An AI colleague that does the work first
SAI
Summary
The model proposes. A person accepts. Only then does it become work. SAI listens to a meeting, offers the work it needs, and writes only accepted results into the notes.
- 4 stages
- 32
- 2
Overview
The model proposes. A person accepts. Only then does it become work. SAI listens to the meeting, offers the work it needs before anyone asks, and runs only what someone accepted. Only the result reaches the notes.
It was built by a team of three. Jaeyong Seong, Jaeyong Bae and Seungbin Oh; I did the UX flow, the design system and the frontend. It was built for the 2026 Wanted AI Championship.
Four stages run while the room talks. Speech is transcribed live. The conversation and the shared material are read to work out what is needed now. Web research and arithmetic run, and the answer comes back with its evidence. A follow-up, "and if the budget were 600,000 won?", is linked to the earlier discussion and recalculated.
People check an offer at a glance and accept it by hand, without stopping the conversation. They join one meeting from their own laptops and phones in the same room, with a single device listening.
What runs each of the four stages
- Transcription – Amazon Transcribe, writing down speech as it happens.
- Judgement – Claude, reading the conversation and the shared material to pick what is needed now.
- Research – Amazon Nova, searching the web and returning sources with the answer.
- Calculation – code execution, so numbers are computed rather than guessed by a model.
- Frontend – Next.js, React, TypeScript, Tailwind CSS; the same meeting room on desktop and mobile.
Problem
- Looking something up stops the conversation. Checking one number or comparing two options means someone turns to a screen, and the meeting waits.
- Using an AI means explaining the meeting again. To ask a chatbot you first summarise what has been said, and writing that summary stops the meeting a second time.
Key decisions
Notes never write themselves – suggest, accept, then note
Text an AI writes on its own only gives people in a conversation more to check.
The AI asks first, works only once accepted, and writes only the result. Suggestions never mix into the note.
No evidence, no suggestion
People accept only when they can see why it's asking now.
A suggestion without a source utterance or material isn't shown, and at most two wait for an answer at once (design value).
What the AI made, it marks with light
Korea's AI Basic Act requires disclosure, and it had to be clear what the AI wrote.
AI output carries a blue glow and an 'AI-generated' badge.
Material waits for the reader
A number can't be read if it changes or disappears while you read it.
New results queue behind a banner instead of replacing the card in view. Unverified values keep an 'unverified' tag instead of vanishing.
Two more decisions
- One accent: blue. Mid-meeting, the screen gets a glance. Kinds and states differ only in grey value, and only what needs pressing now is blue. There is no red.
- Everything happening now, in one island. Recording state and alerts scattered over the screen pull people out of the conversation. One capsule at the bottom holds the recording state and every alert, and a blocked action shows its reason and a button to fix it.
Process
Design system
- Eight principles – space instead of lines, one blue, AI as light, material waits, say what's unknown, lines only when they carry information, never clip a shadow, official logos as they are.
- Tokens – primitive → semantic → component. A 4px grid and only two shadows.
- Components – suggestion cards (offered, working, done, dismissed), three kinds of material card (evidence, calculation, proposal) whose fill shows confidence, the meeting island, a consent gate, QR sharing.
- Live style guide –
/design-systeminside the app shows the tokens and principles as they run.
Verification
The system was checked by tests, not by eye.
- Tests – 32 Playwright tests on desktop (1440×900) and mobile; any console error fails a test. axe checks WCAG 2.1 AA.
- Contrast – colours the check failed (secondary text 4.49, white on the accent 3.88) were fixed to 5.05 and 4.77 (measured).
- Shortcuts while typing Korean – handled by key code, so they work mid-composition.
Development
Development ran through Codex, from requirements to implementation, tests and code review.
Results
It ran on live models, a real browser and a database. All four stages held together as one flow: transcription, judgement, research and calculation, then the offer. Quotes were compared, a budget change was recalculated from a spoken follow-up, and the chosen items were saved to the notes.
Where a person steps in
- The AI offers first – what it heard that might be useful comes out as a card. An offer is only an offer; it never becomes a note by itself.
- Only what is accepted stays – a card someone takes becomes part of the notes. The meeting is never interrupted, and the decision stays with the room.
- Always with its evidence – each card carries the utterance it came from, and confidence shows as the card's fill. What it doesn't know, it says it doesn't know.
Reflection
What held up
- Nothing reaches the notes before someone accepts it. The demo ran research and recalculation several times, and the notes still held only what a person had chosen.
- The screen while the AI works is the product. If people can't see what the model is doing, the wait is just a frozen screen.
What failed
- I built everything first and cut later. After the whole thing was built, close to 9,000 lines came out across 90 files (output). Deciding what was essential first would have saved those lines.
- I judged contrast by eye. Secondary text at 4.49 and white on the accent at 3.88 looked fine on screen. Both were below WCAG AA, and they were only fixed after the check caught them.
What I would test next
- The two-offer limit. It is a design value. No one has counted how many offers a real meeting can hold.
- The acceptance rate. Which offers get accepted and which get ignored is not measured yet. Without that number there is no basis for changing when an offer appears.