When teams connect their help centre to Tagni and the answers come back thin, the cause is almost never the model. It is that the documentation was written for someone browsing, not for someone asking.
The good news is that the changes which make docs work for an agent are the same ones that make them work for a human in a hurry.
Put the answer in the first sentence
A page titled "Managing your subscription" that opens with two paragraphs of context has buried the thing everyone came for. Lead with the answer, then explain.
Compare:
Subscriptions in Tagni are designed to be flexible, reflecting the way modern teams work...
against:
You can change plan at any time from Settings, then Billing. Upgrades take effect immediately and are prorated. Downgrades start at your next renewal.
The second version can be quoted directly to a customer. The first cannot.
One page, one question
Long pages that cover six loosely related topics retrieve badly, because the chunk that matches the query arrives without the context that made it make sense. Split by the question a person would actually ask.
- Good: "Why was I charged twice?"
- Bad: "Billing"
Write out the specifics
Agents cannot infer what you left implicit. If a limit is 5 seats, say 5 seats. If a refund window is 60 days, say 60 days. Phrases like "a reasonable period" or "depending on your plan" are unanswerable, and an honest agent will refuse to guess rather than invent a number.
The corollary: every time you leave something vague in your documentation, you are guaranteeing a human has to handle that ticket forever.
Keep the edge cases on the page
The single highest value paragraph in most help articles is the one nobody writes: what happens when the normal path does not apply. What if the card was already refunded. What if the account is on a legacy plan. What if the customer is in the EU.
These are exactly the questions that reach support, and they are usually documented only inside somebody's head or a Slack thread from March.
Date everything
Add a visible last reviewed date to every page. Retrieval systems can then prefer current material over stale, and your team gets an obvious signal about what needs attention. A page nobody has reviewed in two years is a liability whether or not an AI is reading it.
A short checklist
- Answer in the first sentence
- One question per page
- Concrete numbers, never ranges you have not defined
- Edge cases written down rather than remembered
- A review date on every page
- Tables for anything with more than three variables
Do these six things and the quality of automated answers improves immediately, with no change to the model at all. It is, reliably, the highest leverage work available to a support team.