A redrawn sample. The analysis is opt-in, runs in a queue while the consultant keeps working, and is never shown to the client.
The problem
Consultants already had an offers screen: the bank’s campaign products that suit this client. The bank wanted AI to explain why an offer fits a client and suggest how to start the conversation. Three things made that hard: AI takes minutes, not seconds, and runs in a queue; it costs money per run, so it shouldn’t run by default; and what it says about a client is sensitive, so it shouldn’t appear in front of them by accident.
Opt in, with the cost and the wait up front
The analysis is off by default. Turning it on opens a short dialog: what kind of help (a client narrative, a call script), how long each one will take (“8 ahead · ~4 min”), and a plain note that AI runs cost extra. “Continue without AI” is always there.
Say who will see it
The app works in two modes. In client mode the consultant is with the client in the branch, the client authenticated. In ghost mode the consultant prepares alone, for example before a phone call. The AI analysis belongs to preparation: the dialog and the panel both say it’s for the consultant and isn’t shown to the client. The product owner also dropped the assistant’s persona name early on, so the interface simply says “AI analysis”: a tool, not a colleague.
Never block the consultant
While the analysis runs, a small chip in the header shows the queue (“8 ahead · ~4 min”) and the consultant keeps working with the regular offers. When it’s done, a toast appears on whatever screen they’re on, and the chip turns to “ready”. Try it in the demo above.
Redrawn with invented values
Step 1 of 2: Queued
An explanation they can use, and a way to rate it
The panel reads like a short briefing: what stands out about this client, and one question to open the conversation. Each offer also gets an AI fit score next to the campaign score. Under the panel, thumbs up or down — a thumbs-down asks what was wrong — because the team wanted to learn which explanations helped.
Outcome
Designed end to end as the only designer on the feature, and now coded. It isn’t live yet, so there is no usage data or feedback; the thumbs up and down are in place to collect it from day one.