Someone asks which supplier document belongs in a customer handover pack. The answer exists somewhere across a shared drive, an old email and a colleague’s memory. A virtual assistant finds three plausible files; an AI tool combines them into one confident answer. Neither has established which source applies.
A useful internal knowledge base helps people find an answer, understand its limits and check the supporting source. Start with a small set of recurring questions. A folder containing everything the business has ever saved is a poor starting point for either human lookup or AI-assisted answers.
Organise around questions people actually ask
Collect questions from recent handovers, support requests and internal escalations. Remove unnecessary customer details and group questions by the decision they support: which document to use, who can approve an exception or where to find a current service description.
Choose one workflow for the first release. Write entry titles in the team’s language, including common alternative terms. “Find the current product manual” is easier to recognise than “Document repository administration”. Add product or service identifiers where they change the answer.
Separate general guidance from individual job records. An entry can explain how to find a customer’s approved handover documents without copying those documents into a broadly searchable library. If the missing knowledge is still in the founder’s head, use our VA-led SOP capture process first.
Make source priority and scope visible
Agree which source controls each type of answer. For product instructions, that might be the applicable manufacturer document. For delegated authority, it may be an approved internal policy. For a customer’s agreed scope, the authorised job record may be relevant. Where external requirements apply, the responsible specialist must confirm their interpretation.
Do not give every search result equal authority. Label approved guidance, reference material, drafts and historical records. A helpful email can prompt an update; it does not automatically replace an approved instruction. Equally, an internal summary cannot override an applicable external requirement.
Each entry needs an owner, applicable scope, approval status, effective date where relevant and a last-checked date. Show a review trigger or next review date. A recent review stamp alone does not establish that the answer fits a particular product, contract or job.
Build one answerable entry
Use this fictional supplier-document entry as a starting format. It links the short answer to the evidence and defines where the answer stops.
| Field | Example entry |
|---|---|
| Question | Which product manual should go in this handover pack? |
| Scope | Product family covered by this entry; other products go to the document owner. |
| Short answer | Match the supplied model and revision against the approved document register, then use the linked current file. |
| Required context | Exact model from the authorised job record and any relevant revision information. |
| Authoritative source | Applicable manufacturer document, with the approved register linking to its location. |
| Owner and validity | Document owner; approved version; effective date; last source check; review due date. |
| Stop condition | Model mismatch, missing revision or conflicting sources: request owner review. |
| Access | General lookup instructions available to the support team; customer records remain restricted. |
Put the direct answer near the top, followed by qualifications and source links. Keep each entry focused enough that a reader can tell whether it answers their question. Link to a procedure when the next step requires detailed instructions.
An AI-generated summary should be distinguishable from the approved source. Give the reviewer a direct route to the evidence rather than another summary of it.
Set access before connecting AI search
Decide who may read, edit and approve each collection. Then check what any connected AI tool can retrieve and what it can expose in answers. The person configuring a connector may have broader access than the VA who will use it.
Start with approved general information and keep payroll, identity documents and individual customer records outside that initial collection. Have the system owner verify permissions on search results, excerpts and generated responses, including any indexed copies. A prompt telling AI to respect access rules is not a substitute for enforced permissions.
The OAIC recommends assessing product suitability, human oversight, privacy risks and access to information before adoption. Its best-practice advice is to keep personal information, especially sensitive information, out of publicly available generative AI tools. Read the OAIC’s commercial AI guidance.
Define the tool’s permitted job: find and draft from approved material, with source links. Finding an answer does not grant authority to change a policy, approve a document or send a customer commitment.
Run a small lookup exercise
Take ten actual recurring questions, stripped of unnecessary personal information. Ask the process owner to identify the expected source and acceptable answer before testing. Have a VA unfamiliar with the library try ordinary search, then test any proposed AI search against the same questions.
Include these cases:
- A straightforward question with one approved answer.
- A question expressed using a different everyday term.
- A question missing the model or other required context.
- Two sources that appear to disagree.
- A question whose answer sits outside the user’s access.
Record the source found, whether its scope fits and whether the response should answer, ask for context or escalate. Check that a citation actually supports the answer. A linked document is not proof that the tool interpreted it correctly.
The National AI Centre recommends testing before deployment, ongoing monitoring and human oversight suited to the stakes and autonomy involved. This exercise is one practical starting point, not complete assurance. See the current AI adoption foundations.
Turn missing answers into managed work
Maintain an unanswered-question queue with the question, sources tried, missing context, owner and next action. Keep customer-specific detail in its authorised system. Repeated questions can reveal poor titles, hidden sources or a decision nobody has made.
If an entry has expired or its source has changed, mark it for verification rather than presenting its answer as current. When sources conflict, flag the affected entry and withhold the disputed answer pending owner review. Once resolved, update the entry and check that search and AI retrieval no longer present the obsolete answer. Our solar operations playbook covers maintaining approved instructions after release.
Talk to Nexus Group Global about VA support for organising knowledge, testing lookups and keeping unanswered questions moving to the right owner.
