Skip to main content
Use these techniques to get the best performance out of StoreTalk’s AI staff. The basics get the AI running — this guide covers what separates a mediocre setup from a great one.

​
Knowledge base structure

The single biggest factor in AI response quality is how the knowledge base is structured. Use the patterns below for consistent, accurate answers.
Use # and ## headings to group content. The AI uses headings as context anchors — it knows that a price listed below a heading belongs to the product in that heading.
# Product: Wooden Study Table

## Dimensions
Width: 120 cm | Depth: 60 cm | Height: 75 cm

## Pricing
Standard finish: ₹8,999
Premium finish (walnut): ₹11,999
Custom size (add ₹2,000 to base price)

## Delivery
Delhi NCR: 3–5 business days (free)
Other cities: 7–10 business days (₹499 shipping)
Write FAQs exactly as customers ask them. The AI matches customer questions against the question text in your knowledge base — closer phrasing means better matching.
Q: Can I get the table in a custom colour?
A: Yes — we offer custom RAL colour finishes for an additional ₹1,500. Contact us for samples.

Q: Do you offer EMI?
A: Yes — 0% EMI on all orders above ₹5,000 using HDFC or ICICI credit cards.
If there are things the AI should never mention — discontinued products, prices under renegotiation — add an explicit exclusion section.
## Discontinued products (do not quote these)
- Plastic Chair Model PC-100: discontinued as of March 2025, do not offer or quote
- Budget Desk Model BD-50: out of stock indefinitely, redirect to Model BD-60

​
Persona tuning

The Dos & Don’ts field is the most powerful lever for shaping AI personality and behaviour. Use it to set communication style, enforce business rules, and define escalation triggers. Communication style: 
Do: Use simple, friendly language. Avoid jargon.
Do: Use short paragraphs — 2–3 sentences maximum per reply.
Do: Address the customer by their first name if you know it.
Hard business rules: 
Don't: Offer any discount not listed in the knowledge base.
Don't: Confirm stock availability — say "subject to availability" instead.
Don't: Quote delivery dates — say "typically X days" instead.
Escalation triggers: 
Do: If the customer asks for a bulk order above 20 units, collect their name and requirement and say our sales team will call them.
Do: Escalate to a human if the customer expresses frustration or uses the word "complaint".

​
Escalation logic

AI staff escalates to a human agent when:
  1. The customer explicitly asks for a human
  2. The AI’s confidence falls below the threshold — typically after 2–3 failed answer attempts
  3. A custom escalation rule in Dos & Don’ts is triggered

​
Reducing false escalations

If the AI escalates too often, the knowledge base is likely too thin or too vague. Before adjusting escalation settings:
1

Review escalated conversations

Check the Inbox and filter for recent escalated conversations.
2

Identify the trigger

Read what question or message caused the escalation.
3

Add a direct answer

Add a clear, specific answer for that question to the knowledge base.
4

Re-test

Send the same message again to confirm the AI now handles it correctly.

​
Reducing missed escalations

If the AI is confidently answering questions it shouldn’t — such as confirming prices not in the knowledge base — add an explicit Don’t: 
Don't: If you're not certain of a price from the knowledge base, say "Let me get that confirmed for you" and escalate to a human.

​
Multi-staff routing

When a client has multiple AI staff members, StoreTalk automatically routes each incoming message to the most relevant specialist. To make routing accurate:
  • Give each staff member a distinct, narrow scope — a Sales Executive should not also handle return policies
  • Use one knowledge base file per topic rather than a single large file for everything
  • Set each staff member’s language explicitly — if you have a Hindi-only staff member for regional customers, set them to Hindi only
Run at least 20 test messages after a new setup before handing over to the client. Cover edge cases: questions the AI shouldn’t answer, requests for discounts, complaints, and multi-part questions.