OpenAI's new agent finishes a spreadsheet or slide deck on its own. That's your submittal log and bid tab it's aimed at.

OpenAI launched ChatGPT Work on July 9 — an agent, built on the new GPT-5.6 model family, that takes a single instruction, plans out the steps, works across your files and apps without asking permission at each one, and comes back with a finished spreadsheet, slide deck, document, or basic web app. The examples OpenAI used to demo it — updating a spreadsheet with new data while keeping the existing formatting, turning a screenshot or dashboard into a presentation, drafting a document from source material — are close to a word-for-word description of what a project engineer or document control coordinator does with a submittal log, an OAC meeting deck, or a bid tab.

What did OpenAI actually launch?

ChatGPT Work is a workspace agent layered onto ChatGPT Enterprise, Business, Edu, Pro, and Plus, rolling out starting July 9. It shipped alongside GPT-5.6 — three new model tiers named Sol, Terra, and Luna that we covered on pricing and RFI triage the day before — plus a redesigned desktop app that merges Chat, Work, and Codex into one window. Rollout order: Pro, Enterprise, and Edu got it first on web and mobile; Plus and Business follow "over the next few days," per OpenAI's release. The desktop app has it on every plan already.

The pitch is autonomy, not just chat. Instead of a back-and-forth where you paste in data and ask for edits, you hand the agent a task — "update this cost tracking sheet with the new invoice totals, keep the same tabs and formulas" — and it works the file directly, then hands back a finished version.

Where does this map onto a GC's back office?

None of this is construction-specific — OpenAI didn't build it for the industry. But the actual work being automated is the mechanical layer of jobs that already exist on every GC's staff: the project engineer who reformats the OAC deck every Thursday, the document control coordinator who keeps the submittal log current, the estimator's assistant who refreshes the bid tab when a sub resubmits a number.

Is it accurate enough to trust unsupervised?

Not yet, and OpenAI's own benchmark says so. On SpreadsheetBench — a test built around real-world spreadsheet edits — the agent scores 45.5%, compared to 20.0% for Microsoft's Copilot in Excel. That's more than double Copilot's rate, and it's also a coin-flip-adjacent number to hang a live cost log on. The failure mode that matters most in a bid tab isn't the agent refusing the task — it's a plausible-looking spreadsheet with one wrong cell that nobody catches until the bid is submitted. Independent reporting on ChatGPT's models makes the same point in general: the hallucinations worth worrying about are the ones that read as confident and correct, especially with numbers.

Should a GC or estimator try this now?

If your firm already runs ChatGPT Business or Enterprise, this is worth a narrow pilot — not a rollout. Two things to check first:

1. Enterprise controls. ChatGPT Work includes permissions for which tools and files the agent can touch, plus a Compliance API for monitoring what it did. Confirm those controls are configured before pointing it at anything with contract numbers in it — most AEC document-retention and audit requirements assume a human authored the record.
2. Pick a low-stakes first task. A meeting-minutes draft or a slide deck from dashboard screenshots has a low cost if the agent gets something wrong. A live SOV or certified payroll spreadsheet does not. Start with the former.

The honest read: this is the same office work your admin staff already does, now demonstrably automatable in one shot instead of a chat back-and-forth — but "demonstrably automatable" and "safe to leave unsupervised on a financial document" are two different bars, and ChatGPT Work has only cleared the first one so far.

Forward this to whoever on your team maintains the bid tab template.

Construction AI Brief publishes new analysis three times a week. Subscribe at constructionaibrief.com.