Practical guide · AI at work

Preparing and testing an AI knowledge base

A large folder of documents is not automatically a usable knowledge base. An assistant needs agreed sources, clear boundaries and examples for testing its behaviour. Start with one task rather than all company data.

1. Limit the task before gathering material

Describe the input and result. For example, a team member asks about a service, and the assistant finds information in approved material and prepares a draft answer. This is more specific than “AI should know everything about the company”.

Separately list what is outside the task: approving discounts, changing terms, sending a customer message or editing the CRM. Preparing an answer and taking an external action carry different responsibilities.

Task card: user …; question or input …; permitted result …; source …; prohibited actions …; when a person is needed … .

If an ordinary table or fixed rule solves the task completely, AI may not be necessary.

2. Create a source register

Gather only the material needed for the selected task. Record an owner, version, scope and review date for each source. A date alone does not establish that content is current; someone must confirm it.

MaterialWhat to checkOwner
Service descriptionScope, limitations and applicable termsService owner
Enquiry handling processStages, responsibilities and exceptionsProcess owner
Frequently asked questionsConsistency with current termsAssigned editor
Human handoff rulesReasons to stop and the receiving ownerWorkflow owner

These are example responsibilities, not a requirement to create a separate position for every document. One person can hold several roles when ownership is clear.

3. Remove contradictions and unnecessary data

If two documents state different terms, do not leave the choice to the assistant. Find the owner, identify the current version and mark the outdated material. If the conflict is unresolved, the appropriate result is to state the uncertainty and refer it to a person.

Separate topics with clear headings. Conditions, exceptions and limitations should sit near the statement they qualify. A short rule without an important exception can produce a convincing but incorrect answer.

Do not connect every conversation and file by default. Remove secrets and unnecessary personal data. Source permissions and the assistant’s access rules must be defined separately; access to one document should not automatically expose the rest.

4. Separate knowledge from authority

Source material explains the subject but should not arbitrarily expand the system’s capabilities. An instruction inside an uploaded document is not permission to send messages, disclose information or change records.

  • May read: only agreed sources within the task.
  • May prepare: an answer, summary or suggested next action.
  • Needs confirmation: predefined actions with external consequences.
  • Must stop: no reliable answer, conflicting terms or a request outside its authority.

Stopping conditions must be included in system testing. A request to “never make mistakes” in an instruction is not sufficient.

5. Prepare a set of test questions

Use anonymised ordinary and difficult examples. Write the expected behaviour and the source used to check the answer before running each test. Do not judge quality solely by the politeness or confidence of the wording.

SituationWhat to check
The current document contains the answerThe meaning and conditions are accurate
The required information is missingNo invented fact; clarification or human help is offered
Sources contradict each otherThe conflict is not hidden behind a confident answer
The question involves an exceptionThe limitation is retained
A prohibited action is requestedThe action is not performed
A document contains an unrelated commandIt does not change agreed permissions
The user asks for a team memberThe agreed handoff works and preserves context

Record results separately for accuracy, source support, inclusion of limitations and respect for boundaries. A single overall score can hide an important failure.

6. Decide how the knowledge base will be maintained

Before launch, assign ownership of updates and errors. What happens when a service changes? Who reviews new material? Which scenarios must be rerun before capabilities are expanded?

Start with a limited pilot where the result can be observed and corrected. Do not treat a successful demonstration as evidence that the assistant is ready for every question. Acceptance establishes only the agreed scenarios and conditions.

A useful handover includes the source register, task card, action boundaries, tests, known limitations and route to a person. This makes quality a concrete discussion rather than an impression that the assistant “sounds human”.