Your AI answers are only as useful as your knowledge base

Before rewriting your assistant's greeting, look at the information behind its answers. A polished sentence cannot resolve two conflicting opening times, an expired promotion, or a service description that never says who the service is for.

Treat your knowledge base as a working part of customer operations. It needs clear answers, explicit limits, and someone responsible for keeping it current.

Start with customer questions#

Collect a small sample of recent conversations. Remove personal information before turning examples into shared reference material. Group the questions by what the customer was trying to find out: whether you offer a service, how pricing works, what to bring, or what happens next.

Now write one focused article for each recurring question. A broad “Everything about our business” document is harder for a teammate to review and easier to fill with contradictions. A focused article makes the answer and its boundaries easier to inspect.

Liveline retrieves relevant passages from a workspace's knowledge base to help an assistant answer. Give those passages enough context to stand on their own: name the service and location instead of relying on “this,” “there,” or a heading several pages earlier.

Use a small, repeatable template#

The following is an illustrative internal article for a fictional repair shop. The example is a format to adapt, not a policy to copy unchanged.

Include conditions next to the answer they qualify. If a price applies to one location or excludes materials, keep that information in the same passage. A caveat at the bottom of an unrelated document is easy for a human editor to miss too.

Good source material answers two questions: “What can we say?” and “When does that answer stop applying?”

Resolve contradictions before adding volume#

If the website says you close at six and an imported PDF says seven, adding a third document will not settle the issue. Ask the owner of the policy which source is current, update the source, and remove or correct the obsolete version in the knowledge base.

Use a short review queue:

Uploading more information is easy. Deciding which information your business stands behind is the work that makes it useful.

Test meaning, not matching words#

For each important article, write a few questions that express the same need differently. “Can I drop by?”, “Do you take walk-ins?”, and “Do I have to book first?” should all lead to the appointment policy, while an urgent same-day request may need a person.

Also test a question the material cannot answer. You want an honest boundary and a next step, rather than an invented detail. If you serve customers in English and Spanish, test both languages with natural wording; a translated document is not itself evidence that every response will be correct.

Keep a small list of expected answers and rerun it after changing policies or importing a large document. Record whether the response was supported by the source and whether it respected the conditions.

Make maintenance someone's job#

Give each topic an owner. Review time-sensitive information when it changes, and reserve a recurring slot for unresolved questions from real conversations. Log the source of each policy so the next editor knows who can approve an update.

The goal is not the largest library. It is a dependable set of answers your team can explain, correct, and trust enough to use in a customer conversation.

When the approved material runs out, the next step should be clear. See our human handoff playbook.