AI now delivers finished work. How to review what you didn't build.

OpenAI's new ChatGPT Work does not hand you a draft to finish. It works for hours, pulls from the apps your firm already uses, and hands back the finished job: a built spreadsheet, a full set of slides, a drafted report.

That is a real change in how the work gets done, and it moves the skill that matters. When an agent builds the whole thing, prompting it well is no longer the hard part. The hard part is reviewing work that nobody on your team actually built.

A finished deliverable needs a different check from a draft you shaped yourself. You never watched it being made, so you cannot feel where it is thin. Checking it well means testing three things an agent can get wrong while the result still looks perfect. Look at the inputs it used, the assumptions it built on, and whether it did the job you actually asked.

What changed this week

ChatGPT Work went live on 9 July, powered by OpenAI's latest model, GPT-5.6. OpenAI describes it as taking an outcome, gathering what it needs from your connected apps, breaking the job into steps, and staying with it for hours. What comes back is finished material rather than chat: a spreadsheet, a set of slides, a document, even a working website.

It is the direct answer to Anthropic's Claude Cowork, which has done much the same since January. The same capability is turning up inside tools your staff already pay for, including Microsoft 365 Copilot. So for most firms this is not a product to go and buy. It will simply appear in the software already on the desk.

Whether to switch such an agent on, and what it costs by the task, is its own decision. This piece is about the moment after: it is on, it has produced something, and the work is now in front of you.

Why a finished deliverable is harder to check than a draft

With a draft, you were the maker. You gathered the sources, chose the structure and wrote the argument. When you reread it, you can feel the soft spots: the figure you were unsure of, the section you rushed.

An agent hands you a finished piece of work you did not assemble. It looks complete, and complete is the problem. The rough edges that used to signal a rushed draft are gone: the gaps, the hedged sentence, the note to fill something in later. A confident, well-formatted deliverable can rest on a misread brief or a wrong number, and nothing on the surface tells you which.

This is why reviewing is now the expensive part. Ethan Mollick, who studies how people work with AI, makes the point plainly. The model produces the work in minutes. A person still has to check it, and the checking is where the time goes.

In one set of expert tasks he cites, the AI produced results in minutes while the experts needed about an hour each to check them. The work got faster. The judgement did not.

Three checks for work you didn't build

So the review has to change shape. Reading the deliverable top to bottom and fixing what reads wrong is proofreading, and it will miss the errors that matter. Three checks catch more. Do them in this order.

1. Check the inputs, not just the output. An agent that pulls from your connected apps chose which files, figures and versions to use, and it will not always choose well. Before you read the output, ask what it drew on: this quarter's numbers or last quarter's, the signed contract or an old draft, the right client's data. A polished report built on the wrong source is still wrong, in a way that looks right.

2. Check the assumptions and the logic, not the surface. In a spreadsheet, that means the formulas and the assumptions behind them, not the tidy formatting. In a report, it means the reasoning that leads to the recommendation, not the fluent prose around it. This is the layer proofreading never reaches, and it is where a confident deliverable does the most damage.

3. Confirm it did the job you asked, not a near neighbour. An agent optimises for something that looks finished, which is not always the thing you needed. Check the scope and the question: did it answer what the client asked, cover the period you meant, and stop where your authority to decide begins. A polished answer to the wrong question is the most common way these tools waste a morning.

Who signs off the finished work

None of this replaces the fact-check. When AI puts a specific citation, figure or name in front of a client, that still needs verifying against a real source. That discipline is how you stop AI mistakes reaching your clients. The three checks above sit on top of it, for the case the fact-check was never built for: a whole deliverable the firm did not assemble.

Keep the sign-off human and proportional. A named person still puts their name to anything that leaves the firm, because responsibility does not pass to a tool. Run the full three checks where a mistake is expensive: a client deliverable, a board paper, a set of diligence numbers. Go lighter on low-stakes internal work, and spend your attention where an error would actually cost you.

The firms that get value from these agents will be the ones that know what good looks like and can check for it quickly. That is a skill you can build.

Our 1-to-1 AI coaching for leaders works on exactly this: turning a vague sense that something looks off into a fast, repeatable review your team runs before anything reaches a client. If that would help, book a discovery call.

The next step is small. Take the last thing an AI built for your firm: a deck, a model, a summary. Check its inputs, its assumptions, and whether it answered the question you actually asked. Where you find a gap is where your review needs to live.

Common questions

How is ChatGPT Work different from normal ChatGPT?

It is an agent, not a chat. It takes a whole task and works across your connected apps for as long as it needs. What it returns is a finished deliverable such as a spreadsheet, deck or document, rather than a reply you then turn into work yourself. Anthropic's Claude Cowork and Microsoft 365 Copilot now do much the same, so this is a category, not a single product.

Do we have to check everything an AI agent produces?

No. Match the check to the stakes. Run a full review on anything that reaches a client, a board or a regulator, and go lighter on low-stakes internal work. The aim is to spend your attention where a mistake would be expensive, not to re-read everything.

How is checking a finished AI deliverable different from proofreading?

Proofreading reads the surface. A finished AI deliverable needs three deeper checks: the inputs it chose, the assumptions and logic underneath, and whether it did the job you actually asked. Errors in those places can look perfectly polished on the page, which is why a top-to-bottom read misses them.

Is it safe to let an agent work unsupervised for hours?

It can be, when the task is well scoped and the stakes are contained. The longer an agent runs, the further a wrong assumption made early can carry. So say what "done" looks like, keep the job bounded, and check the result before it goes anywhere. Treat a long run as delegation you are accountable for, not automation you can forget.

Sources

• OpenAI, ChatGPT is now a partner for your most ambitious work, 9 July 2026. Source for what ChatGPT Work does: a GPT-5.6 agent that works across connected apps over long runs and returns finished deliverables.
• Fortune, Anthropic launches Cowork, a file-managing AI agent, 13 January 2026. Source for Claude Cowork as the multi-step work agent ChatGPT Work answers.
• Ethan Mollick, Management as AI superpower, 27 January 2026. Source for reviewing as the time-consuming part of AI work, and for the expert tasks where the AI took minutes and checking took about an hour.