AI Employees Knowledge Base: Prepare Before Setup

An AI employees knowledge base should give each AI employee the facts, rules, language and hand-off instructions needed for its role. Before setup, gather current documents, remove contradictions, separate sensitive information, turn staff know-how into clear answers, then test realistic customer questions. Assign one owner to keep it accurate.

For small businesses, preparation decides whether an AI employee gives a useful answer or a vague one. Sevenfold uses the information you provide through your website, uploaded documents and text content, so the quality of that information directly shapes each employee’s answers.

Key takeaways

• Give each AI employee only the information it needs for its role.
• Choose one current version of every policy, price, process and contact detail.
• Write clear answers, limits and escalation rules instead of relying on staff memory.
• Test real customer questions before launch and review the knowledge base after business changes.

What you’ll need before setting up AI employees

You’ll need the documents and decisions that explain how your business works. Start with the material staff already use, then add the answers customers ask for repeatedly. You don’t need a perfect library on day one. You do need a reliable first version with clear ownership and known gaps.

Prepare:

• Your website pages, service descriptions and current business information.
• Price lists, menus, service areas, opening hours and booking rules.
• Customer FAQs, email replies, call notes and common enquiry responses.
• Policies covering cancellations, refunds, eligibility, delivery or collection.
• Role instructions for the AI employee you’re setting up.
• A list of subjects that require a human hand-off.
• Example customer questions, including awkward or incomplete ones.
• One person responsible for approving updates.

Sevenfold’s AI employees are built around defined roles. Desky handles calls, bookings, routing and reminders. Inky writes blogs, emails and ad copy. Bossie handles calendar and inbox administration. The knowledge base should match the employee’s job rather than dump every business file into one large folder.

1. What should an AI employee knowledge base contain?

An AI employee knowledge base should contain verified information that helps one employee complete its assigned work. Include facts, approved answers, process steps, boundaries and escalation paths. Leave out old drafts, private material the employee doesn’t need, duplicate policies and instructions that conflict with the current way you operate.

A useful knowledge base answers five practical questions:

1. What does the business offer? Include services, products, locations and the customers you serve.
2. How does the work happen? Document booking, enquiry, sales, content or admin processes.
3. What may the AI employee say or do? Set communication standards and role limits.
4. What must it never guess? Name missing information, restricted topics and approval points.
5. When should it hand off? Give a clear trigger and explain what happens next.

Think in content types, not file types. A PDF, webpage or text document is only the container. The useful content inside it might be a fact, a question-and-answer pair, a process, a script or a rule.

A knowledge base is a working instruction set, not a document dump. Its value comes from accurate, role-specific information that an AI employee can use to answer or act. If a file contains three old prices and one current price, uploading the file without clarification creates a decision problem rather than solving one.

For a wider look at role-based AI staff, see how to create custom AI staff for repetitive business tasks.

2. Which business documents should you collect first?

Collect the documents that affect customer answers and daily work before adding background material. Start with current public information, operating rules and frequently used staff references. A short, approved set of documents is more useful for setup than a large archive nobody has checked recently.

Use this collection order:

1. Public customer information: website copy, service pages, contact details and opening hours.
2. Commercial information: prices, inclusions, exclusions, service areas and qualification criteria.
3. Operational rules: booking steps, availability, cancellations, refunds and delivery arrangements.
4. Communication examples: approved emails, call scripts, FAQs and responses to common objections.
5. Role-specific material: sales notes for Sally, call and booking rules for Desky, or content guidance for Inky.
6. Escalation information: who handles complaints, unusual requests, urgent matters and approvals.

Do not treat internal files as automatically suitable. A staff handbook may contain useful procedures alongside personal data, outdated instructions or comments written for humans who already know the context.

Before you add a document, record its owner and review date. If nobody knows who approves a file, nobody can confidently decide whether it should remain in the knowledge base.

The best first upload is the document your team trusts today, not the document with the longest history. Current policies and customer-facing answers should come before old proposals, training archives or general business background. This order gives setup a clear operating base and makes testing easier because every answer has a known source.

