Skip to content
Fritz Gerald ZephirinMenu

Social

The Knowledge Base Is the Product

4 min read

Every few weeks someone asks me which AI support vendor to buy, and they are always surprised by the answer: it matters far less than they think, because they are asking about the renderer when their problem is the source code.

Here is the architecture nobody draws on the whiteboard. An AI support system is two things: a model, which is a commodity you rent, and a knowledge base, which is the company’s written account of what is true — policies, procedures, prices, exceptions, the answer to every question you’ve ever been asked. The model renders the KB at the customer. That’s the whole system. Which means the quality ceiling of your AI support was set years before you bought the bot, by every help article nobody owned, every policy that lives in a senior agent’s head, every page that says the old refund window because nobody deprecates anything.

Before AI, a weak knowledge base was survivable, because the humans reading it were a correction layer — an experienced agent reads a stale article, knows it’s stale, and answers correctly anyway. The bot does the opposite. It renders the stale article fluently, confidently, at scale, around the clock. Pre-AI, bad documentation produced occasional wrong answers with an apologetic tone. Post-AI, it produces a thousand identical wrong answers delivered with perfect assurance — and confident garbage is worse than honest ignorance, because customers act on it. This is the unglamorous truth inside the MIT finding that most AI pilots return nothing: the “learning gap” between buying a tool and getting value is, in support, mostly a writing gap. The companies in the successful five percent fixed their truth layer first.

So treat the knowledge base the way engineering treats code, because it now executes:

Named ownership. Every article has an owner the way every service has an on-call. “The team maintains it” means nobody does.

Review and deprecation. Articles carry review dates and expiry behavior. Stale truth is the worst truth — an article that was right last quarter fails in a way no one audits, because it looks maintained. When policy changes, the old article is deprecated the same hour, not whenever someone remembers.

One source of authority. RAG over your raw tickets and wiki sprawl gives recall, but it cannot tell which of two contradictory answers is policy. Somewhere there must be a canonical layer — smaller, curated, versioned — that outranks everything else when sources disagree. The practice that works is hybrid: canonical KB for authority, wider corpus for coverage, and an explicit rule about which wins.

Evals built from real tickets. Before any model change ships, it answers a held-out set of your actual historical questions, scored against verified resolutions. This is your unit-test suite. Run the AI like a hire; this is its exam.

Failures file KB bugs. Every audited deflection failure, every recontact, every correction a human makes to a bot answer becomes a ticket against an article, with an owner and a due date. The weekly review of “top unanswered and worst-answered intents” is the most valuable meeting in a modern support org — it is the backlog grooming of your truth layer.

There’s a paradox here worth naming, because a thoughtful reader will spot it: if everything written down can be automated, isn’t writing everything down giving away the moat? The resolution is that the writable layer was never the moat — it’s table stakes, and executing it well is now mandatory. The moat is what the writing process reveals you can’t write down: the judgment calls, the client-specific tone, the escalation instincts. You document everything precisely to discover, and then concentrate humans on, the residue that won’t go in the document.

The compounding effect is the payoff. A ticket queue is pure operating expense — every answer consumed once. A knowledge base under this discipline is an asset that appreciates: every failure makes it more complete, every review makes it more current, and every model upgrade renders the same accumulated truth a little better, for free. Your competitors will buy the same bot you did. They cannot buy your KB. Write accordingly.

Sources & further reading