On July 9, OpenAI launched something called ChatGPT Work. Strip away the announcement language and here's what it is: you give it an outcome, and it goes away — sometimes for hours — and comes back with finished spreadsheets, slide decks, documents, even working web pages.
That's a different product than the one you've been using. For three years, the deal with AI chat was: you ask, it answers, and then you go do the work. It gave you an outline; you built the deck. It suggested a formula; you fixed the spreadsheet. The last step — the actual deliverable — was always yours.
This one keeps the last step. It breaks a big job into smaller ones, works through them across your connected files and apps, and hands back the artifact. And the notable part for a small business: the new desktop app that ships it puts this on every plan, including the free one. This isn't a $200-a-month enterprise feature you can ignore. It's in the tool half your competitors already have open.
It's also not one company's experiment. Anthropic shipped the same category with Claude's Cowork mode; Notion, Google, and Microsoft are all converging on it. When every major vendor lands on the same shape at the same time, that's not a feature — that's where the tools are going.
What changes for you
The practical shift is in what you ask for. Most business owners still use AI like a smart colleague you consult: "what should my prices be?", "how do I write this policy?" Consulting is fine, but it leaves the slowest part — production — on your desk.
The new move is to assign deliverables. Not "help me think about my pricing" but "compare my price list against these four competitors and give me a spreadsheet with a recommendation column." Not "give me ideas for the client update" but "draft the client update as a document, using the notes in this folder." Outcome in, artifact out.
Stop grading AI on how good its answers are. Start grading it on whether you'd ship what it hands you.
How to try this without getting burned
Start with work that's recurring and low-stakes. The weekly numbers summary. The monthly competitor check. The first draft of the proposal you write from the same skeleton every time. Recurring work is where you'll actually feel the hours come back, and low-stakes means a bad first run costs you nothing.
Hand it the inputs, not just the request. These agents do their best work when they can see your real files — the spreadsheet, the folder of notes, the past examples. An agent guessing at your business produces generic work. An agent reading your actual numbers produces yours.
Review everything before it leaves the building. Finished-looking is not the same as correct. A polished deck with one wrong number is more dangerous than a rough draft, because you're less likely to check it. The agent produces; you approve. That review step is your job now, and it's not optional.
Write the assignment like you're delegating to a person. What the outcome is, what inputs to use, what format you want, what "done" looks like. The quality of what comes back tracks the quality of the brief almost exactly — we wrote about why the brief is the real skill if you want to go deeper.
The honest caveat
Agents that run for hours can also be wrong for hours. Early users report the same pattern every agent launch produces: impressive on structured, well-fed tasks; shaky when the goal is vague or the data is messy. The cost of experimenting keeps falling — the models behind this now start around a dollar per million tokens — but the cost of shipping unreviewed work stays exactly what it's always been.
So the takeaway isn't "hand your business to an agent." It's this: pick one recurring deliverable this week — a report, a summary, a first draft — and assign it end to end instead of asking questions about it. Look at what comes back. That one experiment will tell you more about where this fits in your business than any launch post, including this one.