On July 9, OpenAI launched ChatGPT Work: a GPT-5.6-powered agent built to take a goal, work on it unattended for as long as it needs, and hand back finished output — docs, slides, sheets, or working web apps — rather than a conversational answer you still have to act on. It arrived alongside GPT-5.6 itself, and OpenAI merged its Codex coding app directly into the ChatGPT desktop client the same day. The framing matters: this isn't a chat feature with more autonomy bolted on. It's OpenAI explicitly positioning ChatGPT as a system you assign work to, not a system you converse with.
What actually shipped
ChatGPT Work can navigate websites, work with uploaded files, connect to third-party data sources like email and document repositories, fill out forms, and edit spreadsheets, while OpenAI says the user remains in control throughout. It can spend hours on a single project if the task genuinely requires it, rather than returning a quick partial answer and stopping. The headline capability is Scheduled Tasks: the agent can perform an action once, repeat it on a schedule, run in response to an event, or monitor for changes over time using connected apps and a built-in browser. OpenAI's own examples are telling — refreshing a recurring meeting agenda from new Slack updates, checking dashboards each morning and sending a summary of what changed, updating a presentation automatically when new feedback arrives by email. These are exactly the categories of recurring, low-judgment busywork that eat disproportionate time in knowledge work.
Cross-device continuity is part of the pitch too: start a task on a phone, review a draft on the go, check progress on a longer-running workflow between meetings, then pick it back up on the web at a desk. On desktop, the merged Codex-ChatGPT app lets the same agent work across local files and apps or use its built-in browser to reach websites and online tools. Rollout started with Pro, Enterprise, and Edu accounts on web and mobile, with Plus and Business access following in subsequent days.
Why "finishes the job" is a bigger shift than it sounds
Most AI assistants to date have optimized for turnaround speed on a single prompt-response exchange. ChatGPT Work optimizes for something different: task completion over an extended, potentially multi-hour window, with the agent making its own intermediate decisions about how to get there. That's a meaningfully different reliability bar. A chat response that's 80% right is easy to spot-check and correct in seconds. A multi-hour autonomous task that's 80% right might have already sent an email, edited a shared document, or filled out a form with a subtly wrong number — and the review burden shifts from "does this answer look right" to "can I reconstruct everything this agent touched and verify each step." Teams adopting this class of tool need review processes built for that shift, not the review habits built around chat-based assistants.
This is not happening in isolation
ChatGPT Work lands in the middle of an accelerating pattern across the industry: AI tools moving from single-turn chat toward autonomous execution loops that run with less human oversight per step. Coding agents have already made this shift this year, taking on larger chunks of implementation work with less turn-by-turn review. ChatGPT Work extends the same pattern to general knowledge work — documents, spreadsheets, scheduled monitoring — rather than just code. If your organization has already had to rework code review and test-suite assumptions around autonomous coding agents, expect a parallel adjustment for document workflows, approval chains, and data access scoping as agent-style tools like ChatGPT Work get adopted more broadly.
How to actually pilot this without losing control
Start with tasks that have a clear, checkable output, not open-ended judgment calls. A recurring status report pulled from a known set of sources is a good pilot task because the output is easy to verify against ground truth. A task requiring subjective prioritization or stakeholder judgment is a poor first pilot, because errors are harder to catch and more consequential if they slip through.
Scope data access before you scope the task. Because ChatGPT Work can connect to email, documents, and third-party apps, the access grant itself is the actual risk surface, not just the task output. Apply the same least-privilege thinking you'd use for a new employee's system access — grant only what a specific task genuinely requires, and audit those grants on a schedule rather than leaving them open indefinitely.
Treat Scheduled Tasks as production automations, not chat conveniences. A task that runs unattended on a recurring schedule deserves the same change-management discipline as a cron job or a scheduled report in any other system: documented ownership, a way to disable it quickly, and monitoring for when it silently starts producing wrong output rather than failing loudly.
Build a verification step into every recurring task, not just the pilot phase. The risk with "finishes the job" automation isn't that it fails obviously — it's that it succeeds most of the time and fails quietly on the occasions it doesn't. Spot-check scheduled outputs on a regular cadence even after a task has been running reliably for weeks, because the failure mode to worry about is drift, not day-one errors.
Map who's accountable when the agent's output is wrong. If a scheduled task sends a stakeholder update, edits a shared document, or fills out a form that feeds a downstream process, someone needs to own the outcome the same way they would if a junior employee had done the work. Assign that ownership explicitly before rollout, not after the first mistake surfaces.
What to actually watch next
The interesting test for ChatGPT Work isn't the launch — it's what happens over the next two or three months as Plus and Business users get access and the range of tasks people assign to it broadens past OpenAI's curated examples. Watch specifically for how organizations handle the access-scoping question, since that's the part most existing IT governance frameworks aren't built for yet. A tool that can independently browse the web, edit files, and act on schedules is a different governance object than a chat interface, even when it's wearing the same product name.
If your organization is evaluating ChatGPT Work or a comparable agent product, the honest first question isn't "what can it do" — OpenAI's examples answer that clearly enough. It's "what's the smallest, most verifiable task we can hand it that would still tell us something real about how it fails," because that's the pilot that actually de-risks the rollout that follows.