OpenAI Just Gave ChatGPT a Job. Should It Have Yours?

OpenAI has launched ChatGPT Work, an agent mode that promises to take a goal, disappear for a few hours, and come back with a finished spreadsheet, deck or report instead of a chat reply. That's a bold claim for anyone still copy-pasting between five tabs to build a valuation pack. Here's what it actually is, what it's promised to do, and how it stacks up against the two tools you're more likely already using.

What ChatGPT Work actually is

Launched on 9 July, ChatGPT Work sits alongside ordinary ChatGPT as a separate mode, rather than replacing it. You still have Chat for quick questions and drafting. Work is for when the job is bigger than a single prompt can handle.

The mechanics are worth understanding properly, because they explain both the promise and the catch. Here's how a job actually runs:

• You give it an outcome, not an instruction. Something closer to "pull together this month's variation summary" than "write me a paragraph"
• It gathers context first. It pulls from whatever you've connected, Slack, Gmail, Google Drive and a wide directory of other apps, so it isn't working from a blank page
• It shows you a plan before doing anything. The job gets broken into steps so you can see what it intends to do and redirect it if it's heading the wrong way
• It works independently once approved. It can keep going for hours on a single task rather than needing you to babysit each stage, checking in for approval on anything significant along the way
• It hands back a finished artefact. A spreadsheet, a slide deck, a written report, rather than a chat reply you then have to turn into something usable yourself
• It can run on a schedule. Work can repeat a task or trigger when something changes, so in theory a recurring report could be waiting for you every Monday morning without you asking again

A few practical points before you get carried away:

• It's bundled into existing Plus, Pro, Business and Enterprise plans rather than sold separately, so no new line item needed if your practice already has ChatGPT seats
• The rollout is staged: Pro, Pro Lite, Enterprise and Edu users get it first, with Plus and Business following over the following days and weeks
• One of its more ambitious features, publishing a finished output as a live web page ("Sites"), isn't available in the UK, EEA or Switzerland at launch, with no announced timeline for when that changes

Worth checking exactly where your firm's plan sits in that rollout, and what's actually switched on for your seat, before you promise any of this to a colleague.

Where Copilot and Claude fit

ChatGPT Work is app-agnostic, it works across whatever you connect it to. Microsoft and Anthropic have both been building towards the same promise from different starting points.

Microsoft Copilot

• Lives inside Word, Excel, PowerPoint, Teams and Outlook rather than working as a standalone destination
• Grounded in your organisation's own Microsoft Graph data, so it already knows who you work with, which files are live, and what meetings you've had
• Hard to replicate that context elsewhere if your firm lives in the Microsoft stack, and a lot of you do
• Microsoft has started offering Claude as a selectable model inside some Copilot agent experiences, so Copilot and Claude aren't as separate as they used to be (worth verifying against your own tenant, as availability varies by licence and hasn't reached everyone yet)

Claude

• Reaches this through Cowork, an agentic desktop app built for knowledge work rather than coding
• Can pull in tools like Chrome, Excel and PowerPoint to complete a task
• Also offers Projects, for grounding a working AI in a set body of documents (a live index, not a hard file cap, so it holds up as your document set grows)

What to watch out for

Every one of these tools will happily produce a document that looks finished. It isn't finished until a QS has checked it. The RICS Responsible Use of AI standard, mandatory from March next year, requires documented human oversight and dip sampling of AI outputs, and a valuation pack or variation summary produced by an agent working unsupervised for two hours is exactly the kind of output that needs a human pass before it goes anywhere near a client. Treat all three tools as very fast first drafters, not sign-off authorities.

Also worth flagging to your IT or compliance lead before rolling any of these out at practice level, given they're all reaching into connected accounts and files by design.

Try this today

Pick one task you do every month that involves pulling information from more than one place, a variation summary built from a folder of emails and a spec, for example. Run it through whichever of these three you already have access to at work. Time how long it takes you to get from goal to a document you'd actually be happy sending, including the check you do afterwards. That number is the useful one, not the tool's own claims.

How Gather integrates with it

Every agent above is only as good as what it can see. Hand ChatGPT Work, Copilot or Claude a goal like "build this month's valuation pack" and, right now, the first thing any of them will do is ask you for the data, which usually means you exporting site diary entries, progress records and photos out of whatever platform holds them before the agent can even start.

Gather's plugin closes that gap directly:

• It runs on an MCP server, the same connection standard that ChatGPT Work's app directory, Copilot's Graph grounding and Claude's Projects are all built around. We covered what MCP means for quantity surveyors in more detail separately
• That means Gather's site records become a data source these agents can query natively, rather than a spreadsheet you have to prepare first
• Ask Claude, Copilot or ChatGPT to pull together a progress summary or a variation pack, and if Gather's connected, it can go straight to the contemporaneous site record, photos and all
• No more starting from whatever you happened to have exported that week

It's the same underlying shift as the rest of this piece, just pointed at construction data specifically: less time assembling the inputs, more time checking the output before it goes anywhere near a client.