Plan requirement
| Subscription | Suite Team or higher, Support Team or higher |
| Access | Admin |
Letting the agent answer from your articles instead of fixed text. The setting is small; the preparation is not.
Turn them on
- In the AI agents workspace, select the agent, click Content in the sidebar and select Knowledge sources.
- Click Edit sources. In the field at the top, choose Help center and select the help centre you want the agent to read, or choose External content and select a source that a crawler or connector has brought in.
- To let it read only part of that, go to Content > Search rules and click Create search rule. A rule filters on Article, Locale or Labels, with operators such as Equals, Contains and Includes none.
- Test with real questions, then read the first fifty transcripts.
Two settings sit under Content > Advanced settings, both on by default: Send a waiting message while generating knowledge answers, whose wording follows the agent's persona and cannot be dictated, and Display sources for generative replies, which shows the customer which articles the answer came from. Turn the second off and you lose the cheapest way for a customer to check the answer themselves.
Scope the knowledge deliberately
A whole help centre is the easy choice, and it is rarely what you want. Internal notes, outdated articles and anything ambiguous should be outside the agent's reach, and search rules on labels are the practical way to do it: label the articles the agent may use, then rule on that label.
One caveat: label-based rules do not work for Confluence sources.
Narrowing the source is the most effective quality control available here, and it takes minutes.
What makes an article usable
The answer stated plainly, near the top, in a section whose heading matches the question. Articles that build to a conclusion over five paragraphs get summarised badly, because the summary is of the whole thing.
Where it goes wrong
- Two articles disagreeing. The agent has no way to know which is current, so it may use either.
- Articles that assume context a reader has and the agent does not.
- Long combined articles, where the relevant paragraph is one of twenty.
Fixed text still wins where wording matters
Prices, terms and guarantees stay as written replies. Generative answers are accurate in substance and variable in phrasing, and some sentences need to be the same every time.
Read transcripts, not settings
The only way to know whether this is working. Fifty conversations tell you which articles are being used, which are being misread, and which question has no article at all.
Comments
0 comments
Article is closed for comments.