AI · Knowledge management · 2026-11-21
Internal AI assistants: what should you check before a pilot?
A practical way to evaluate an internal knowledge assistant: trusted documents, access checks, difficult questions and clear decision criteria.
Start with a useful decision, not a demonstration
An assistant answers three questions correctly during a demonstration. Is it ready for a team to use? Not yet: you need to check what happens when documentation is missing, two procedures contradict each other or the user has different permissions from the person who built it. The useful question is: within which scope can we trust this assistant, with what limitations and what supervision?
Here is our proposed preparation method for an assistant that helps people find internal procedures. It does not change records in business tools or make decisions on behalf of an accountable manager. That limited scope makes an initial assessment easier. This is a recommended working approach, not a claim about results achieved for a customer.
1. Assemble a small set of reliable references
Choose a specific need, such as directing employees to the correct equipment request procedure. Gather the necessary documents and assign an owner to each item. Check its validity date, intended audience and which procedure takes precedence. Clearly identify outdated versions; the assistant should not have to guess which policy the organisation actually follows.
For each important question, prepare a short record containing the expected answer, supporting document, applicable conditions and contact for exceptions. If nobody can write a reference answer, the first problem is documentation. Fix it before changing the assistant’s instructions. This preparation also helps the people who currently answer these questions manually, making the work useful beyond the pilot itself.
2. Check access with representative users
Microsoft states that the documented SharePoint integration for Copilot Studio agents using the standard harness returns only content the user can access. This behaviour depends on the integration method and configured authentication; do not assume it applies to every uploaded file or connector.
For your pilot, define a simple matrix: user role, permitted documents and information that must remain inaccessible. Test with representative test accounts under your organisation’s rules, rather than relying solely on an administrator account. Also check that the user can open the source link. A document title or excerpt can itself contain sensitive information, so review the complete response rather than only its main recommendation.
3. Challenge the assistant with difficult questions
Entirely fictional example: an organisation prepares twenty questions about equipment requests. Twelve concern documented procedures, four require clarification, two have no answer in the sources and two concern information restricted to another role. This distribution illustrates a test set; it is neither an official threshold nor a guarantee of quality.
Define the expected behaviour for each case before testing: answer with the right reference, ask a clarifying question, acknowledge missing information or direct the user to the right person without exposing restricted content. Microsoft provides repeatable test sets in Copilot Studio. Supplement their results with human review of important cases: an overall score cannot, by itself, validate an operational answer.
4. Make a decision and plan maintenance
Classify discrepancies by their consequences: sending someone to the wrong team, providing incorrect instructions, citing outdated material or disclosing information. We recommend preventing expansion of the pilot if inappropriate access is observed, and agreeing other acceptance criteria with business owners before testing. After a fix, rerun the affected questions and a stable set of reference questions to detect regressions.
Finally, identify who updates documents, handles reported answers and can suspend the pilot. Keep a record of tested versions and decisions. ITSMWORK supports AI use-case scoping, knowledge governance and evaluation criteria. A focused workshop can produce the scope, access matrix and question set needed to make an initial decision supported by evidence.