Knowledge base automation from solved tickets

In March, support answers the same question about exporting invoices 140 times. Forty agents type their own version of the answer, and three of the versions are wrong. The help center has no article on it, because nobody owns noticing which questions keep coming back. In April, the AI assistant on the help center can’t answer it either, because it can only use what the help center says.

The help center is where customer support gets to answer a question once, and it only covers what someone has written down.

What is knowledge base automation?

Knowledge base automation finds solved tickets that no help article answered, and groups them by the question the customer was asking. It drafts an article from the replies that solved it, and publishes only after a named owner reviews the draft. Then it watches whether tickets on that topic drop.

Why the knowledge base falls behind

1. Nobody sees the repeats. Each agent answers the question in front of them, and no report shows that the same question came in 140 times.
2. Customers ask the same thing many ways. Grouping by keyword misses most of the repeats.
3. Writing articles is extra work. It competes with the queue, and the queue wins.
4. Drafts get published without review. A draft built from agents’ replies can carry an old workaround or a detail about one customer’s account.
5. Nobody checks whether the article worked. Without measuring repeat tickets afterwards, the team can’t tell a good article from an ignored one.

What it costs

Every repeat question is agent time spent on something the help center could have answered. It also limits the AI assistant. An assistant that answers only from the help center can’t answer what the help center doesn’t cover, so every missing article is a ticket the assistant has to hand over.

What to fix first

1. Find solved tickets with no linked article. Each one is a vote for an article that doesn’t exist yet.
2. Group by the question, not the words. Cluster tickets by what the customer was trying to do.
3. Draft from the replies that solved it. Use what worked, not the model’s own idea of your product.
4. Never publish without review. A named owner who knows the product checks every draft.
5. Measure repeat tickets after publishing. It’s the only proof the article worked.

The guide to building ticket to help article drafting has the prompts for each step.

Where it fits

New articles are what AI knowledge agents answer from, so this is how the assistant at the front of support gets better. Bugs that keep generating the same question should also go through the escalation process, because an article explains the problem while engineering fixes it.

For the whole customer support process, from first message to fix, see the customer support process.