The team did not need a better help centre. It needed customers to get the answer from the help centre without having to search it - and to be told plainly when the help centre did not have one.
The starting position
The company sells scheduling and dispatch software to service businesses. Its help centre was good - 340 articles, kept current by the support team with each release - and under-used. Customers emailed instead, and about 610 tickets a month were questions a help-centre article already answered: how to set up recurring jobs, why a technician could not see a booking, how to export a timesheet.
An earlier experiment with a scripted chatbot had been switched off after a quarter. It handled the dozen intents someone had written and replied "Sorry, I didn't get that" to everything else.
What they set up
One SupaCortex project on the Growth plan, set up by one senior agent over two afternoons.
- A crawler feed on the help centre, using the TheDocs preset, re-read every six hours so a changed article is answered from within hours of publishing.
- Help-centre articles marked trust tier 1 (official); a folder of older release notes uploaded as tier 2.
- The chat widget in public mode on the help centre and in the Help panel of the web app, with both sites listed as allowed origins.
- A greeting and four quick actions for the most common topics, in the product's colour.
- Thumbs up and down on every answer, with the citations shown beneath it.
What went wrong first
During the trial the agent had also uploaded a folder of agent-only troubleshooting notes into the same project, to see how it handled them. It handled them well - which was the problem: in testing, the widget answered a question about a known sync bug with the internal workaround and its ticket reference. The notes were moved to a separate internal project, used by the support team through workspace chat, before the widget went public.
The second problem was contradictory content. Three release notes described the old behaviour of the timesheet export, and in the first week the assistant cited one of them over the current article. The help-centre article already outranked them as tier 1, but on one phrasing the release note matched the question's words far better. The team deleted release notes older than a year.
Running it
In week one, 68% of widget questions were answered with a citation; the rest got an honest "I couldn't find that" rather than a guess, and the help centre's own contact page was a click away. The project's unanswered questions, grouped by wording, became the support team's writing list: 38 new or rewritten articles over ten weeks, most of them short.
By week ten, 81% of questions were answered with a citation. Thumbs-down answers settled at about one in twenty rated answers; almost all of them traced back to an article that was ambiguous about which plan a feature belonged to.
The result
How-to tickets fell from about 610 a month to about 330 - a 46% fall - while billing and bug tickets, which the assistant does not cover, stayed roughly flat over the same period. No agents were added, and the team's first-response times improved because the remaining queue was harder questions rather than more of them.
What did not move: tickets where the customer needed someone to look at their own account. The team's next step is identified mode in the web app, so the assistant knows which account is asking and can look up a booking through a connector - with a confirmation before it changes anything.
ChatGPT
Claude
Perplexity
Grok