When should the AI quote the Handbook?
Never. Handbook procedures shape what the AI does — they’re instructions, not customer-facing content. The customer should never see Handbook text quoted back at them.
This is the single most important thing to internalize about the Handbook.
Why this matters
The Handbook is full of language that would feel wrong if a customer saw it. Procedures look like:
When a customer reports a duplicate charge: 1. Confirm identity by asking for the order number plus the email on file. 2. If verified, refund both charges and offer a 10% goodwill credit.
That’s a runbook for the AI. A customer reading “if verified, refund both charges and offer a 10% goodwill credit” would feel like they’d been handed an operator script — because that’s what it is.
Atender’s AI agents follow Handbook procedures internally and respond to customers in their own natural language. The AI:
- Reads the relevant procedure for the situation it’s in.
- Uses the procedure to decide what to do.
- Responds to the customer in conversational language that reflects the procedure’s intent.
It doesn’t quote.
What if I want the customer to see specific words?
If you want the customer to see specific words verbatim, those words belong in a Knowledge Base article, not a Handbook procedure. The split:
- Customer-facing content → Knowledge Base. Written for the customer. The AI can quote, link, or summarize freely.
- AI behavior shaping → Handbook. Written as instructions. The AI follows them; the customer never sees them.
If a procedure includes a sentence like “tell the customer that…” that’s a sign you’ve put customer-facing content in the wrong place. Move that sentence to a KB article and reference it from the procedure (“Use the language in the duplicate-charge KB article when explaining the refund.”).
Confidential procedures
Procedures marked Confidential deserve extra care. They’re meant for the AI’s internal decision-making — escalation contacts, sensitive thresholds, policy details that shouldn’t surface broadly. The same rule applies: the AI follows the procedure, doesn’t quote it.
If you’re worried that the AI might leak confidential content into a customer-facing reply, the safest test is to ask: “Could the customer infer this from the AI’s natural response?” If yes, restructure the procedure so the actionable instruction is separated from the sensitive context.
A practical example
Consider this Handbook procedure for a VIP customer escalation:
When a Tier-1 VIP customer reports a service outage: 1. Acknowledge the VIP status first thing. 2. Offer a 30-day service credit if asked, otherwise don’t volunteer it. 3. Transfer to the on-call manager (extension 4012, available 24/7). 4. Tag the conversation
vip-escalation.
What the customer sees in chat:
“Thanks so much for being a long-time customer — I’m sorry you’re hitting this issue. Let me get this in front of one of our managers right now so we can get you sorted quickly.”
What the customer never sees:
- The exact tier label.
- The extension number.
- The “if asked, otherwise don’t volunteer it” condition on the credit.
- The escalation tag.
The procedure shaped what the AI did. The customer experienced a thoughtful response.