A knowledge base used to have one audience: humans reading answers. Now it has two, and they consume knowledge differently. There's the end user self-serving — reading an answer, scanning a doc, deciding what to do. And there's the AI agent — retrieving passages to ground a response, or calling your knowledge base as a tool mid-task. Designing for only the human leaves the agent under-served; designing only for the agent produces content humans find robotic. The good news: with a few deliberate choices, the same content serves both well. Here's how.
How the two audiences differ
End users (self-service) consume knowledge as presentation. They want a readable answer, scannable structure, and enough surrounding context to feel confident. They tolerate — even prefer — some prose, examples, and reassurance. They navigate, skim, and judge. Their failure mode is "I couldn't find it" or "the answer was unclear."
AI agents (agent-assisted) consume knowledge as retrieval units. An agent doesn't read your beautifully formatted page top to bottom — it retrieves the most relevant passage and grounds an answer on it. It needs each passage to be self-contained (because it's retrieved out of context), factually clean (because it'll be quoted or paraphrased), and accompanied by a source (because a good agent cites). Its failure mode is retrieving a fragment that's ambiguous alone, or a passage that mixes fact with fluff the model then repeats.
The subtle tension: humans like context and warmth; agents like clean, self-contained, factual chunks. But these aren't actually opposed — they converge on the same underlying quality.
The content that serves both
The reassuring finding is that well-structured content is good for both audiences. The properties agents need turn out to be properties that also make content better for humans:
- Self-contained sections. Agents need passages that stand alone when retrieved; humans benefit from sections they can land on and understand without reading everything above. Both are served by writing sections that don't depend on distant context. This is the biggest shared win.
- Clear, descriptive headings. Agents use structure to chunk cleanly; humans use headings to scan. Question-shaped, task-oriented headings help both find the right part.
- Facts stated plainly. Agents ground more reliably on clear factual statements than on hedged, buried ones; humans get clearer answers too. Verbose, meandering content hurts both — it muddies the agent's chunk and bores the reader.
- Cited, verifiable claims. Agents should cite sources; humans want to verify. Content organized so claims trace to sources serves the agent's citation need and the human's trust need at once.
Write for clarity, self-containment, and structure, and you've largely satisfied both audiences with one body of content.
Where they genuinely diverge — handle it in delivery, not content
A few needs do differ, and the clean way to handle them is at the delivery layer, keeping the underlying content single-sourced:
- Presentation vs. raw passage. Humans want rendered, formatted, navigable pages; agents want raw retrievable text. Solve this by separating content from presentation — the same source content renders as a help article for humans and is available as retrievable passages for agents. Don't maintain two copies; maintain one source with two surfaces.
- Interaction model. Humans browse and search; agents call a tool. Expose the knowledge base both ways: a search UI / help center for people, and a tool interface (an API, or an MCP server) for agents. Kognita does exactly this — one ingested knowledge base, surfaced as human-facing search and as an MCP tool an agent can call — so both audiences hit the same content through the interface each prefers.
- Verbosity expectations. Humans sometimes want more hand-holding; agents want the core fact. A good retrieval + generation layer handles this dynamically — the agent retrieves the clean passage and can elaborate as needed — rather than requiring you to write two versions.
The design principle
Don't design two knowledge bases. Design one well-structured source of truth, delivered through two surfaces. Write content that's clear, self-contained, factual, and cited — which serves humans and agents alike — and then expose it both as a human-navigable search experience and as an agent-callable tool. The mistake is thinking self-service and agent-assisted need different content; they mostly need different interfaces to the same high-quality content.
This also future-proofs you. The agent-assisted audience is growing fast — more and more knowledge gets consumed by an AI acting on a user's behalf rather than by the user directly. Content structured well enough for agents to retrieve and cite cleanly is content ready for that shift, and it happens to be better for your human readers today. Build for the passage and the page, from one source, and you serve everyone who'll ever ask your knowledge base a question — human or not.