3. How do you clean and organise the information?

Clean the knowledge base by removing duplicates, marking outdated material and resolving contradictions before setup. Give each important rule one approved wording. Organise content by AI employee and task, then label anything still waiting for a business decision instead of allowing the AI employee to interpret it.

Use a simple review pass for every item:

• Keep: The information is current, useful and approved.
• Rewrite: The information is correct but vague, incomplete or written as internal shorthand.
• Merge: Two documents cover the same rule and should become one source.
• Remove: The item is obsolete, duplicated or irrelevant to the role.
• Escalate: The business hasn’t decided what the correct answer should be.
• Restrict: The material contains personal or confidential information that the AI employee doesn’t need.

Watch for words such as “usually,” “as needed,” “soon” and “we may.” They often hide decisions. Replace them with a condition, an action or a hand-off rule.

For example, “We sometimes take urgent bookings” leaves too much open. A better instruction names what counts as urgent, who checks availability and what the AI employee should tell the customer if no slot is available.

The NIST generative AI risk guidance recommends identifying and managing risks linked to inaccurate or misleading generated content. For a small business, that starts with ordinary housekeeping: current source material, clear ownership and a process for correcting wrong answers.

4. How should you handle confidential and personal information?

Separate information by purpose before adding it to the knowledge base. An AI employee should receive the customer and business facts needed for its role, not every record your team can access. Remove unnecessary personal details, restrict sensitive material and create a hand-off for requests that need human judgement or secure handling.

Create three practical categories:

Don’t paste a customer list into a general knowledge document just because it contains useful context. Extract the business rule instead. For example, record the approved lead qualification criteria, not every person who has previously met it.

You should also tell the AI employee what it must not disclose. “Don’t share internal notes” is a start, but a stronger rule identifies the type of information, the response to a request and the person or team who should take over.

The Federal Trade Commission’s guidance on privacy and confidentiality commitments reinforces a plain rule: businesses should handle information according to the promises they make about it. If you’re unsure whether material belongs in the knowledge base, ask a qualified privacy or legal professional before adding it.

5. How do you turn staff knowledge into clear AI instructions?

Turn informal staff knowledge into short, explicit instructions by recording the decision behind each answer. Ask experienced team members what they check, what exceptions they watch for and when they ask someone else. Then rewrite that knowledge as conditions, actions, approved language and hand-off points.

A practical interview prompt is:

• What does the customer usually ask?
• What information do you need before answering?
• What answer is approved?
• What exceptions change the answer?
• What must be recorded or confirmed?
• Who takes over if the request falls outside the normal process?

Convert the result into a consistent format:

Situation: A customer asks to change an appointment.
Check: Confirm the booking details and the business’s change policy.
Action: Offer the permitted options and confirm the selected time.
Limit: Don’t promise an unavailable slot.
Hand-off: Transfer the request if it requires an exception or approval.

This format helps a receptionist, sales employee or support employee follow the same reasoning each time. It also exposes missing decisions before customers find them.

Staff expertise becomes useful to an AI employee only after the hidden decisions are written down. People fill gaps with memory, context and judgement. An AI employee needs those gaps named through rules, examples and escalation instructions. The writing work is where much of the setup value sits.

6. What rules and escalation paths should you define?

Define rules for accuracy, authority, tone and hand-off before an AI employee starts customer-facing work. The employee should know what it can answer from the knowledge base, what it can do using its role, what it must refuse and when a human should take over.

Write four groups of instructions:

1. Answer rules: Use approved business information and say when the answer isn’t covered.
2. Action rules: State what the employee may book, route, remind, draft or record.
3. Tone rules: Describe the shop’s voice with examples rather than broad labels such as “professional.”
4. Escalation rules: Name the trigger, the human destination and the message given to the customer.

Sevenfold’s published knowledge states that its AI employees use the information supplied through websites, uploaded documents and text content. When something isn’t covered, they don’t invent details. They say so and offer to transfer the customer to a human. Your own instructions should make those limits easy to follow.

