Why
The demo's whole appeal is that it collapses an afternoon of dashboard-hopping into one sentence. That collapse is also the risk: the agent shows its "thinking" as friendly captions — "reviewing company context," "comparing traffic, conversion and revenue" — never as artifacts a second person could check. When it says "root cause: checkout leakage plus an invite-link bug" (00:24–00:33), nothing on screen shows the SQL it ran against Snowflake, the Tableau view it read, or which Teams thread it is quoting. That is the difference between an analyst and a black box that talks like one. Hwang Sok-yong's point about the AI era — that the scarce skill becomes the quality of the question, not the answer — cuts the other way here too: a good question deserves an answer you can re-derive, not just restate.
How it works
The question and the visible "thinking"
The PM's question is open-ended and business-shaped, not a query. The agent responds with a sequence of labeled steps rather than a single black-box answer: reviewing company context (00:15), then comparing traffic, conversion and revenue in Snowflake, cross-checking a Tableau sales dashboard, and reading Dash XV threads in Microsoft Teams (00:16). This is the "agent mode" pattern — an LLM driving connectors across a company's existing tools instead of a bespoke integration per data source.
From a number to a named cause
The agent surfaces one comparison — product-page traffic is the highest of this quarter's launches, but conversion is 1.3% against a 2.5% launch average (00:17) — and then names why: checkout leakage plus an invite-link bug, not just when the drop started but the mechanism behind it (00:24–00:33). Naming a mechanism from a single metric comparison is the strongest claim in the video and the one with the least shown evidence. A gap between traffic and conversion is consistent with dozens of causes; picking one and presenting it as the root cause needs the underlying rows, not just the gap.
One click into a deliverable, then a living dashboard
A single instruction — "add this analysis, the checkout issue and the recommended actions to the Dash XV launch-week Google Slides deck" — updates the deck directly (00:35–00:38). The agent then renders an interactive web dashboard for the team (00:41–00:56); by voice, "show traffic here instead of conversion" re-lays the cards live (00:44–00:48), and the dashboard picks up brand colors, theme and typography automatically (00:49–00:52). The recovery model quoted — 157 customers recovered, 39 remaining (00:28–00:34) — is presented with the same confidence as the traffic number, but it is a projection, not a measurement, and the video never marks that distinction.
Scheduling the recap
The closing instruction, "every Monday morning refresh this dashboard and post key changes to Slack #product-leads" (00:58–01:00), turns a one-off investigation into a standing job: the weekly post carries numbers, a summary and a dashboard link (01:01–01:02). This is the same move as a shipping container's standard interface — one connector-and-schedule combination replaces bespoke integrations to Snowflake, Tableau, Teams, Slack and Slides — and the same trade-off: a standard interface hides what is inside the container, and here it hides which query ran and which rows came back. YC's "own the outcome, not the task" framing applies too: scheduling the post is easy; owning whether next Monday's number is still true is the actual job.
Where it lands in Jayverse
- Auditor: the evidence panel is the deliverable, not the sentence. Every agent finding the Auditor records should carry (source, query, row count, time window) as structured fields, not prose — "checkout leakage" is a claim until the query and the rows it returned are attached.
- Verex: a ledger question needs a block range, the same way a Snowflake claim needs a query. "Why did volume drop" against Verex's order book or settlement history should answer with the block range and the rows it summed, not just a number.
- Knowledge Notes: the weekly Slack post is a scheduled ratio, the same shape as this site's daily history file. A number that reappears every Monday only stays honest if the query behind it is pinned — the same discipline
learning-greed-with-no-time's scoreboard depends on. - Eng: interview question — "how would you make an AI analyst's answer auditable?" The honest answer is this demo's missing half: attach source, query, row count and time window to every claim, and let the schedule re-run the query, not just repeat last week's sentence.
