The Knowledge Base Catch-22

Most AI support tools need your help center to work. Intercom's Fin searches your docs. Zendesk's AI Agent references your knowledge base articles. If your docs are thin, inaccurate, or non-existent, these tools give bad answers.

But here's the irony: the teams that need AI support the most — small startups, solo founders, early-stage companies — are the same teams that don't have a 200-article knowledge base. They haven't had time to write one. They're too busy building the product.

So there's a gap. The teams that need help with support are locked out of the tools that could help because they haven't written enough help articles yet.

How Classification Skips the Knowledge Base

Intent classification doesn't reference your docs. It doesn't need them. Here's why.

The classifier is pre-trained on a taxonomy of 315 support intents. When a customer writes "I was charged twice," the model doesn't search your help center for an article about double charges. It recognizes the intent: billing_dispute.

What happens next is up to you. You set a routing rule: "When intent is billing_dispute, do X." X can be:

• Send a template response ("We'll investigate your charge. Here's a case number: [auto-generated]")
• Notify your Slack channel
• Create a ticket in Linear or GitHub
• Escalate with priority flagging
• All of the above, in sequence

You write the response template. You control the action. The AI's job is just to understand what the customer wants — not to generate an answer.

What You Need Instead of a Knowledge Base

If you don't have docs, here's what you do need:

5 to 10 response templates. Short, direct answers for your most common questions. You can write these in 30 minutes by looking at your recent support emails and writing a good answer for each repeated question.

Routing rules for each template. "When the customer asks about pricing, send this response." "When they report a bug, create a GitHub issue."

A catch-all rule. "For everything else, notify Slack." This is your safety net for anything the automation doesn't cover.

That's it. No 200-article knowledge base. No hours of content writing. No ongoing maintenance of help center docs.

But Should You Still Write Docs?

Yes, eventually. Good documentation reduces support volume over time because customers find answers themselves. But you don't need docs to start automating support. You can set up classification-based automation today and write your knowledge base gradually, as you learn what questions come up most often.

Start with classification and templates to handle the immediate pain. Build out docs as you go. Each doc you write reduces your support volume by a tiny amount. Over 6 months, those tiny amounts compound.

The Setup Flow

Here's the 15-minute version:

Minutes 1 to 3: Sign up, get your API key, install the widget on your site.

Minutes 3 to 8: Write 5 response templates for your top 5 question types. Keep them short — 2 to 4 sentences each. Include a link to your contact email for follow-up if the template doesn't fully answer the question.

Minutes 8 to 12: Create routing rules. Match each template to the right intent. Set confidence thresholds at 80 to 85%.

Minutes 12 to 15: Connect Slack for catch-all notifications. Test with 3 to 5 sample messages.

Done. You now have AI-powered support that works without a knowledge base, handles your most common questions, and notifies you about everything else.

Comparing the Two Approaches

Knowledge-base AI (Fin, Zendesk AI):

• Setup time: 1 to 4 weeks (writing docs + configuring AI)
• Ongoing cost: $0.90 to $1.50/resolution + time maintaining docs
• Accuracy: depends on doc quality (40 to 90%)
• Works without docs: no

Classification-based AI:

• Setup time: 15 minutes
• Ongoing cost: $0.20 to $0.30/resolution
• Accuracy: 92% intent classification out of the box
• Works without docs: yes

The knowledge-base approach is better when it works. But "when it works" requires significant upfront investment. Classification gets you running today.

When to Graduate to Docs-Based AI

Consider adding a knowledge base and an LLM-based support tool when:

• You have 500+ support interactions/month
• Your questions are becoming more varied and don't fit neat templates
• You have the time or staff to write and maintain 50+ help articles
• Customers frequently ask multi-step "how do I" questions that need detailed explanations

Until then, classification with templates gets the job done.