A customer support assistant can retrieve a polished answer and still give the wrong price, deadline, or policy. The hard part is deciding which information is approved for the customer, which version applies, and what to do when the source is missing or contradictory.
Short answer
Start with the questions your team actually receives. Give each answer an owner, an authoritative source, an effective date, a review trigger, and a permitted audience. Test retrieval against ambiguous and outdated questions before allowing an AI assistant to answer. When evidence is missing, route the question to a person.
An AI knowledge base is more than a folder of PDFs. Intercom documents a distinction between customer-facing help articles and the broader knowledge hub available to its AI products. Microsoft documents knowledge sources for Copilot Studio and a setting that can restrict responses without a source. Those are product capabilities, not proof that a particular answer is correct or that every channel preserves its citations.
Decide what may become an answer
Build the first inventory from support conversations, approved policies, and product records. Do not copy an entire inbox into an AI tool: it may include private customer details, exceptions that were granted once, and superseded promises.
| Candidate information | Publish for self-service? | Let AI answer? | Control required |
|---|---|---|---|
| Stable opening hours, verified for a location | Usually | After source check | Location and holiday exception |
| Current product instructions | Usually | After owner approval | Version and applicability |
| Price, promotion, stock or delivery promise | Only with a current authoritative source | Only when current state is retrievable and confirmed | Expiry, region and escalation |
| One customer’s refund exception | No | No | Keep in the authorized case record |
| Legal, safety or account-specific decision | Only an approved general explanation | Route to an authorized person | Identity and case review |
This matrix is an editorial decision model. It does not grant legal permission to use personal data or publish a policy.
A copyable answer contract
Create one record per answer, rather than one document for every topic. The contract below is a Pulse model that makes changes reviewable:
question_id: KB-014
customer_question: "Can I change an appointment?"
approved_answer: "Use the booking link or contact the team."
authoritative_source: "approved_booking_policy_v3"
source_owner: "operations"
audience: "public"
region_and_language: "US / English"
effective_from: "date and time zone"
expires_or_review_when: "policy change or 90-day review"
status: "draft | approved | suspended | retired"
last_approved_by: "authorized role"
escalate_when: "booking is already in progress or policy is unclear"
The exact fields are less important than their effect: a reviewer can identify which answer was active when a customer saw it. Do not expose internal notes or customer records through a public source.
Run a five-question release test
Before connecting the approved set to a chatbot, ask it five fictional questions and inspect both answer and cited source:
- A common question with one valid source: does the response match the approved text?
- A question with an old and a new policy: does it use the effective version?
- A regional or language variation: does it avoid borrowing a rule from another market?
- An account-specific request: does it hand off rather than infer private facts?
- A question with no approved source: does it say it cannot confirm and route the case?
The release fails if any material answer is unsupported, uses a retired source, leaks internal content, or gives a definitive answer where the contract requires a human. Fix the source or routing and repeat the failed case. Do not count a plausible-sounding answer as a pass.
Keep changes from becoming silent errors
The owner should suspend a record when a price, service area, policy, product version, or regulatory requirement changes. Publish the replacement only after approval and a regression test against the old question. Maintain a short change log with record ID, former version, new version, approver and affected channels. Check that cached answers and public pages also reflect the change.
Measure the share of answered questions with a valid source, handoffs with a clear reason, corrections reported by staff, and time from source change to published correction. A rising automation rate alone can hide more wrong answers.
For the customer conversation itself, use the WhatsApp service control flow to define handoff and external-send boundaries. For a suspected wrong answer already sent, use the automation failure and continuity sheet to identify affected customers before correcting anything.
Limits and sources
This guide designs a small, testable publishing process. It does not evaluate any chatbot’s accuracy, replace a privacy review, or claim that a vendor will enforce the proposed contract automatically. Intercom’s content guide supports organizing and reviewing help content; Microsoft’s knowledge source documentation describes grounding options and limitations. The inventory matrix, record contract, release test and change-control sequence are original Pulse editorial models that need validation in the chosen system.