[object Object]

Freddy Self-Service is not a “switch on and forget” feature. The teams hitting 30%+ deflection share three habits: aggressive source curation, intent confidence thresholds tuned to brand risk, and a hard escalation contract.

Curate sources, do not dump them

The default behavior crawls every published Freshdesk solution article. That is a mistake. Tag a freddy-source: true custom field on articles you actually trust, then filter the corpus via the Self-Service settings → Source filters panel. Pages older than 18 months get auto-deprioritized.

Source priority order:
1. Solution articles tagged "verified-2026"
2. Public release notes
3. Help center FAQs
4. Community posts (read-only, never primary)

Set the confidence floor by category

A billing question deserves a 0.85 confidence threshold. A “how do I reset my password” gets 0.6. Configure per-category thresholds via the Freddy Admin → Confidence rules screen. Anything below the threshold falls through to a live agent without showing a half-broken answer.

The escalation contract

Self-Service should always offer a one-tap “Talk to a human” link in the response card. Use the escalate_to_agent action with the originating intent attached so the agent sees what the bot tried first.

{
  "action": "escalate_to_agent",
  "context": {
    "intent": "refund_request",
    "confidence": 0.71,
    "attempted_articles": ["sol_3382", "sol_4921"]
  }
}

Measure the right deflection metric

Resolved-without-agent is vanity. Track CSAT-on-deflected-conversations and 24-hour reopen rate. If reopen rate exceeds 8%, your confidence floor is too low.

Retraining cadence

Pull the weekly “low confidence” export from the Freddy dashboard. Group by intent, write three new training utterances per cluster, and republish. Monthly, prune sources that produced any negative feedback above 5%.

What to do this week

Tag your top 50 solution articles as verified, set a 0.8 confidence floor on billing and account categories, wire the escalation context payload, and export last week’s low-confidence questions to seed retraining.

[object Object]
Share