When a customer asks a question, the AI answers from the knowledge sources you give it — starting with the desk’s own knowledge base — and cites what it used. Anything it can’t ground in those sources, it hands to your team instead of guessing.
Writing your knowledge base
Each desk has its own knowledge base, and you edit it inside the desk — nothing else to install and no other product required.
- Open it. In the desk’s inbox, click the Knowledge base icon — the book, next to the settings gear.
- Add an article. Click the blue + button, give it a title, write the body in Markdown, and click Create Document.
- That’s it. The AI can use it on the very next customer message. Edit or delete any article from the three-dots menu beside its title.
Write the way a customer asks, not the way your team talks internally: one article per question, a plain title like “Opening hours” or “How refunds work”, and the answer in the first paragraph.
Customers can ask in any language
The AI finds the right article by what the question means, not by matching its exact words. So a question asked in Spanish finds an article you wrote in English or Japanese, and a customer who asks “can I skip a delivery?” finds your article on subscriptions even if it never uses the word “skip”. The AI then replies in the customer’s own language, translating faithfully from your article and adding nothing it doesn’t say.
One article per question is still the shape that works best: ten short articles are easier to match, and easier to keep accurate, than one long article covering ten topics.
The source types
This desk’s knowledge base
On by default, and the one you write yourself. If you already keep documentation in an AB Projects project, you can point this at that project instead — Use another workspace’s wiki instead in Settings, blank means this desk’s own.
Public help content
Help-centre articles, like this one. Leave the channel id blank for the default help channel, or point it at your own. Good for how-tos and FAQs your customers already read.
Each source can be turned on or off, and repointed, per desk — you do it in Settings, under Knowledge sources. Because it’s per desk, a billing desk and a technical desk can answer from completely different material.
How answers stay trustworthy
The AI answers only from the sources you give it, and lists the articles it drew from beneath each reply. Those citations are links: click one and it opens the article it came from, so you can check the wording a customer was shown. If the sources don’t cover the question, the AI says so and routes the ticket to a person rather than inventing an answer — see how a ticket works.
Test it before a customer does
You don’t have to wait for a real question to find out whether an article works. Ask the AB Support bot in Microsoft Teams the same question a customer would — in their language, not yours, if they write to you in another one. It reads this desk’s sources and answers from them exactly as the customer-facing AI would, and shows you which article it used.
If the bot says it didn’t understand, that is the gap. Write the article, ask again, and you will see it answer.
Keeping them healthy
- Watch the escalation rate. A low Handled by AI % on the dashboard usually means thin sources, not a bad AI. The fix is another article, not a setting. AB Support’s Home page warns you about this on its own, once a desk has handled enough tickets for the number to mean something.
- Watch the helpfulness score. Lots of answers with a poor rating means the AI is confidently answering from content that’s wrong or unclear — fix the source, not the prompt.
- Let escalations tell you what to write. Tickets the AI couldn’t answer are a ready-made list of the articles your knowledge base is missing.
- Review after any change customers care about. New pricing, a changed policy, a discontinued product — update the source the day it changes, because the AI will keep quoting the old one until you do.