My role
- Visual design of components and addition to consistent AI design language
- Accessibility audit of positive and negative color association and general color choices across application as a result
- Marketing imagery for launch of feature
The problem: feedback without a fuse
Survey platforms are good at collecting signal and bad at forcing action. A dissatisfied customer's survey would land in an inbox alongside dozens of positive ones, and unless someone was actively reading every response, it sat there. The cost of that silence is asymmetric: a happy customer ignored is a missed thank-you; an unhappy customer ignored is a churned account and a public one-star review.
The businesses we serve (remodelers, builders, home services companies) don't have CX teams. The person reading survey feedback is often the owner, between site visits. The design problem wasn't "show more data." It was: how do you get the right feedback in front of a busy person, with enough context to act, before the window closes?
The shape of the solution
The feature works as an escalation arc:
1. Detection. AI monitors incoming survey feedback in real time and flags responses showing dissatisfaction.
2. Comprehension. Every flagged survey comes with an AI-generated summary of the customer's experience and the summary carries context from that customer's past surveys, not just the latest response. The owner doesn't have to reconstruct what happened. 3. Learning. In Analytics, flagged feedback rolls up into ranked recurring themes, comparable across teams and locations. Individual rescues become organizational insight: not just "this customer is unhappy," but "installation delays are our most common failure mode in the [region] market."
Meet people where they already are
At-Risk summaries surface everywhere the owner already looks: email digests (key issues + recommended next steps), the survey inbox (Surveys → Inbox → At Risk), and an AI insights widget on the main dashboard sitting alongside "contact at risk" follow-up cards. This was deliberate. Forcing a new destination would have meant flags going unseen so we made the new feature go viral in known locations.
AI you can steer, not a black box
Users can select or write their own themes that the AI tags surveys with (Account → AI). A roofing company and a kitchen remodeler don't fail in the same ways, and a fixed taxonomy would have forced both into the same generic buckets.
We didn't hand users a blank slate or lock them into our taxonomy. The settings page offers suggested themes they can adopt in one tap, alongside a free-text field for their own. Guidance without prescription. And because many of our customers run multi-location or multi-brand operations, theme configuration is shared at the account-family level, so the whole organization speaks one vocabulary in its reports.
One piece of a coherent system
At-Risk Insights ships inside GuildQuality Assist, alongside At-Risk Alerts, Smart Survey Responses, and the Action Toolbar. Designing it meant designing for the suite: shared patterns for how AI-generated content is presented, labeled, and edited, so users build one mental model instead of four.