Overview
The Smart Policy Recommender analyzes your interaction history data to identify contact frequency patterns and recommend optimal contact policies. It looks at how often customers are contacted, through which channels, and how they respond — then suggests policies that reduce fatigue and improve engagement.
Navigate to AI > Policy Recommender in the sidebar.
What It Analyzes
The recommender examines three dimensions of your interaction data:
- Frequency bands — How contact frequency correlates with response rates. It identifies the “sweet spot” where customers are contacted often enough to stay engaged but not so often that they disengage.
- Channel performance — Which channels (email, SMS, push, web) produce the best response rates and at what frequencies.
- Time-based patterns — How response rates change based on time-of-day, day-of-week, and recency of last contact.
Dual-Tier Analysis
The recommender supports two analysis tiers:
| Tier | When Used | Method |
|---|
| LLM | Default, or when ML Worker is unavailable | Samples interaction data and uses the configured LLM to identify patterns and generate recommendations |
| ML Worker | When ML Worker is connected and dataset exceeds 5K rows | Runs statistical analysis using sklearn — frequency binning, correlation analysis, and optimal threshold calculation |
Reading the Frequency Charts
The recommender displays an interactive frequency analysis chart:
- X-axis — Contact frequency (contacts per time window)
- Y-axis — Response rate (percentage)
- Curve — Shows how response rate changes as contact frequency increases
- Optimal zone — Highlighted region where the frequency-to-response ratio is highest
- Current average — A marker showing your current average contact frequency
The chart helps you visually confirm the AI’s recommendations before applying them.
How Recommendations Work
After analysis, the recommender generates one or more contact policy recommendations. Each recommendation includes:
- Policy name — A descriptive name (e.g., “Email frequency cap — 3 per week”)
- Channel — Which channel the policy applies to (or “all” for global policies)
- Frequency limit — The recommended maximum contacts per time window
- Time window — The period over which the limit applies (e.g., 7 days, 30 days)
- Expected impact — Estimated improvement in response rate
- Confidence — Statistical confidence in the recommendation
- Supporting data — The frequency bands and sample sizes behind the recommendation
Applying Recommendations
When you click Apply on a recommendation:
- A new Contact Policy is created in draft status
- The policy is pre-configured with the recommended frequency limit, time window, and channel scope
- You are taken to the Contact Policies page to review and activate it
Applied policies are always created as drafts. They do not affect live decisioning until you explicitly activate them.
Best Practices
- Run analysis regularly — Customer behavior changes over time. Re-run the recommender monthly or after major campaign changes.
- Start with global policies — Apply broad frequency caps first, then refine per-channel.
- Compare tiers — If both LLM and ML Worker are available, compare their recommendations. The ML Worker is more precise for large datasets; the LLM can surface qualitative insights the ML Worker misses.
Advanced Parameters
Each policy analysis run can be fine-tuned using the Advanced Parameters panel. Expand it to adjust:
| Parameter | Default | Description |
|---|
| Daily Cap | 3 | Max messages per customer per day |
| Weekly Cap | 10 | Max messages per customer per week |
| Monthly Cap | 30 | Max messages per customer per month |
| Lookback Days | 90 | How far back to analyze interaction history |
| Min Sample Size | 100 | Minimum interactions before analysis is meaningful |
Large Dataset Warning
When the interaction history contains 5,000 or more rows, a confirmation dialog appears before analysis begins. The dialog shows:
- Accuracy comparison — ML Worker uses logistic regression and statistical analysis on the full dataset vs. LLM sampling
- Estimated cost — Token count and approximate cost if proceeding with LLM
- Speed comparison — ML Worker processes locally in seconds vs. LLM round-trip
You can choose Use ML Worker (recommended for large datasets) or Proceed with LLM (uses sampled data).
For large interaction histories, the ML Worker produces more precise frequency recommendations because it processes the full dataset. See ML Worker Setup for deployment instructions. Next Steps
