AI Add-Ons

AI Add-Ons

Use these when the app can guide, summarize, or automate.

The menu

• AI assistant drawer
• Ask-this-page chat
• Explain this metric
• Summarize selected rows
• Generate report
• Recommend next action
• Natural language filters
• AI search
• Prompt history
• AI-generated chart insight
• Tool call timeline
• Confidence or source display
• Human approval step

The prompt

Add an AI assistant experience to this dashboard.

The assistant should:
- Understand the current page context (route, visible data, current filters)
- Summarize visible data on demand ("summarize this view")
- Explain a metric in plain language ("why did revenue drop?")
- Suggest next actions tied to what's on screen
- Let users ask follow-up questions in a conversation
- Show its work via a tool-call timeline (which queries it ran, which sources it used)
- Cite sources for any claim, with hover preview
- Require human approval before taking an action that touches the world

UI requirements:
- Right-side drawer that toggles open/closed
- Persists across pages (assistant remembers context within a session)
- Empty state with 3 example prompts based on the current page
- Streaming response area with proper status indication
- Message actions: copy, regenerate, thumbs up/down

Use Vercel AI Elements where they fit. Use ai-sdk for streaming.

Design the UI first, then implement the smallest useful version. Verify
streaming works on slow connections (test with throttled network).

The dual-mode assistant pattern

Most great AI surfaces have two modes:

1. Inline AI. Subtle, embedded in the existing UI. “Explain this metric” links on KPIs. Auto-summary banners. Smart defaults.
2. Conversational AI. A chat panel where the user can dig deeper.

The inline version is the daily driver. The chat version is the escape hatch.

What “human approval step” means

Anytime the AI is about to do something — send an email, run a refund, change a setting — show what it’s about to do and require explicit approval.

Pattern:

"I'm about to [send this email / refund $X / publish this draft]. Continue? [Approve] [Cancel]"

This is a security pattern as much as a UX one. The AI should never act on its own interpretation of the user’s intent without confirmation for non-trivial actions.

When NOT to add AI

• Your data isn’t worth summarizing (dashboards with 3 metrics).
• Users explicitly don’t want it (regulated industries, audit trails).
• You can’t ground the AI in reliable data (it’ll hallucinate).

If your data is messy or sparse, fix that first. AI on bad data multiplies the problem.