Product Prototypes
All prototypes
ReCo · AI interfaces

AI Chat Interface

Creators are sitting on a lot of data — subscriber stats, broadcast performance, engagement trends — but most of it lives behind reports they have to know to look for. The result is that useful signals go unnoticed, and good decisions don't get made. An in-product AI chat interface lets creators ask questions in plain English and get answers that are grounded in their actual account data.

This prototype explores what that interface looks like inside Kit: where it lives, how it responds, how it handles uncertainty, and whether conversational AI is the right model for creators who aren't data analysts.

Open prototype
What's in this prototype
💬
Ask in plain English
No menus, no reports — just type your question. The chat understands Kit concepts: subscribers, broadcasts, sequences, automations. No data literacy required. The prototype shows three different query types to test which feels most natural.
📊
Data-backed answers
Data queries return structured responses — a mini stat card, a trend indicator, and a clear timestamp ("Based on last 30 days · May 2026") so you know exactly what the data covers. Numbers without a source erode trust; numbers with a source build it.
✨
Content on demand
Ask for subject line ideas, email hooks, or calls to action. The prototype returns 3 options with a short rationale for each — surfacing the thinking rather than just the output, so creators can make an informed choice rather than just picking one at random.
🎯
Insight to action in one click
When the AI surfaces an audience insight — "437 subscribers haven't opened anything in 90 days" — a single action button lets you create a segment from that result immediately. The insight and the action stay connected rather than requiring a separate workflow.
🤔
Honest about what it doesn't know
When a query is ambiguous or the AI's confidence is low, it surfaces the most likely things you were trying to ask as clickable options — never a dead end. The prototype includes a "fall-forward" fallback state and a structured clarification card to test which handles uncertainty better.
📍
Sidebar placement
The chat panel lives in a persistent right-side drawer — open alongside whatever you're viewing, without obscuring the data. Charts and stats stay visible while the conversation happens. This prototype tests whether the panel model feels natural or intrusive for creators.
Design & product decisions
Right-side panel over modal or floating button
A floating action button suggests reactive, occasional use — "I have a problem." A modal obscures the data the creator is trying to ask about. A right-side panel makes the AI feel like a persistent collaborator that's always available without taking over the screen.
This is the dominant pattern in mature product-level AI implementations. The panel stays alongside the current view, shows conversation history, and can be collapsed when not needed. One important caveat: HubSpot Breeze, Figma AI, and GitHub Copilot Chat are all desktop-first, power-user tools. Whether this pattern translates to Kit's creators — who often access the product in shorter, lower-intensity sessions — is an open question this prototype is specifically designed to test.
Competitor analysis — HubSpot Breeze · Figma AI · GitHub Copilot Chat
Context-aware prompt suggestions on open
When the chat panel opens, it surfaces 3 suggested prompts based on what the creator is currently looking at. On a broadcast stats view, suggestions might be "Why did open rate drop this week?" or "How does this compare to my last 3 sends?" rather than generic examples.
This solves the articulation barrier — the gap between what a creator wants to know and their ability to phrase it as a data query. For Kit creators who aren't data analysts, being shown what's possible is more useful than a blank input. The concern is whether suggestions feel helpful or prescriptive. If reviewers find the chips constraining rather than orienting, the fallback is a blank input with example prompts available on demand — suggestions become opt-in rather than the default state.
HubSpot Breeze · NNGroup prompt controls research
Fall-forward fallback over generic error states
When the AI can't answer a query confidently, instead of "I don't understand, try rephrasing," it surfaces the 2-3 most likely things the creator was trying to ask as clickable chips. Picking one restarts the query with a clearer intent signal rather than requiring the creator to start from scratch.
Research on chatbot fallback design shows nearly half of users hit an uncertain state in any given session — treating fallback as a core experience rather than an edge case substantially affects perceived quality. A dead end destroys trust; a forward path preserves it.
UX Content Collective — fallback design patterns
Surface AI interpretation before showing results
For ambiguous queries, the AI surfaces its interpretation before generating output: "You asked about opens — I'm showing open rate (unique opens / delivered). Is that right?" The creator can correct the interpretation before acting on the data, rather than discovering a mismatch after the fact.
In analytics contexts, a wrong interpretation can lead to a wrong decision. Making the interpretation explicit and correctable — using a coloured highlight to distinguish it from the response — adds a step but removes a significant failure mode. This is worth testing against user patience.
Notion AI — visual differentiation pattern
Confidence language, not confidence percentages
When the AI's data coverage is partial or a query spans a time window with incomplete data, the response uses natural language to signal this: "Based on what's available..." rather than displaying a percentage score. This is a deliberate design principle: Kit's creators are not data analysts, and surfacing a number like "67% confident" creates anxiety without helping them decide what to do next. Precision matters less than trust.
HubSpot's Breeze Copilot includes a Sources section listing the records and connected apps referenced in each answer. This prototype includes a lightweight version — a timestamp and data scope label — to test whether that level of attribution is useful or adds noise. If reviewers want more precision, a detail mode could surface confidence scores as a secondary option.
Design principle — prioritising non-technical clarity over precision
Questions for the team
Things to explore and validate