Capture → Bizzy chat

How the owner adds a transaction with AI. The old flow was a staged form; the one we ship now is a conversation with Bizzy — say it, and it's logged.

Before
9:41
After
9:41

Existing screen. A "Quick Capture" form: a dark-teal "AI Powered" header, four co-equal method tiles (Upload / Camera / Voice / Text), a big empty "Nothing added yet" list, and a separate "Run AI Extraction" button. You pick a method, stage items, then trigger a batch — and wait.

Proposed screen. The Bizzy chat. You tell Bizzy what happened in plain language — or snap a receipt — and it logs it and replies at once, ready to correct. Built from Bizzy + ChatBubble.

What changed.

  • A staged form (add sources → Run Extraction) became a conversation with immediate feedback.
  • Four co-equal method tiles became one natural-language message, with camera and voice tucked into the composer as secondary.
  • The dominating "Nothing added yet" empty state is gone — the thread opens with Bizzy telling you exactly what to say.
  • The "Run AI Extraction" jargon and its deferred payoff became an instant, plain-language confirmation you can fix in place.

Why it changed. The old screen made the owner do the AI's job — choose a format, stage inputs, then trigger extraction — and then wait for a batch result with no way to converse or correct it. A chat flips that: the owner speaks naturally, sees the entry the moment it's logged, and fixes it in-line. It also lets the AI show up as a helper (Bizzy), not a machine you feed and crank.

Which rule applies.

  • Conversation over form for AI capture — one natural-language input beats a wall of co-equal method tiles.
  • Immediate feedback over a deferred "Run" — show the result as it happens.
  • Plain language over jargon — "Sold 2 trays of eggs" beats "Run AI Extraction."
  • Make results correctable in place — never a silent batch the owner must dig through.
  • The AI wears Bizzy — a named coach, not an anonymous engine.

User outcome. The owner just says "sold 2 trays of eggs for 7,000" and it's logged — no method-picking, no staging, no waiting on a batch.

Edge case fixed. Bizzy got it wrong. The reply invites a correction ("tap if I got the category or amount wrong"), instead of a silent extraction the owner has to hunt through and fix after the fact.

Reusable rule

For AI capture, prefer a conversation to a form. Let the owner say it in their own words, answer instantly in plain language, keep photo and voice as secondary in the composer, make every result correctable — and let Bizzy front it.