How is this different from a database lookup?

A lookup returns stored answers; a knowledge-based agent derives new answers by applying rules to facts.

Is this what chatbots do with company documents?

It is a relative: retrieval plus a language model plays a similar role, but without the formal guarantee of explicit inference.

What is the main running cost?

Knowledge maintenance: keeping facts and rules current, owned and tested.

AI Glossary / Agents and automation / Intermediate

Knowledge-based agents

Definition

A knowledge-based agent is an agent that keeps an explicit store of facts and rules about its world, called a knowledge base, and reasons over that store to decide what to do, rather than reacting only to the current input.

What it means

Most agents react: they see input, they produce output. A knowledge-based agent also remembers and deduces. It is told facts, such as which customers are on which plan, and rules, such as a customer on the premium plan gets priority routing, and it derives what follows for the case in front of it.

The power of the design is that knowledge is stated, not buried in code. A new rule can be added in something close to plain language, and the agent's behavior changes without rewriting its machinery. The knowledge base becomes a readable, auditable description of what the system believes.

The design also has a known failure shape. If the knowledge base contains a wrong fact or a missing rule, the agent reasons correctly from wrong premises and reaches a confident, wrong answer. Maintaining the knowledge base is the real work.

Modern language-model systems borrow the pattern: retrieval over curated documents plays the role of the knowledge base, and the model plays the role of the reasoner. The old lesson still applies: what the system is allowed to believe decides what it can sensibly say.

For a business, knowledge-based thinking turns into a discipline: write down the facts and rules the system may rely on, give them owners, and review them like financial records.

How it works

The classic architecture has two parts. The knowledge base stores sentences in a formal language: facts about the world and rules connecting them. The inference engine derives new facts from old ones by applying the rules, answering questions or choosing actions.

The agent cycle then reads naturally: tell the knowledge base what was perceived, ask it what to do, act, and record the result. Because each step is explicit, the chain of reasoning behind a decision can be inspected, which is why the design has long been used where justification matters.

Practical systems bound the reasoning. Unlimited inference is slow, so real systems limit depth, cache conclusions or restrict the rule language, trading completeness for speed.

Modern retrieval-augmented generation echoes this: documents are retrieved into the model's context and the model reasons over them. The retrieval quality plays the role the rule quality once did.

Three practical examples

Policy advice: a system stores product terms and eligibility rules, then answers whether a given customer qualifies, showing the rules it used.
Diagnostics: a support tool stores symptoms and known fixes, and reasons from a reported symptom set to the most likely cause to check first.
Configuration: a sales tool stores compatibility rules between components and rejects combinations that violate them before a quote goes out.

Think of it this way

Think of a meticulous company handbook plus a careful administrator. The handbook holds the facts and rules; the administrator reads the situation, finds the relevant pages and applies them, noting exactly which pages justified the decision.

The limit: the administrator can notice the handbook is out of date and ask about it. The agent cannot; someone must own the handbook, or the system will keep applying last year's rules with perfect logic.

Worked example

answering eligibility questions from stated rules

Suppose a grants body gets thousands of questions about whether applicants qualify. The criteria are published but tangled, and staff answer inconsistently.

A knowledge-based approach encodes the criteria as explicit rules and facts: turnover thresholds, registration requirements, sector exclusions. An applicant's details are entered, and the system derives eligible, ineligible or needs-evidence, showing the exact rules behind each answer.

When criteria change, the rule owner edits the knowledge base and the new logic is tested against last year's decided cases. The teaching scenario is invented; the auditable chain from answer to rule is the point.

Consistency is the payoff: every applicant is judged by the same stated rules, and every answer carries its justification.

Illustrative case study

A fictional bank builds a knowledge-based adviser for account fees. It answers confidently and cites its rules, which impresses everyone until a fee-change month: half the knowledge base now describes the old schedule, and the system reasons flawlessly to wrong answers.

The fix is process, not technology: every knowledge entry gets an owner and a review date, and the fee schedule gets a single source that the knowledge base must reference rather than copy.

The invented lesson: knowledge-based systems are only as current as their maintenance. Reasoning quality cannot rescue stale facts.

Common misconceptions and mistakes

The knowledge base maintains itself. Facts and rules decay; each needs an owner and a review date.
Explicit rules make the system inflexible. The rules are easier to change than buried code; the discipline is changing them deliberately.
Retrieval from documents is the same as knowledge. Retrieved text is evidence, not verified facts; someone must still vouch for what goes in.

Business use

Use knowledge-based designs where answers must be justified: eligibility, compliance, configuration, diagnostics and policy advice.

Start by inventorying the rules your experts actually apply, assign each an owner, and test the encoded version against historical decided cases before letting it advise.

Check your understanding

A rules-based fee adviser gives correct reasoning from an outdated fee schedule. Where exactly did the system fail, and what process prevents a repeat?

Answer: The knowledge base was stale, not the reasoning; assign every knowledge entry an owner and review date, and reference single sources rather than copying values.

Related terms

Learn next

AI for Business Owners covers putting company knowledge to work safely in AI workflows: https://aiforbusinessandlife.com/academy/ai-for-business-owners. It is an adoption course, not a logic-programming course. [C2]

Sources

[S1] Russell and Norvig, AIMA glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Knowledge base Next: Knowledge-based configuration