The usual reason an AI support agent gives poor answers is not that it lacks adjectives. It is that the business has several contradictory documents, outdated prices and no agreement about what the bot is allowed to say.
Inventory questions from actual conversations
Review a sample of recent support tickets and enquiries. Remove personal details and group the questions by the outcome the customer wanted. "Where do you deliver?" and "Can I get this in my area?" may be one intent. "Can I change my delivery address?" is different because it could modify an existing order.
For each group record question variants, the approved answer, the system or document that confirms it, and the human owner who can approve changes.
Give every answer a small contract
A support-ready article needs a clear title, last verified date, responsible team, supported customer scenario and explicit exceptions. State what must never be promised. For example, a shipping article might say "delivery estimates are confirmed at checkout" rather than citing an attractive but unverified two-day guarantee.
Example answer record
- Intent: request a quotation
- Approved reply: explain information needed for a quote
- Never do: generate or guarantee a price
- Escalate when: customer requests custom terms
- Owner: sales lead
- Review: on every pricing change
Separate retrieval from permission
Finding a piece of content does not give the assistant authority to act on it. A knowledge article may describe refunds, but that does not mean a bot should issue refunds or promise approval. Actions that change accounts, reveal personal information or move money need separate authentication and authorization controls.
Instructions found in retrieved content can also be malicious or simply wrong. The OWASP guidance on prompt injection explains why external text must not be treated as trusted instructions.
Build confidence rules around missing knowledge
Write what the bot does when two articles disagree, when no source matches, and when the customer asks for a human. A safe response can ask a specific clarifying question, quote an approved policy, or escalate with context. It should not hide uncertainty behind a confident sentence.
Try realistic variations: spelling errors, multiple questions in one message, screenshots referenced but not supplied, and messages in language your support team actually receives.
Create a maintenance habit
Review failed and corrected conversations weekly. Tag their underlying causes: missing article, stale policy, poor routing or an integration failure. Assign each correction an owner and a date. Track whether the same question fails again after changes.
Start with 20 well-maintained answers before trying to ingest a shared drive full of stale files. A small trusted library typically produces a more controllable pilot than an uncontrolled document dump. Learn how that fits into an AI WhatsApp agent setup.
Use an answer card, not a document dump
An AI support system is easier to govern when each answer has a deliberate scope. Instead of uploading an old 80-page policy document and hoping the right paragraph is retrieved, break frequent questions into individual reviewed answer cards. Each card should say who it applies to and when it must not be used.
| Field | Example value | Why it matters |
|---|---|---|
| Question | Do you offer weekend appointments? | Matches customer wording |
| Approved answer | Saturday by arrangement, subject to confirmation | Avoids invented availability |
| Exception | Public holidays require manual confirmation | Prevents overpromising |
| Owner | Scheduling manager | Someone can approve changes |
| Review trigger | Booking policy changes | Keeps the answer current |
Test ambiguity before launch
Customers rarely use your internal category names. A request such as "Can someone come tomorrow?" might mean a repair visit, a sales demo or a delivery. Write at least five alternative phrasings for each high-volume question, including abbreviated language, typos and a question with missing details. A reliable agent asks for the missing detail instead of deciding for the customer.
Build a small rejection set too. It should include a question with no approved answer, a request for a special discount, an attempt to overwrite the agent's instructions, and a question about another customer's account. If the system provides confident answers anyway, the knowledge base is not production-ready.
Keep ownership after the setup is finished
Choose one person who can approve new answers and another who can review handoff failures. Hold a short weekly review during the first month: which questions were unanswered, which answers were corrected, and which topics should stay human-only? Track decisions in the same change log as the answer cards. For the complete deployment sequence, use our AI WhatsApp setup checklist.
Reference notes
- OWASP Prompt Injection
- NIST AI Risk Management Framework
This is a practical editorial guide, not a claim that any setup guarantees a specific outcome. Verify platform features, regulations and prices before making business decisions.
