feature. Display: show once on next session start.

Log suggestion_delivered events in PostHog with {suggestion_id, suggestion_category, suggestion_text} for each delivery.

Monitor for 7 days: query PostHog to check if each user performed the suggested action after the delivery timestamp.

Calculate acceptance rate: users who adopted / users who received. Pass threshold: >=30%. If pass, document top-performing suggestion categories and proceed to Baseline. If fail, diagnose: relevance, timing, or complexity issue.