Blog · AI for small and midsize businesses
An AI assistant for your company manuals and policies
An internal assistant should help employees find a supported answer and inspect its source. Start with a narrow set of approved manuals, define who can read them and test questions the material cannot answer. Useful retrieval depends as much on ownership as on the model.
In this article
An employee asks how to request a replacement device. The answer exists, but finding the right manual and current paragraph takes longer than asking a colleague. An internal AI assistant can help when it retrieves relevant approved information and explains it with a source employees can inspect.
Follow the question to the evidence
Retrieval-augmented generation, usually called RAG, means finding relevant information before asking a language model to draft an answer. The Microsoft overview describes this grounding pattern. Retrieval does not guarantee that the retrieved passage or final interpretation will be correct.
- The employee signs in and asks a question.
- The application searches only sources that employee is permitted to use.
- Relevant passages and document identity support a proposed answer.
- The answer names its source and preserves important conditions.
- Missing or conflicting information leads to clarification or referral.
Choose a small source set
Begin with general procedures that have clear owners and modest consequences. Exclude personnel records, disputed policies and live account information from the first scope unless the design explicitly covers them. The policy owner confirms current versions; the technical owner verifies access before information reaches the model.
Microsoft's secure RAG guidance states that grounding information must be authorized for the user. Asking the assistant to keep restricted documents secret is not a substitute for enforcement in the application and data systems.
A hypothetical office-equipment assistant
Suppose a services business starts with device-request, reimbursement-submission and office-access procedures. A question about requesting a laptop should cite the current request route. A question about an individual's reimbursement status should explain that personal status is outside the document assistant and direct the employee to the approved system or owner.
Prepare a worksheet with question, employee role, expected source, required conditions and acceptable referral. Include ambiguous requests, absent information and an outdated document that should not be used. Label examples synthetic unless they come from permitted business records.
Evaluate finding and answering separately
If the answer is wrong, inspect whether the right passage was retrieved. If retrieval succeeded, examine whether the answer omitted conditions or added unsupported detail. Microsoft's evaluation guide recommends assessing individual stages as well as the final experience.
Measure supported answers, useful referrals, time to an accepted answer and reviewer corrections. Assign someone to review unanswered questions and source changes. Start with supervised use before widening the domain.
Before the first live trial, ask a new employee to follow an answer back to its source without help from the developer. Can they locate the relevant conditions and identify the owner? If the citation opens an inaccessible file or an unrelated section, the answer is not yet practically verifiable. Fix that path before increasing the number of documents.
For the broad technology comparison, see chatbots, RAG and agents. To define one practical source set and test workflow, discuss a bounded company knowledge assistant.