A knowledge base for WhatsApp customer support: what to include and how to maintain it

A knowledge base is the set of business-approved information an AI agent draws its answers from: policies, products, timelines, reference prices and rules for what to do in each case. If the agent cannot find something there, it has two options: say it does not know and hand the conversation to a person, or improvise. What you write, and what you leave out, decides which one happens.

That is why the knowledge base matters more than the model. Two stores can run the same AI and get opposite results: the one with shipping, returns and warranty rules written down clearly gets correct answers; the one that kept everything «in the team's heads» gets generic or invented ones. AI does not fix missing information. It exposes it.

It should include what customers actually ask, not what you wish they asked. The most useful starting point is your own history: open your last few weeks of WhatsApp conversations and note the questions that repeat. They almost always fall into five groups: products (materials, sizes, compatibility, availability), shipping (timelines, costs, zones, tracking), payments (accepted methods, installments, invoices), returns and exchanges (windows, conditions, who pays for shipping) and business details (hours, addresses, contact channels).

Add the rules that are judgment rather than facts: when you offer a discount and when you do not, what happens if an order arrives damaged, which exceptions you accept on a return. Those decisions are the ones that eat your team's time when they get asked again and again, and the ones that benefit most from being written down.

To see how this content is organized inside the product, heysteff's knowledge base lets you load and maintain this information so Steff uses it when replying.

Each entry should answer one question, with the answer in the first sentence and the conditions after it. «Exchanges are accepted within 30 days with the receipt and an unused product» is more useful than a legal-style paragraph where the number sits in line five. If an answer has exceptions, separate them clearly: general rule first, exceptions after.

Write the question the way a customer would ask it, not the way your team names it. Customers say «how long until it arrives?» and «can I swap it if it doesn't fit?», not «dispatch lead time» or «reverse logistics policy». Including two or three phrasings of the same question helps retrieval find the right entry even when the customer uses different words.

A common mistake is mixing several topics in one document. A file called «General information» holding shipping, warranties and hours together makes the right answer harder to find and a contradiction easier to create. Splitting by topic is tidier for you and more precise for the agent.

Finally, write in the tone your brand would use with a customer. Short, warm and specific answers read better on WhatsApp than formal text copied from a website, and they set the standard the agent will follow when it phrases its own replies.

The main cause of wrong answers is not that the AI «hallucinates» but that the knowledge base holds two versions of the same rule. If the old entry says 15 days for exchanges and the new one says 30, the agent can quote either. The practical rule is that each fact lives in one place and, when it changes, you replace the entry instead of adding another next to it.

Assign an owner per area: someone in logistics for shipping, someone in finance for payments, someone in product for product details. They do not need to write; they only need to confirm the content is still valid. Schedule a short review whenever you launch a campaign, switch a carrier or change a policy, because those are the moments the knowledge base goes stale without anyone noticing.

Promotions work the same way: an offer with an end date should have that date written in and should be removed when it expires. A forgotten promotion in the knowledge base is a promise your business will have to honor or explain.

Whatever the knowledge base does not cover should end with a person, not with improvisation. Decide in advance which topics the agent never resolves alone (legal complaints, health cases, refunds outside policy, upset customers) and set up escalation so the conversation reaches your team with context already summarized. We cover it in our piece on AI to human handoff on WhatsApp without making customers repeat everything.

If you are just starting, copilot mode is a careful way to test the knowledge base: the AI drafts the reply from your information and a person approves it before it is sent. Every correction you make shows where content is missing or excessive. You can see how it works in copilot mode, and once corrections get rare, move the simple topics to automatic mode.

You improve the knowledge base by reading the conversations the agent could not resolve. Every handoff to a person is a signal: either information was missing, it was badly written, or the case truly needed human judgment. A weekly review of those conversations, asking one simple question («which entry would have prevented this?»), improves the knowledge base more than any mass rewrite.

Looking at the most repeated topics helps here. If a question shows up twenty times and is always escalated, it deserves its own entry; if an entry is never used, it is probably redundant or poorly written. We explain how to run that analysis in how to analyze customer conversations on WhatsApp, and a related read on why first response time matters shows why filling these gaps pays off.

A concrete example: a clothing store notices that every week several people ask whether they can exchange an item bought on sale. The policy exists, but it is buried inside a long «terms» document. The team creates a dedicated entry with the direct answer and the conditions, and handoffs on that topic drop without touching anything else.

How many questions should a WhatsApp knowledge base have? There is no right number. Start with the 20 or 30 questions that repeat most in your conversations and expand based on the handoffs you see. A short, precise knowledge base beats a huge, contradictory one.

Can I use my existing documents, like PDFs or policy pages? You can start from them, but it pays to rewrite the important parts as short entries, one per topic, with the answer up front. Long documents written for legal reading tend to bury the fact the customer needs.

What happens if a customer asks something that is not in the knowledge base? A well-configured agent should recognize it lacks that information and hand the conversation to a person instead of inventing an answer. Defining that behavior is part of building the knowledge base, not an extra.

• heysteff knowledge base
• Copilot mode
• AI customer support
• WhatsApp AI agent