Write AI Assistant Instructions That Handle Ambiguity

Separate behavior from business facts

Assistant instructions work best when they define repeatable behavior rather than duplicate a changing handbook. Put the company’s current prices, eligibility rules, addresses, and service details in approved knowledge sources. Use behavioral guidance to say how the assistant should answer, what evidence it needs, when it should ask a question, and when it must stop. This distinction makes maintenance clearer: a policy change calls for a source review, while a new support boundary calls for an instruction review.

Begin with a short response pattern. For example: answer the customer’s direct question from an approved source; mention the relevant condition; ask one clarifying question if that condition is unknown; and refer the request when it depends on an exception. A pattern is more actionable than a pile of broad rules such as “be helpful, accurate, friendly, and professional.”

Add an explicit stop condition. “Do not guess” is a useful principle, but staff need to know what happens next. The assistant might say that the available information does not establish the answer and explain how the customer can contact the team. Specify only a route and commitment the business has verified. Avoid promising that a person will reply within a particular time unless the organization has approved and can meet that service level.

Make ambiguity visible

Use examples to expose ambiguity in the source or instruction. Suppose a customer asks, “Does the basic plan include exports?” The source may say that exports are available on some plans without naming which ones. The assistant should not infer the answer from a plan name. The instruction can require it to state that the available material is inconclusive and seek clarification or a human review, while the business owner corrects the underlying plan comparison.

In Resonoon, role and purpose guidance is part of assistant setup, and approved knowledge is managed separately. Keep the response pattern in the role guidance and put authoritative facts in the knowledge material. The editor supports PDF and Markdown uploads. Review the files before relying on them; an uploaded document is not automatically current or correct.

Write instructions in positive operational terms where possible. Instead of “never be vague,” say “name the source-backed condition and ask which case applies when the customer has not supplied it.” Instead of “do not handle refunds,” clarify whether the assistant may explain the published refund policy and where it must refer exception requests. The distinction prevents a boundary from becoming an unnecessary refusal of ordinary information.

Test an ambiguous case and an escalation

Prepare a fixed pair of tests. The ambiguity test should have a plausible answer only if an important detail is known. The escalation test should ask for a decision the assistant has no authority to make. Record the expected answer, the missing detail or boundary, and the appropriate next step before testing.

Use the assistant test page to run the exact prompts. Examine whether the assistant asks a relevant question rather than several unrelated ones, and whether it explains a limit without sounding as if the customer is at fault. For the escalation case, check that it does not invent an approval, promise an exception, or imply a response time. A correct stop is a successful outcome when the question needs a person.

When a test fails, diagnose the cause before editing. Missing or conflicting facts belong in the source review. A missing stop condition belongs in behavioral guidance. An escalation with no assigned owner is an operational gap for the business. Change one cause, rerun the unchanged prompt, and record the result. Do not continually rewrite the question until the answer happens to look acceptable.

Maintain instructions as policy changes

Instructions should be reviewed when responsibilities, policies, or source ownership change. Keep a dated copy and a small regression set with common, ambiguous, and escalation requests. Store this record where the team can compare the wording and test results from successive reviews. Ask a policy owner to approve instructions that affect refunds, safety, regulated advice, or commitments to customers.

A concise instruction set is easier to audit than an exhaustive page of overlapping rules. If two instructions conflict, resolve the conflict instead of adding another sentence. Then test both the ordinary path and the stop condition. For broader examples of support context, visit Resonoon’s AI customer support use case. The goal is a response pattern that employees can explain and customers can rely on, within the evidence and authority actually available.

Have a colleague read the instructions without seeing the draft history. Ask them to predict what the assistant should do in an ordinary case and an exception. If their interpretation differs from the author’s, clarify the wording before launch.