“I don’t know” is useful only when it tells you what happens next. An assistant that refuses every difficult question is frustrating. One that fills every gap with a confident sentence is worse. The design problem is deciding which missing facts should stop the task and which can be handled with a clearly stated assumption.
For a practical workflow, define an unresolved state before you need it. Give that state a reason, a record and a next step.
Separate an answer from a decision to act
A writing assistant can offer three possible headings while remaining uncertain about which a reader will prefer. An assistant changing a delivery address should not guess which customer record is intended. The acceptable uncertainty depends on the task and the consequence of being wrong.
Start by listing actions the workflow can take. Mark those that publish information, spend money, expose private material or change a record. Then identify which facts must be established before each action. This is a proposed design exercise, not a universal safety certification.
Use reasons that help someone resolve the problem
| Reason | Example | Useful next step |
|---|---|---|
| Missing source | The supplied file does not contain the price | Ask for the relevant document or omit the price |
| Conflicting evidence | Two instructions give different dates | Identify the authoritative version |
| Ambiguous subject | Two customers share a name | Request an approved distinguishing detail |
| Unavailable tool | The account lookup failed | Report the failure rather than invent the result |
| Outside authority | The task requests an unapproved change | Route to the authorized person |
A technical failure is not evidence that an item does not exist. “The search returned no result” and “there are no matching records” are different statements, especially if the search timed out or covered only part of a system.
A worked example: two dates in a message
Imagine a fictional note: “The draft is due Thursday. The launch date will be confirmed after review.” Someone asks an assistant to put the launch on the calendar. Thursday is nearby, but it is a draft deadline. The workflow should preserve the distinction and ask for the launch date rather than quietly reuse the only date it found.
A useful response is specific: “The note gives Thursday for the draft, but no confirmed launch date. I can create a draft-deadline reminder or wait for the launch date.” That explains the gap and offers bounded choices. It does not claim the assistant has checked another calendar unless it actually has.
Do not manufacture a confidence cutoff
A percentage printed by an assistant is not automatically a calibrated probability. Before using any score to trigger actions, establish what it means in your particular system and test how it relates to actual outcomes. A confident wrong answer is precisely why a separate evidence check matters.
NIST’s Generative AI Profile discusses erroneous content presented confidently as a risk. It does not give a single confidence threshold that makes every use case safe. The policy you choose should reflect the task, evidence and consequences.
Test both answering and stopping
Create paired examples. In one, the required fact is present; in the other, remove it. Add cases with irrelevant dates, similarly named people and failed tools. Score whether the workflow completed the supported task, paused for the right reason and avoided unnecessary refusal when the evidence was sufficient.
Record the cost of each kind of error. A needless clarification can slow a harmless draft. An unsupported automatic action can create a much larger problem. Treat them separately rather than hiding both inside one accuracy percentage.
Make the handoff easy to read
An unresolved record should contain the question, the missing fact, the sources already checked and the next authorized step. Keep sensitive content out of broadly shared logs. If a person supplies the missing fact later, record that input and rerun the relevant check.
The evaluation scorecard can help track recovery and evidence. For the source-checking step, use the citation review worksheet. A good pause is a usable part of the workflow, not a dead end.
AI-generated conceptual illustration; the scenarios are hypothetical.
