Before you spend time evaluating AI support vendors, improve your knowledge base. This is not romantic advice. It is engineering reality: an AI agent is a retrieval-and-reasoning system, and the quality of the retrieval is bounded by the quality of what it retrieves. We have seen teams deploy sophisticated AI tooling on top of documentation that was two years out of date, then blame the AI when it gave wrong answers. The culprit was the content, not the model.
What "Hygiene" Actually Means in a Knowledge Base Context
Knowledge base hygiene has four dimensions, and most teams only think about one of them:
Coverage
Does every major ticket category have a corresponding article? This is the most obvious gap, and it is almost universally present. In our experience auditing knowledge bases before deployment, teams typically find that 20-30% of their top inbound ticket categories have no dedicated documentation. Customers are asking questions that have no written answer anywhere — and they are routing to humans because the documentation gap left no other option.
Run this audit: export your last 60 days of Zendesk or Intercom tickets, pull the top 20 categories by volume, and map each one to a knowledge base article. Any category without a clear mapping is a coverage gap that needs to be filled before you deploy AI assistance.
Accuracy
Are the articles factually correct today, not as of when they were written? This is the dimension that degrades fastest. A SaaS product that ships every two weeks will have broken documentation within a month if no one is maintaining it. We once found a help center article from 2023 describing a UI flow that had been redesigned twice since then — the article was still the top search result for its topic, and agents were citing it confidently. Every wrong answer the AI gives on a documented topic traces back to a stale article.
Findability
When a customer asks a question, can the AI locate the relevant article? This depends on how the article is titled and structured. An article titled "Integration Settings" is harder to match against the query "how do I connect Slack" than an article titled "Connecting Slack to Your Workspace." Titles should use the language customers actually use, not the language your product team uses internally.
Completeness
Does each article answer the question fully, or does it leave edge cases unaddressed? Partial answers are one of the most common causes of re-submitted tickets. The customer reads the article, follows the steps, hits a snag on an edge case the article did not cover, and submits another ticket. The fix is adding an "If this doesn't work" section to every procedural article, covering the 3-4 most common failure points.
The 60-Day Hygiene Audit
Here is the practical process for a team of 3-5 people with a moderate-sized knowledge base (50-200 articles):
- Pull your top 30 ticket categories by volume. These are the articles you need to get right first. Everything else is secondary.
- Assign each category to a team member who owns it. Not "the support team" — a specific person. Ownership is what drives completion.
- Review each article against the current product. Open the product in one window, the help article in another, and walk through every step. Where the product has changed, update the article.
- Add articles for categories with no coverage. Target 200-400 words per article: overview of the topic, step-by-step resolution, common failure points, and what to do if the steps do not resolve it.
- Standardize article titles. Every procedural article should start with a verb: "How to...", "Connecting...", "Resetting...". This improves matching accuracy significantly.
This process takes 2-4 hours of dedicated work for a team covering 30 ticket categories. It should be done before AI deployment, and it should be repeated quarterly — or whenever a major product change ships.
The Maintenance System That Prevents Drift
A one-time audit is not enough. Knowledge bases drift as products ship. The teams that maintain good AI performance over time have a maintenance system, not a one-time cleanup event.
Changelog-triggered reviews
Every time a product feature changes, flag the corresponding help articles for review. In practice, this means adding a step to your engineering release process: when a PR ships a user-facing change, the PR author notes which help articles need updating, and that review appears in the support team's queue within 24 hours. This is the single most effective maintenance habit we have seen.
Re-submission rate as a signal
Monitor re-submission rate weekly — tickets that are marked resolved (by AI or human) and then re-opened within 48 hours. Re-submissions cluster around articles that are incomplete or inaccurate. A spike in re-submissions for a specific category is almost always a signal that the corresponding article needs attention.
Monthly coverage drift review
Each month, compare your top inbound ticket categories against your article index. New ticket categories appear constantly as your product evolves and your customer base grows. A category that generates 30+ tickets per month with no corresponding article is both a support load and a documentation gap.
"We spent two months frustrated by AI answers that were technically coherent but factually wrong. Every single case traced back to a stale article. The fix was not in the model — it was in the documentation."
What Good Coverage Actually Looks Like
For reference: a knowledge base with good coverage for an AI support deployment has the following characteristics:
- Top 20 ticket categories each have at least one dedicated article
- Articles reviewed and verified against current product within the last 60 days
- Procedural articles include both the successful path and common failure resolutions
- Article titles use customer vocabulary, not internal product terminology
- A review was triggered for any article touching a feature that shipped in the last quarter
Teams that reach this state before deployment consistently see higher deflection rates and better CSAT on agent-resolved tickets than teams that deploy first and fix documentation later. The order matters. Knowledge base first, AI second. Not the reverse.
Getting Started This Week
If your team has not done a knowledge base audit recently, start with three actions this week:
- Pull your top 10 ticket categories from the last 30 days.
- Open each corresponding article (or confirm one does not exist).
- Check each article against the current product — note anything that is outdated or missing.
That audit takes under two hours and gives you a clear priority list for the documentation work that will have the most impact on your AI deployment quality. Do this before anything else.