For example, Desky may need to answer calls, book appointments, route enquiries and send reminders. A booking rule should state which details to confirm, not only “book appointments politely.” A sales employee may follow up leads, but its qualification criteria and approval limits still need to be written down.

The NIST AI Risk Management Framework core puts responsibility, measurement and ongoing management at the centre of AI use. Give your team a named person who can review an answer, approve a change and decide when an issue needs attention.

7. How can you test the knowledge base before launch?

Test the knowledge base with real questions, incomplete questions, conflicting questions and questions the AI employee should hand off. Check not only whether the answer sounds good, but whether it uses the current rule, avoids guessing and takes the correct next step for that role.

Build a test set from:

• Recent customer emails and call topics, with personal details removed.
• Questions new staff often ask during training.
• Requests involving prices, availability, cancellations or exceptions.
• Misspelled, short or unclear customer messages.
• Questions outside the business’s services.
• Requests for confidential information.
• Situations that need a person to approve the next action.

Score each answer against four checks:

Keep a defect log. Record the question, the answer, the source it should have used, the correction and the owner. This turns testing into knowledge base management rather than a one-time approval ritual.

Sevenfold’s onboarding guidance describes briefing the AI team, connecting business tools, setting follow-up sequences and going live. Use the briefing stage to fix knowledge gaps before launch, not after a customer receives a wrong answer.

8. Who should own and update the knowledge base?

One person should own the knowledge base, even if several people contribute content. The owner approves changes, checks review dates, removes old information and keeps a record of unresolved questions. Each role can have a subject-matter reviewer, but responsibility for the final version should be unmistakable.

Set a simple update routine:

• Review time-sensitive details after every change to prices, services, hours or policies.
• Capture questions that the AI employee couldn’t answer confidently.
• Check hand-offs after staff, phone or booking processes change.
• Retire old documents instead of leaving competing versions in place.
• Re-run the test set after a major update.
• Keep a short change log with the date, editor and reason.

A knowledge base should change when the business changes. That includes a new service, revised cancellation rule, seasonal opening time or different approval process. If the source changes but the AI employee’s instructions don’t, the business has created two versions of the truth.

Read how to manage AI employees in a small business for the wider operating work that follows setup. The knowledge base is one part of that work, but it is the part that controls what the employee can reliably say.

Frequently asked questions

What is an AI employee knowledge base?

An AI employee knowledge base is an organised collection of business facts, documents, Q&A pairs, role instructions, communication standards and escalation rules. It gives an AI employee the information needed for its assigned work. In Sevenfold, supplied website content, uploaded documents and text content provide the business knowledge employees use.

How do I add a knowledge base to an AI employee?

Prepare and approve the role-specific information first, then provide the relevant website content, documents and text material during setup. Separate current facts from internal rules and mark subjects that require a human. Test the employee’s answers before going live, especially around prices, bookings, exceptions and hand-offs.

What information should be included in an AI employee knowledge base?

Include current services, prices, locations, hours, policies, FAQs, role instructions, approved language and escalation paths. Add process details only when that employee needs them. Exclude duplicate drafts, unnecessary personal information, obsolete policies and any material the employee has no reason to access.

How does an AI employee use a knowledge base?

An AI employee uses the supplied information to answer questions and follow role instructions. The exact work depends on the employee. Desky can handle calls, bookings, routing and reminders, while Inky writes blogs, emails and ad copy. If a detail isn’t covered, the employee should avoid inventing an answer and offer a human hand-off.

How do I keep an AI employee knowledge base accurate and up to date?

Assign one owner, record review dates and update the source whenever a price, service, policy, process or business detail changes. Keep a list of unanswered questions and test corrected content against real customer examples. Remove old versions so the AI employee has one approved rule to follow.

Prepare your AI workforce with the right information

Good setup starts before the AI employee is configured. Give each role a clean body of knowledge, write down the decisions staff normally keep in their heads, set firm boundaries and test the answers that matter most to customers.

Browse Sevenfold’s AI employee profiles to see how defined roles fit different business functions, then contact Sevenfold to discuss your requirements and setup.