Verified and unverified
Verified on 2026-09-19: OpenAI ships ChatGPT for business with connectors to Google Drive, Slack, SharePoint, GitHub, Snowflake and similar enterprise systems, plus scheduled tasks and agent-mode/"record mode" style features; Snowflake, Tableau, Microsoft Teams, Google Slides and Slack are real, documented integrations in that product family. Taken from the summary and not independently checked: the product name "ChatGPT Work" and "data agent" as a distinct SKU — this is a marketing demo, and the underlying data and all numbers shown (the 1.3%/2.5% conversion figures, the 157/39 recovery model) are synthetic; whether the rendered dashboard is a persisted, shareable artifact or a session-only view; the exact timestamps.
Sources: YouTube — OpenAI, "Data agent in ChatGPT Work" · related items: claude-for-cfos-verify-not-summarize, Tech #102 (MLflow tracing, LLM-as-judge), Tech #106 (harness engineering: schedule + connectors need an eval), yc-pick-one-idea-go-deep, design-constraints-layers-interfaces-choice, hwang-sokyong-read-classics-ai-era (Life), learning-greed-with-no-time.
Key expressions
| Expression | 뜻 · 쓰이는 자리 |
|---|---|
| PM | Product Manager(제품 매니저) · 비즈니스 질문을 던지는 화자. "a PM types" |
| root cause | 근본 원인(증상이 아니라 원인) · 진단의 목표어. "root cause: checkout leakage plus an invite-link bug" |
| black box | 블랙박스(내부를 알 수 없는 시스템) · 검증 불가능한 답을 비판할 때. "a black box that talks like one" |
| checkout leakage | 체크아웃 이탈(결제 단계에서 이용자가 빠져나가는 현상) · 전자상거래 전환율 문제의 이름. "checkout leakage plus an invite-link bug" |
| connector | 커넥터(외부 서비스와 에이전트를 잇는 통합 모듈) · 에이전트가 여러 도구에 접근하는 방식. "connectors across a company's existing tools" |
| scheduled task | 예약 작업(정해진 주기로 자동 실행되는 작업) · Monday 리포트의 메커니즘. "scheduled tasks" |
| agent mode | 에이전트 모드(LLM이 도구를 직접 조작하는 동작 방식) · 이 데모가 시연하는 패턴 이름. "This is the 'agent mode' pattern" |
| SQL | Structured Query Language(구조화 질의어, 데이터베이스에 보내는 질의문) · 근거로 요구되는 산출물. "the SQL it ran against Snowflake" |
| LLM | Large Language Model(대형 언어 모델) · 에이전트를 구동하는 기반 모델. "an LLM driving connectors" |
| SKU | Stock Keeping Unit(재고 관리 단위, 여기서는 제품 라인 구분) · "Data agent"가 별도 상품인지 여부. "as a distinct SKU" |
| dashboard-hopping | 대시보드 순회(여러 대시보드를 오가며 확인하는 일) · 에이전트가 대체하려는 수작업. "an afternoon of dashboard-hopping" |
| cross-tool investigation | 도구 횡단 조사(여러 툴에 걸친 조사) · 이 워크플로의 핵심 동작. "cross-tool investigation, one-click deliverable" |
| one-click deliverable | 원클릭 산출물(한 번의 지시로 완성되는 결과물) · 덱 갱신·대시보드 생성을 가리킴. "one-click deliverable, scheduled recap" |
| standing job | 상시 작업(한 번이 아니라 계속 도는 업무) · 예약된 Slack 게시를 가리킴. "turns a one-off investigation into a standing job" |
| own the outcome | 결과를 책임진다(과제 완료가 아니라 결과의 지속을 책임진다는 뜻) · YC식 프레이밍. "own the outcome, not the task" |
| re-derive | 다시 도출하다(같은 과정을 밟아 같은 결론에 이르다) · 재서술(restate)과 대비. "re-derive, not just restate" |
| structured fields | 구조화된 필드(자유 텍스트가 아니라 정해진 항목으로 저장되는 데이터) · Auditor가 요구하는 기록 형식. "as structured fields, not prose" |
| block range | 블록 범위(블록체인에서 시작·끝 블록 번호로 지정하는 구간) · 원장 질문의 근거 단위. "should answer with the block range" |
| session-only view | 세션 한정 뷰(그 접속 동안만 존재하고 저장되지 않는 화면) · 영구 산출물과 대비. "a session-only view" |