Build status: EvidCue is an illustrative concept website, not a live answering service. The workflow below is a design proposal, not a feature demonstration.
Approved answer cards, retrieved evidence, and AI drafts are not the same thing
Short answer: An approved card is wording a responsible team has explicitly authorized for a defined question. A retrieved passage is material to inspect, not permission to repeat it as an approved answer. An AI draft is generated wording that a person must verify before use. These are different authority levels, even when each appears on the same screen.
What each route should tell the speaker
| Route | Useful label and next action |
|---|---|
| Approved card | “Approved wording.” Check the question still fits the card’s scope and that its owner, source and version remain current. If not, do not stretch the card. |
| Retrieved passage | “Source evidence—not approved wording.” Show the document, section and relevant passage. Read its qualifications and date before summarizing it. |
| AI draft | “Generated draft—review required.” Check the interpretation of the question, compare each assertion against its sources and identify unsupported inferences. Do not present the draft as approved wording. |
Microsoft describes retrieval-augmented generation as retrieval, augmentation and generation from source content; an index can preserve document titles, URLs or file names for citation. Source access and context still matter. This is background on the design pattern, not evidence that EvidCue has implemented it.[1]
A practical handoff example
Suppose a caller asks whether a documented exception applies to their particular case. A narrowly scoped approved card might state only the standard rule. A retrieved policy passage may describe the exception but omit a key condition. A generated comparison might offer a plausible interpretation. The safe handoff is to show which layer supplied each item, check the missing condition, and say “I’ll verify the exception” if it is unresolved—not to turn the plausible interpretation into an authorized answer.
Four checks before designing a live-cue workflow
- Permission: Is live assistance allowed in this conversation and by the organization that controls the content? Do not use it where external assistance is forbidden.
- Authority: Who approves spoken wording, and how will obsolete or out-of-scope cards be retired?
- Provenance: Can the speaker open the actual source and see its owner, version, date and limitations? A citation alone is not approval.
- Human review: Who must review a generated draft, and what happens when no source supports a confident answer? NIST's AI RMF emphasizes defined human roles and oversight; it does not certify this proposed product.[2]
This distinction is not theoretical for contact centers: Cognita describes how agents may struggle to access policy and troubleshooting material while actually on a call. That is its reported problem framing, not evidence of demand for EvidCue or of any EvidCue customer.[3]
Where the separate phone fits—and where it does not
EvidCue proposes displaying these cues on a phone during permitted conversations. An ordinary share of a computer window need not include a separate phone, but cameras, phone mirroring, whole-device capture or other setups may expose it. The separate device does not change the rules of the conversation or the speaker’s responsibility. The three-route illustration on the concept homepage is not connected to a functioning standalone answering service.
Sources
- Microsoft Learn, “Retrieval augmented generation (RAG) and indexes”, accessed 25 September 2026.
- NIST AI RMF, Appendix C: AI Risk Management and Human-AI Interaction, accessed 25 September 2026.
- Cognita Knowledge Management, “KM for Contact Centers”, published 2 July 2025; accessed 25 September 2026.