Evan Dunkel

Ambient is a concept for support specialists who need to check an AI answer before it reaches a customer. I brought the draft, the supporting excerpt, and the send decision into one proposed review flow. The intended outcome is easier verification without losing an edit or switching between tabs.

Role: Independent product designer. Team: Solo concept. Timeline: October 2026. Platform: Web. Status: Concept; not shipped.

I owned the problem framing, the comparison of directions, and the interaction specification. Research, implementation, and product results remain proposed. The photos in this case study are placeholders.

A support specialist needs to resolve a customer’s question while taking responsibility for the guidance they send. In this scenario, the draft is in one tab and the relevant policy is in another. A fluent response can conceal an outdated rule, a missing source, or a policy that does not apply to this customer.

The starting brief was to make the assistant’s answers easier to read. I reframed it around making individual claims easier to verify: better formatting cannot establish whether the evidence supports the answer. That shift makes review the main task rather than a final check after the assistant appears to be finished.

This is a design scenario, not a finding from interviews or support data. Before treating it as a product opportunity, I would examine real review tasks and ask specialists where they lose time or miss conflicting guidance. Doing nothing could leave the specialist repeating the same search, or sending a polished answer without noticing the gap.

Constraints: One question at a time; existing document search; no automatic sending. Principles: Evidence over apparent certainty; human review over automatic completion.

I compared three directions against one question: can the specialist understand why this answer is safe to send?

Chat only: The familiar, compact option. It keeps reading easy, but puts citations below the answer and makes the specialist reconstruct the connection between a claim and its source.

Automatic reply: The boldest shortcut. It could remove the review step entirely, but that also removes the specialist’s opportunity to catch an unsupported answer before it reaches a customer.

Answer and evidence: My chosen direction. The draft stays readable while the relevant excerpt opens alongside it. It uses more space and requires a responsive layout, in exchange for keeping the evidence close to the decision.

I kept the scope to one claim and one excerpt. Multi-document comparison could help with complex questions, but would make the first review flow harder to evaluate.

I chose sentence-level references over a source list at the bottom of the response. Selecting a reference would reveal the excerpt, document title, and update date while preserving the draft. The specialist should be able to inspect the source and return to the exact claim they were checking.

The trade-off is density: references interrupt otherwise smooth prose, and an excerpt needs room. I accepted that cost because the connection between a statement and its support matters more than making the answer look finished.

This choice is a hypothesis. I would test whether reviewers catch stale or conflicting guidance more reliably than they do with end-of-answer citations; opening more references alone would not demonstrate better judgment.

I separated preparing the answer from sending it. The assistant proposes an editable draft; the specialist decides when it is ready. An edit should stay intact while the specialist opens a source or returns from an unavailable reference.

I left out a confidence score. A precise-looking number would offer reassurance without explaining which claims are supported, so missing evidence should remain visible as “No source found.” That adds an unresolved state rather than allowing polished prose to hide uncertainty.

Before implementation, I would align with PM on when unsupported content should block sending and when an explicit warning is sufficient. That policy should shape the interaction; the interface should not quietly make it for the team.