An internal AI assistant should be designed to recognize missing authority, conflicting information, restricted content and decisions that still belong to qualified people.
An internal AI assistant should refuse to answer when it cannot find an approved source, when available sources conflict, when the employee lacks permission to see the information or when the question requires legal, financial, personnel, safety or executive judgment.
Refusal does not always mean ending the conversation. A well-designed assistant can explain why it cannot answer, identify the appropriate person or process and preserve enough context for an effective handoff.
The safest assistant is not the one that answers every question. It is the one that knows which questions it is authorized and qualified to answer.
When Is an Approved Answer Missing?
A knowledge assistant should answer from information the organization has reviewed and approved for that purpose.
Suppose an employee asks:
“What is our policy for reimbursing mileage when a personal vehicle is used for an emergency delivery?”
The assistant finds the standard mileage policy, but the document does not address emergency deliveries. It should not infer that the normal rule applies. It should explain that the approved source does not cover the exception and direct the employee to the designated finance or operations owner.
This response is more useful than a plausible guess. It identifies a gap the organization can correct while preventing an unsupported answer from becoming an unofficial policy.
The same principle applies when important knowledge has never been documented. An AI assistant cannot reliably retrieve an approval rule, escalation path or exception procedure that exists only in a manager’s memory.
What Should Happen When Sources Conflict?
An assistant should not silently choose among conflicting documents.
Imagine that SharePoint contains two purchasing procedures. One says department managers may approve purchases up to $2,500. Another sets the limit at $5,000. Both appear current, and neither identifies the other as superseded.
The assistant cannot determine which limit leadership intended simply by comparing the language. It should state that the available sources conflict, identify the documents involved and refer the question to the knowledge owner.
The conflict should then become a tracked governance issue. The appropriate owner must determine which procedure is authoritative, archive or correct the other version and publish a clear effective date.
The AskMaisy article on using Teams feedback to improve weak or incorrect AI answers explains why these failures should become correction tasks rather than disappear into chat history. Different problems—such as conflicting documents, missing knowledge, broken citations and incorrect permissions—require different remedies.
When Should Permissions Stop the Answer?
An assistant must not reveal information from a source the employee is not authorized to access.
For example, a general employee might ask:
“What salary range has leadership approved for the new operations director?”
Even when the answer exists in a restricted HR or executive document, the assistant should not summarize it, reveal its title or hint at its contents if the employee lacks permission.
Permission-aware retrieval should use the employee’s authenticated identity and the access rules applied to the source. The assistant should not create a second security system that overrides SharePoint or Microsoft 365 permissions.
A refusal in this situation should be carefully worded. Saying “You do not have access to the executive compensation plan” may itself reveal the existence of sensitive information. A safer response may say that the assistant cannot provide an approved answer from sources available to the user and identify the appropriate department for questions about the subject.
Which Decisions Should Always Remain With People?
Some questions require judgment, accountability or professional authority rather than document retrieval.
An assistant may retrieve an approved workplace policy. It should not independently decide whether an employee must be disciplined.
It may explain the organization’s contract-review process. It should not decide whether a contract is legally acceptable.
It may display financial approval rules. It should not authorize an unusual expenditure.
It may retrieve an emergency procedure. It should not decide whether a developing situation presents an immediate threat.
It may explain the board’s succession process. It should not select or evaluate the next executive director.
Microsoft’s Copilot Studio agent-design framework recommends defining what an agent may do autonomously, what requires human approval or review and when the agent must escalate or defer.
The boundary should reflect the consequences of being wrong. Legal, HR, financial, safety, clinical, compliance and executive matters generally require more human oversight than routine questions about forms or operating procedures.
Should the Assistant Refuse When a Question Is Ambiguous?
Sometimes the problem is not missing information but an unclear request.
An employee asks:
“Can I approve this?”
The assistant does not know what “this” refers to, which department the employee represents, the amount involved or which approval policy applies.
It should ask a focused follow-up question rather than guess. When clarification still does not establish the correct context, the assistant should escalate.
Ambiguity becomes especially important when the assistant can perform actions. Sending a draft link is relatively low risk. Changing a customer record, submitting a payment request or notifying an external party may be difficult to reverse.
The higher the consequence, the stronger the need for clarification, confirmation and human approval.
What Is the Difference Between Refusal and Escalation?
A refusal means the assistant will not provide the requested answer or action.
An escalation gives the employee a safe next step.
A useful escalation might:
- Identify the responsible role or department
- Link to an approved request or review process
- Explain which information the employee should provide
- Preserve the question and relevant conversation context
- Create a controlled correction or support task
- State whether urgent action is required under an approved procedure
Microsoft Copilot Studio includes system topics for fallback and escalation. Its documentation explains that the Escalate topic can be triggered when a user asks to speak with a person or when the agent calls the escalation event. A live-agent handoff can also transfer conversation history and relevant variables to a connected engagement service.
Not every internal organization needs a staffed contact center. Escalation could instead mean directing the employee to a manager, opening a Microsoft Teams support channel or creating a Power Automate task for the knowledge owner.
How Should Refusal Rules Be Designed?
Refusal and escalation rules should be defined during discovery, not improvised after launch.
For each knowledge area, the organization should decide:
- Which questions the assistant may answer
- Which sources are authoritative
- Which roles may access them
- Which exceptions require human review
- Which decisions the assistant must never make
- Who receives escalated questions
- What response is required for urgent situations
- How failed answers become content improvements
The AskMaisy practical introduction to Copilot Studio for nonprofits illustrates why a focused agent with a defined audience is safer than one expected to answer every organizational question. Approved sources, permissions, content ownership and continuing governance must be established before the conversational layer can be trusted.
NIST’s AI Risk Management Framework also calls for organizations to identify AI capabilities that require human oversight and to establish oversight practices based on the system’s context, limitations and risks.
When Should an Internal AI Assistant Refuse to Answer?
It should refuse when no approved answer exists, sources disagree, access is restricted, the request remains ambiguous or the decision requires accountable human judgment.
A responsible refusal should not leave the employee stranded. It should explain the limitation without exposing sensitive information and direct the question to the appropriate person or process.
An internal assistant earns trust not by sounding certain about everything, but by answering from approved knowledge and stopping when that knowledge no longer supports a safe response.
