AI Configuration
AI Configuration
Supportly works best when the assistant is treated like a constrained support operator. Strong guidance, clean retrieval sources, and explicit escalation boundaries matter more than clever phrasing.
Model routing and failover
Supportly supports standard-provider models plus managed Pro presets. Pro presets keep upstream provider details hidden and can automatically fail over to a workspace-selected standard fallback model when retryable Pro errors occur.
Fallback redirects create internal-only conversation notices for operators. They appear in dashboard timelines and ticket views, but stay hidden from customer-visible widget history and transcripts.
System prompt best practices
Write the system prompt as a compact operating policy. Say what the assistant should optimize for, what it must never do, and when it should escalate. Avoid vague instructions like be helpful or always delight the customer unless they are followed by concrete rules.
Prompt example
Start from a practical baseline like this and then tune it using real conversations.
1You are Supportly for Acme Support.23Answer only using the workspace knowledge base and the details provided by the customer.4If the customer asks for refunds, billing reversals, account ownership changes, or anything you are not confident about, escalate to a human agent.5Keep answers concise, clear, and procedural.6Ask one clarifying question when required, but do not loop.7If a request requires internal action, say so plainly and hand off.- Keep the persona name aligned to the brand or support team name.
- If AI-assisted closure is enabled, train the assistant to offer closure only after the issue is clearly resolved.
- Require the customer to reply with the exact word close before ending the ticket automatically.
- Prefer operational verbs like verify, ask, escalate, summarize, and confirm.
- Review prompts after actual escalations rather than tuning in the abstract.
Knowledge base tips
Retrieval quality depends on clean source material. FAQ entries should answer one question fully. Uploaded documents should be scoped to one policy or workflow whenever possible.
If an answer requires hidden context to make sense, it is not a good knowledge base candidate yet. Rewrite it so the retrieved chunk can stand on its own.
| Source type | What works well | What to avoid |
|---|---|---|
| FAQ | Short question plus direct answer | Bundling multiple workflows into one answer |
| Uploaded docs | Single-purpose policies and SOPs | Long mixed-topic exports with stale sections |
| Escalation notes | Patterns observed from handoffs | Private internal details that should never be surfaced to customers |
Escalation rules guide
Tune escalation using confidence threshold, sentiment detection, maximum AI turns, and explicit trigger phrases. These controls should reflect the actual risk tolerance of your support team.
Escalation rules payload
The dashboard UI edits the same shape shown below.
1{2 "confidence_threshold": 72,3 "sentiment_detection": true,4 "max_ai_turns": 8,5 "escalation_keywords": [6 "speak to an agent",7 "manager",8 "real person",9 "cancel my account"10 ],11 "allow_full_ai_resolution": true12}Good review loop
Look at escalated tickets every week. If the AI escalates too early, improve the knowledge base. If it escalates too late, tighten the prompt and lower the confidence tolerance.