Most teams install an AI support tool expecting immediate results. The AI starts answering questions on day one, but the resolution rate is low — maybe 30%, maybe lower. They assume the AI isn't good enough and either give up or keep fiddling with prompts.
The problem usually isn't the AI. It's the content.
The resolution rate formula
AI resolution rate = (quality of your documentation) × (coverage of your documentation) × (AI configuration)
Most teams optimize for AI configuration (prompt tweaking, confidence thresholds) when they should be optimizing for documentation quality and coverage first.
Week 1: audit your questions
Before writing a single article, spend a week just reading support conversations. Look for:
- Questions that come up 3+ times - Questions where the answer is already in your docs but hard to find - Questions where there's no answer in your docs at all
Categorize them into three buckets: Already covered, Partially covered, Not covered.
Week 2: fill the gaps
Start with the "Not covered" bucket. These are the questions where your AI currently has no information and will either refuse to answer or hallucinate.
Write concise, factual articles for each. A good support article isn't a marketing page — it answers one question, directly, with examples.
Week 3: improve coverage
Now tackle the "Partially covered" bucket. These are often the hardest — there's something in your docs, but it's buried in a long guide or uses different terminology than the question.
Refactor these articles. Add direct answers at the top. Use the same language your customers use.
Week 4: measure and iterate
By week 4, you should be seeing 60-70%+ AI resolution on questions that have corresponding articles. The remaining 30% typically falls into:
- Truly novel questions (edge cases, bugs, new features) - Account-specific questions that require looking up data ("What's my current usage?") - Sensitive escalations that you want humans to handle regardless
That last category is fine — a well-configured AI agent that resolves 70% of tickets and correctly escalates the rest is genuinely useful.
The ongoing loop
Resolution rate isn't a one-time metric. Use the analytics dashboard to see what questions the AI is failing on each week, and treat content creation as an ongoing habit, not a one-time project.