The most common reason AI customer service underperforms is not the AI model. It is the knowledge base underneath it. When a business uploads vague, outdated, or incomplete information, the AI gives vague, outdated, or incomplete answers. When the knowledge base is thorough and current, the AI responds with the accuracy of your best-informed team member — at any hour.
The strongest public example of this comes from Breathe, a UK-based HR software company. Their AI resolution rate started at 56%. After systematic knowledge base improvements — rewriting articles, creating response snippets, and adding guidance for edge cases — the resolution rate climbed to 82% within nine months and eventually reached 88%. The AI did not change. The information feeding it did.
This guide provides a practical framework for building and maintaining an AI knowledge base, based on documented best practices from Zendesk, Intercom, Gorgias, and real implementation data.
What Should You Upload First?
The temptation is to make the AI conversational and friendly before making it accurate. That is backwards. Upload canonical, policy-grade content first — the factual information that resolves the most customer queries.
Priority 1: Policies and high-volume service information. Return and cancellation policies. Shipping and delivery terms (timelines, fees, zones). Payment methods accepted. Operating hours and holiday schedules. Contact information and location details. These answer the questions customers ask most frequently and generate the highest volume of repetitive messages.
Priority 2: Product and pricing information. Product specifications, materials, sizing. Current pricing (including any promotions). Service descriptions and what is included. Package or tier comparisons. This information drives pre-purchase decisions and directly affects conversion.
Priority 3: Edge cases and workarounds. Known issues with products or services. Common troubleshooting steps. Exceptions to standard policies. What to do when something goes wrong. These are the queries that trip up poorly configured AI — and the ones that matter most when they occur.
Priority 4: Channel-specific guidance. How to book via WhatsApp. How to use the web chat. How to reach a human for complex issues. This helps customers navigate the specific channel they are using.
Priority 5 (last): Tone and sales optimisation. Brand voice guidelines, upsell suggestions, greeting customisation. Important, but only after the factual foundation is solid.
How Should You Structure Knowledge Base Articles?
One intent per article. Zendesk's documentation notes that article titles strongly affect whether the AI retrieves the correct article for a customer's query. A broad article titled "Everything About Delivery" performs worse than three specific articles: "How much does shipping cost?", "Where is my order?", and "Can I change my delivery address?"
Title articles as questions. Match how customers actually ask. "What is your return policy?" retrieves better than "Returns Policy Information" because customers type questions, not category labels.
Lead with the answer. The first sentence should directly answer the question. Supporting details, exceptions, and links come after. AI systems extract from the beginning of articles — if the answer is buried in paragraph four, the AI may return the context without the answer.
Keep articles concise. Each article should be 100–300 words. If an article needs more than 300 words, it probably covers multiple intents and should be split.
How Often Should You Update the Knowledge Base?
The evidence is clear: treat the knowledge base as a living system, not a quarterly project.
Immediately: After any pricing, policy, or product change. If your menu changes, your hours change, or your return policy changes, the knowledge base must reflect it the same day. An AI quoting yesterday's price is worse than no AI at all.
Weekly: Review unresolved or low-confidence conversations from the past week. Identify questions the AI could not answer or answered poorly. Write or update articles to address them. This takes 15–30 minutes and produces the highest return on maintenance time.
Monthly: Audit the top recurring topics. Are there new questions emerging that your knowledge base does not cover? Are existing articles still accurate? This takes 30–60 minutes.
Quarterly: Full sweep for broken links, duplicated content, stale articles, and deprecated products or services. This takes 1–2 hours.
Gorgias specifically warns that articles must be accurate, published (not draft), and kept current when policies, product names, availability, or links change. Intercom describes knowledge base upkeep as a daily responsibility, not a periodic task.
What Does a Good Knowledge Base Look Like in Practice?
Here is an example structure for a retail business:
Shipping & Delivery (5 articles): How much does shipping cost? / How long does delivery take? / Do you ship internationally? / Can I change my delivery address after ordering? / Where is my order?
Returns & Refunds (4 articles): What is your return policy? / How do I start a return? / When will I receive my refund? / Can I exchange instead of returning?
Products (variable): One article per product category or per frequently asked product question. "What sizes are available in [product]?" / "What materials is [product] made from?" / "Is [product] in stock?"
Payments (3 articles): What payment methods do you accept? / Can I pay in installments? / My payment failed — what should I do?
General (3 articles): What are your opening hours? / Where are you located? / How do I contact a human?
That is approximately 15–20 articles covering the most common queries. Most SMEs can build this in 2–3 hours. Expansion happens organically as the weekly review surfaces new questions.
The Before and After: What Knowledge Base Quality Actually Changes
| Metric | Poor Knowledge Base | Well-Maintained Knowledge Base | Source |
|---|---|---|---|
| AI resolution rate | 56% | 82–88% | Breathe / Intercom (2025) |
| Automation rate at 30 days | Low (unspecified) | 30% of interactions | Dr. Bronner's / Gorgias (2025) |
| Automation rate at 60 days | Low (unspecified) | 45%+ of interactions | Dr. Bronner's / Gorgias (2025) |
| AI accuracy score | 3.55 out of 5 (Jan) | 4.08 out of 5 (Oct) — 14.9% improvement | Gorgias AI Agent (2025) |
The pattern is consistent: AI performance is a knowledge base quality problem, not an AI capability problem.
How Does This Work with Omago?
Omago, an AI agent platform that helps SMEs automate customer conversations across WhatsApp, Telegram, and web chat, uses your uploaded knowledge base to generate responses. You can paste text, upload documents, or link your website — the AI indexes the content and uses it as its source of truth.
The conversation flow builder complements the knowledge base: while the knowledge base handles open-ended questions ("What's your return policy?"), conversation flows handle structured journeys ("I want to book an appointment" → guided step-by-step qualification). Together, they cover the full range of customer interactions.
Their team currently provides hands-on setup support during the onboarding period, including knowledge base configuration — useful for business owners who want to ensure the initial setup is thorough.
Frequently Asked Questions
How many articles do I need to start?
Start with 15–20 articles covering your top questions. You can identify these by reviewing your last 50 WhatsApp or website messages — the same 10–15 questions will repeat. Build articles for those first, then expand based on what the AI escalates.
What format should knowledge base articles be in?
Plain text is best. Write in clear, simple language — not marketing copy. Avoid jargon unless your customers use it. Structure as question (title) → direct answer (first sentence) → supporting details → exceptions or edge cases.
Can I just link my website instead of writing articles?
Many platforms allow this, and it works for a quick start. However, dedicated knowledge base articles perform better because they are structured around customer questions rather than marketing messages. Your website says "We offer premium delivery options" — your knowledge base should say "Standard delivery costs $5 and takes 3–5 business days. Express delivery costs $12 and arrives next business day."
What happens if the AI encounters a question not in the knowledge base?
A well-configured AI will acknowledge the gap and offer to connect the customer with a human rather than guessing. This is the correct behaviour — it protects your customer experience and flags which topics need new articles. Review these escalations weekly to continuously expand coverage.
Is it worth hiring someone to build my knowledge base?
For most SMEs, no. The business owner or manager knows the answers to customer questions better than any external writer. The value of doing it yourself is that the content matches your actual policies, products, and tone. The 2–3 hours of initial effort pays dividends for months.
Sources: Intercom/Breathe case study (2025), Gorgias AI Agent documentation and blog (2025), Zendesk help center optimisation guides, Dr. Bronner's/Gorgias case study, Baymard Institute checkout research (2024).