01 the business
Online course platform
35 staff, 80,000 learners
The help centre was written at launch and has drifted out of date. The best answers live in closed tickets that only one or two agents remember. Learners keep asking the same questions.
02 before
The problem
Your best answers are buried in closed tickets that only one agent remembers. The help center was written at launch and has drifted out of date. Nobody has time to write new articles, so customers keep asking the same questions.
- Dozens of common questions have no article.
- Some articles contradict what agents now tell customers.
- Nobody has time to write.
03 the build
What the AI does
- Step 1: Groups recently solved tickets in Zendesk, Intercom or Freshdesk by the question being asked.
- Step 2: Checks your existing help center to find questions with no article, or with an article that gives a different answer from your team.
- Step 3: Drafts a new or updated article from the replies that solved the problem, in plain steps and your house style.
- Step 4: Strips out customer names, order numbers and other personal details from the examples.
- Step 5: Sends each draft to the article owner in your help-center tool or Notion, with links to the tickets it came from.
paused: waiting for a person to approve
How it was put together
- Grouped three months of solved tickets by the question being asked.
- Compared the groups with the help centre to find gaps and contradictions.
- Drafted articles from the replies that actually solved the problem.
- Removed personal details and linked each draft to its source tickets.
04 human in the loop
Where a person stays in charge
A named editor on your support team reviews every draft for accuracy and tone. Nothing is published to the help center until they approve it.
- A named editor approves every article before it is published.
- Personal data is stripped from every example.
- Each draft shows the tickets it came from, so the editor can check.
05 rollout
From kickoff to live in 2 weeks
- Week 1Connect the help desk and help centre; agree house style.
- Week 2Produce the first gap report and ten draft articles.
- MonthlyRun the gap report and draft updates every month.
06 after
What changes
The help center keeps up with what customers really ask. Agents link to an article instead of rewriting the same answer, and knowledge stays with the business when people move on.
Assume 20 common questions with no article, each asked 40 times a month at 5 minutes per reply, and articles that deflect a quarter of them.
20 × 40 = 800 tickets a month. 800 × 25% = 200 fewer tickets. 200 × 5 minutes = about 17 hours of agent time a month.
What to measure
- Tickets on topics that now have an article.
- Article views before contacting support.
- Articles updated per month.
- Editor time per article.
07 evidence
What the research says
Across 5,179 customer support agents, an AI assistant raised issues resolved per hour by 14% on average and by 34% for novice and lower-skilled agents.
NBER, Generative AI at Work (Brynjolfsson, Li and Raymond) (opens in a new tab) · 2023
Knowledge workers spend roughly a fifth of their working week searching for and gathering information.
McKinsey Global Institute, Capturing business value with social technologies (opens in a new tab) · 2012
08 lessons
Mistakes to avoid
- Publishing without an editor.
- Writing for the agent, not the customer. Use plain steps.
- Forgetting updates. Old articles cause tickets too.
09 questions
Common questions
- Which help centres does it support?
- Zendesk Guide, Intercom Articles, Freshdesk, Help Scout, Notion and Confluence.
- Will it change existing articles directly?
- No. It proposes changes with a clear list of what changed, for the editor to accept.