ChatGPT Work Links Analytics, Computer Use, and Code Fixes
OpenAI’s Dominik Kundel presents ChatGPT Work as a mode for delegating work across connected apps, files, local projects, browser sessions, computer use and Codex, rather than as another chat interface. The walkthrough’s central claim is that ChatGPT can carry context through a full work chain: preparing recurring briefings, generating launch materials, analyzing product data, publishing a dashboard and, in the sharpest example, tracing a Mexico onboarding anomaly to a phone-validation bug and preparing the code fix for review.
The sharpest demo is a full loop from metric anomaly to code fix
Dominik Kundel presents ChatGPT Work as a new ChatGPT mode across web, desktop, and mobile: one place to delegate work across connected apps, scheduled routines, local files, local projects, published sites, the browser, the computer, and Codex. The most concrete demonstration is not a chat response. It is a work loop: launch analytics surface a country-specific problem, ChatGPT investigates the product experience in a simulator, identifies the root cause, edits code, and prepares the change for review.
The fictional company is Hudson Bank, a neobank launching a mobile app. After the launch, the metrics, user feedback, team conversations, and launch plans are treated as scattered inputs that ChatGPT can bring together. Because the request is a data question, Kundel says ChatGPT “automatically brought in the right skills and workflows.” The output includes activation rates, launch signups by region, and an onboarding funnel with 21,300 launch signups, 20,380 phone starts, 18,330 verification completions, 15,342 profile completions, and 10,384 identity verifications.
The launch looks healthy overall, but Mexico stands out. ChatGPT flags the issue directly: “Mexico warrants investigation: activation is 18.8%, and invalid phone-format handling affects 49.4% of Mexican signups.” A country table then shows Mexico with 3,400 signups, an 18.8% activation rate, 44.4% verification completion, and 72.2% phone validation failures.
| Country | Signups | Activation rate | Verification completion | Phone validation failures |
|---|---|---|---|---|
| Mexico | 3,400 | 18.8% | 44.4% | 72.2% |
To confirm what customers are experiencing, Kundel switches to Codex inside ChatGPT and uses an iPhone simulator. He takes an “appshot” by pressing Command and Command, describing it as “a smarter screenshot” that gives ChatGPT the app on screen and the context it needs to act. The prompt asks ChatGPT to review phone onboarding in Mexico, compare it with Germany, and click through the main screens.
ChatGPT then triggers background computer use. Kundel points out that the moving cursor is not his: “This is ChatGPT using my app.” When there is no dedicated plugin, he says, ChatGPT can use the computer directly; the same applies in the browser through a Chrome extension for tasks such as copying and pasting data, browsing sites, or checking a workflow.
The investigation confirms that customers are getting stuck when they enter a valid Mexican phone number. ChatGPT’s reasoning log identifies the root cause as isolated: Mexico is the only country whose validator always returns false, even though formatting and length limits already support 10 digits. Kundel asks ChatGPT to prepare a fix for the phone number validation bug in Mexico.
| Step | Evidence shown |
|---|---|
| Data anomaly | Mexico activation is 18.8%; phone validation failures are 72.2%. |
| Simulator investigation | Valid Mexican phone numbers are rejected during onboarding. |
| Root cause | The Mexico validator always returns false despite 10-digit formatting support. |
| Code fix | The Mexico branch changes from `return false` to `return count == 10`. |
The code diff shown in hudson_phone_validation_main.swift is small but decisive:
case .mexico:
- return false
+ return count == 10
Once the fix is ready, Kundel says he can review the diff inside Codex and publish the pull request so the team can ship a new version of the app.
Work starts as connected context, then becomes recurring responsibility
The basic ChatGPT Work interface keeps the familiar ChatGPT shape but adds a switch from “Chat” to “Work.” In Work, Kundel gives ChatGPT an operating brief rather than a question: “Take a look at my calendar, Slack, and Drive, and help me prepare for my day.” The resulting daily preparation view is organized around “Today’s 2 priorities,” “Rest of today,” “Tasks for you,” and “Worth skimming.” Kundel says it found the relevant notes, meetings, and preparation material for the day.
That daily brief then becomes recurring work. By voice, Kundel asks ChatGPT to act as his “chief of staff,” deliver a briefing every morning, and include World Cup scores for Germany and other matches he cares about while the tournament is going on. ChatGPT confirms the schedule: “Done — you’ll get a chief-of-staff briefing every morning at 8:00 AM, covering calendar, Slack, Drive, priorities, preparation, and recommended order.”
The same work thread appears on mobile, where the daily brief is available away from the desk. For local files and projects, Kundel moves to the desktop app. Its sidebar shows projects and active work, including “openai,” “developers-website,” “Hudson Bank,” and “Codex Launch.” Kundel’s point about the desktop app is local context: “I can give it the context, hand off the work, and then come back when it needs me or has something ready.”
Plugins are presented as the connective layer. Some connect ChatGPT Work to applications; others bundle role-specific workflows for sales, data science, marketing, and more. The assistant’s capabilities change with the job, without requiring the user to leave the main work surface.
Brand inputs become campaign assets, then a review deck
The marketing example starts with Hudson Bank preparing to launch its mobile app. The launch site carries the line “Built for wherever you’re going next.” The visuals are described as generated with ChatGPT and image generation from a logo and brand guidelines. Kundel emphasizes the combination of range and consistency: many creative directions, all still aligned with the brand.
The campaign work then moves from brand inputs into ad concepts. ChatGPT creates mood boards for a mobile launch campaign, including a “Hudson Takeover” concept and a Piccadilly Circus placement. Kundel says he can imagine many ideas in less time than it would take to explore one manually, then asks ChatGPT to turn the concepts into a deck for the launch review meeting.
The generated presentation is a campaign review deck for “Hudson Bank MOBILE APP LAUNCH,” with “NEW YORK LONDON PARIS” and three campaign territories: “City in Motion,” “Quiet Confidence,” and “Hudson Takeover.” Kundel calls it ready for the launch review meeting. The work moves from brand assets to campaign visuals to mood boards to a shareable presentation without leaving the thread.
A dashboard becomes both a shared artifact and an ongoing task
For the analytics work, Kundel chooses a dashboard rather than a deck because the launch metrics need to stay useful after the immediate review. ChatGPT pulls the metrics into a dashboard running in an in-app browser on localhost, with a “Mobile app launch” view and an activation rate of 37.8%.
He annotates the dashboard by voice: “Add a one line summary on how the launch is going.” The dashboard updates with: “Verification and post-activation engagement are strong overall; activation remains the main opportunity, most notably in Mexico.”
Publishing is handled as another delegated task. Kundel asks ChatGPT to publish the dashboard, share the link in a team channel, and make sure the metrics update every day. The dashboard is then shown on a live ChatGPT team site domain for the Hudson Bank mobile launch.