AI Can Now Do Your Work While You're in a Meeting. Here's What Decides Whether It's Any Good.

As of July 9, you can hand ChatGPT a project brief, connect it to your Slack workspace and Google Drive, and come back hours later to a finished spreadsheet, slide deck, or document. Not a draft to fix — a deliverable, produced while you were doing something else.

ChatGPT Work, available now to paid subscribers, is the first mainstream tool that turns AI from a typing assistant into a project executor. The question is not whether to use it. The question is: do you know how to hand something off well enough for this to go right?

What happened this week

Three things converged that prove the same principle.

OpenAI launched ChatGPT Work on July 9 — an autonomous agent that connects to Slack, Google Drive, and SharePoint, breaks projects into steps, and delivers finished output after running unattended for hours (OpenAI, July 9). The same week, Simon Willison documented the Bun runtime's complete rewrite from Zig to Rust using 64 concurrent Claude agents: 6,502 commits, 1 million lines of code, in 11 days, with the existing test suite serving as automated verification (Willison, July 8). And Anthropic published the Alberta Government case study: 50 autonomous agents scanned 466 million lines of legacy code in 20 hours — work that would have taken 6.5 years manually — with human approval required before any deployment (Anthropic, July 6).

Same week. Three different scales. Same result: it worked because someone defined what "done" looked like before the work started.

The Bun rewrite did not succeed because the prompts were clever. It succeeded because tests caught errors automatically. Alberta did not succeed because the agents were trusted unconditionally. It succeeded because humans reviewed output before it went anywhere.

What changes about your Monday

Three things are now worth treating differently:

Delegation is the frontier skill. Writing a good brief for a human — scope, goal, constraints, what success looks like — is now the same act as writing a good brief for an AI agent. The professionals who are already good at handing work off to colleagues will have an immediate advantage. Those who hand off vague instructions will get vague output back, faster and at greater scale.

You can only automate what you can verify. OpenAI's own safety guidance for ChatGPT Work warns that the system can produce "finished mistakes at a scale that is harder to catch" (OpenAI, July 9). The Bun rewrite worked because tests caught errors. If you cannot define how to check the output — specific criteria, not "it looks right" — you are not ready to delegate it.

Specification beats prompting. Ethan Mollick's Wharton GAIL research, published July 7, confirmed that prompt tricks produce zero measurable gains. Chain-of-thought phrasing, tipping, threatening, expert personas — none of it shows up in output quality. What works is giving AI clear goals, constraints, output shape, and acceptance criteria. That is a writing discipline, not a technical one.

Priya: the operations manager who got a wrong answer that looked right

Priya runs operations at a 90-person logistics company. She manages the weekly carrier performance report — pulling data from three spreadsheets, formatting it into a summary, distributing it to four directors every Friday.

She connected ChatGPT Work to her Google Drive and tasked it: "Build the carrier report from this week's data." Forty minutes later it returned a formatted document. The layout was correct. Two of the three data sources were pulled accurately. The on-time delivery rate — the number the directors watch most closely — was calculated from the wrong column, because the spreadsheet structure had changed two weeks earlier.

The document looked finished. It was wrong in one specific way she would have caught in thirty seconds if she had written down what she planned to check when the output came back. She had not. She sent it before looking closely.

The fix is straightforward: before delegating any multi-step task, write the three specific things that would make you reject the output. Not "check it" — the named criteria that determine whether it is right or wrong. That list is what the Bun team had in their test suite. Alberta had it in their human review protocol.

Marcus: the analyst who briefed it like a client project

Marcus is a junior analyst at a 15-person strategy consultancy. He builds competitive landscape documents — each one takes half a day to research, structure, and draft.

Before using ChatGPT Work, he wrote a brief the way he would for a junior colleague: three named competitors, four specific dimensions (pricing model, target segment, key differentiator, recent product changes), a defined output format, and a note that sources must be cited. When the document came back, he reviewed it against his four dimensions line by line before it went to the partner.

The brief took him eighteen minutes. The document came back usable. His review took fifteen minutes instead of the usual two hours.

The difference between Marcus and Priya is not that Marcus used better AI. It is that Marcus wrote down what the output had to contain — and checked against that list, not his general sense that it looked professional.

The one-sentence version

AI can now deliver finished work while you are in a meeting — whether that work is any good depends entirely on how clearly you described what you wanted before you left.

<BlogPracticeSection />