A whole industry has grown up around prompts — courses, gurus, “prompt engineering” on résumés. And I will say something this industry will not like: in enterprise AI, the knowledge base matters more than the prompt. Much more.
The prompt is how you asked. The knowledge base is what the system knows about your company. You can phrase the question perfectly, but if the agent does not know your regulations, routes, templates and decision history — it will answer beautifully and “in general.” In personal correspondence that is harmless. In a business process, an outdated price in a commercial proposal or a cancelled regulation in an agent’s answer is no longer a style issue — it is a risk with consequences.
Last summer MIT published a study I have written about: 95 percent of enterprise AI pilots produced no measurable effect. Note that the cause is almost never the model — everyone’s models are roughly the same. The cause is data and how it is governed: the agent simply has nothing to build an accurate answer from.
Let me tell you how we went through this ourselves. When we started rebuilding Documentolog around AI, the first infrastructure project was not a model and not an agent — it was the corporate knowledge base. Boring, long, unspectacular — no demo at the end of the week. But it is what turns an agent from “answers roughly right” into “answers by your rules.” Today all our digital employees run on it, and it is the core of the methodology we use with clients.
What a working knowledge base looks like in practice — no magic, just items you can check in your own company. Every document has an owner and metadata. Policies, roles, prices and regulations are current — not “somewhere on a shared drive” but with a review date. Documents are linked to each other and versions are controlled. There is an update process after every change: the regulation changed — the base learned about it the same day, not a quarter later.
And the most important part, which we did not arrive at immediately: freshness must be someone’s job, not “everyone’s a little.” For us it is a dedicated system agent — the Knowledge Steward. It checks knowledge for freshness and consistency, tracks decay, and prevents the other agents from citing anything that has been cancelled. A company’s knowledge is a living organism: without this control the base goes stale within months, and the agents start confidently answering with yesterday’s truth. This, incidentally, is the answer to the hallucination question I get at every meeting: hallucinations are cured not by a perfect prompt but by a controlled source of truth the agent is required to check against.
There is a simple test of whether your company is ready for agents. Ask yourself: if your key regulation changes tomorrow, how many days until every place where it is recorded “knows” about it? If the answer is “no idea” or “a quarter,” you do not need another AI tool yet — it will eloquently recite outdated rules.
The prompt is technique. The knowledge base is the foundation. And a house, as everyone knows, is built starting from the foundation.
