EvidCueConcept notes

Build status: EvidCue is an illustrative concept website, not a live answering service. This checklist is guidance for designing a permitted workflow, not a feature demonstration.

How to review an AI answer draft before speaking

Short answer: Confirm the question, inspect the underlying sources, separate direct evidence from inference, check whether the wording is approved for this situation, and withhold any unresolved claim. A fluent draft is not an approved answer. Use assistance only where the conversation’s rules permit it.

A five-check review

  1. Did it hear the right question? Restate the actual question in a sentence. If transcription lost a name, negation, date or qualifier, fix that before looking at the proposed answer.
  2. Can you open each cited source? Check the original document, relevant section, date, owner and access permission. A title or citation marker alone does not show that the passage supports the draft. Microsoft notes that a retrieval index can carry titles and URLs for citations, but even a grounded response can be wrong when retrieval is incomplete or irrelevant.[1]
  3. Which statements are quoted, inferred or unsupported? Mark each substantive clause as directly stated in the source, a human interpretation, or unresolved. Do not convert an inferred exception into a policy.
  4. Who may authorize this wording? A relevant passage may be context without being approved spoken language. Ask who owns the answer and whether this case falls inside an approved card’s scope. NIST recommends clearly defining human decision and oversight roles in operational AI; it does not validate this proposed workflow.[2]
  5. Can the speaker responsibly use it now? If a key condition is missing, say you will verify it rather than reading the draft aloud. Avoid entering confidential conversation or customer material into tools without the organization’s permission and data-handling controls.

An example: a policy exception

A customer asks whether an exception applies today. A draft says, “Yes, this account qualifies.” The source says only that some accounts may qualify after a separate review. The missing review is decisive: the speaker can explain that an exception exists, but should not promise eligibility before checking it.

Why the label matters

An approved card is prewritten wording within a defined scope. A retrieved passage is source material to inspect. An AI draft is generated synthesis requiring review. Keep these authority levels visible and distinguish a source’s statement from the system’s interpretation. See our guide to cards, retrieval and drafts for the wider routing concept.

Limits of the separate-device idea

The EvidCue concept proposes showing cues on a phone during authorized conversations. An ordinary computer-window share need not include a separate phone, but camera views, mirroring, whole-device capture and other setups may expose it. The device changes neither the conversation’s rules nor the speaker’s responsibility. The homepage illustration is not connected to a standalone three-route service.

Sources

  1. Microsoft Learn, “Retrieval augmented generation (RAG) and indexes”, accessed 26 September 2026.
  2. NIST AI RMF 1.0, Appendix C: Human-AI Interaction, accessed 26 September 2026.