Your AI Support Is Only as Good as Your Knowledge Base

When we started working on AI customer support for FeedLog, I thought the AI model would be the difficult part.

It wasn't.

The bigger problem was much more boring:

The knowledge base.

Most small SaaS products don't have perfect documentation

AI support demos usually assume something like this already exists:

Perfect documentation
        ↓
AI understands everything
        ↓
Great customer support

Real products look more like this:

README
+ outdated docs
+ random FAQ
+ information buried in the website
+ things only the founder knows

Then we give all of that to an AI and expect perfect answers.

Not surprisingly, that doesn't always work.

So who should build the knowledge base?

The obvious solution is to ask the user:

Upload your documentation.

Or:

Create your knowledge base.

Technically, that works.

From a product perspective, I'm not convinced.

Most founders don't wake up thinking:

"I'd love to spend today writing 40 FAQ entries for my AI support agent."

If using AI support creates another maintenance job, we've missed part of the point.

We're experimenting with a different approach

Instead of asking users to build everything manually, we're thinking about how much of this process can happen automatically.

For example:

1. Give FeedLog your website URL.
2. Scan the website, docs, FAQ and other public pages.
3. Generate an initial knowledge base.
4. Let the user review it instead of writing it from scratch.

Then something more interesting can happen later.

Imagine five customers ask the same question and the AI can't answer it.

Instead of silently failing five times, the system could say:

Five users asked about this. Your knowledge base doesn't have an answer yet.

Then generate a suggested knowledge entry.

The founder only needs to approve it.

Build less. Confirm more.

This has become one of the product principles we're exploring:

Don't make users build the knowledge base. Make them confirm it.

There's still a lot to figure out — especially around accuracy, outdated pages and deciding what information should actually become knowledge.

But I increasingly think the quality of AI customer support will depend just as much on how knowledge is created and maintained as on the AI model itself.

If you've built an AI product around a knowledge base, I'd love to know how you're handling this.