The AI coworker: when the assistant does the whole job

Why it matters

You have spent this course learning to prompt, chain and judge. All of that assumes you are in the loop: you ask, it answers, you check. The new products blurring into view in 2026 change that shape. You hand over a whole task, a deliverable rather than a question, and the assistant works across your files, inbox, calendar and connected tools until it has something finished, often while you are doing something else. This lesson teaches the category so you recognise it under any brand name, and the questions that keep it safe at work.

From answering to delivering

You already met the idea of an agent in the last lesson: an agentic system works towards a goal over several steps, decides what to do, and uses tools rather than only producing words. An "AI coworker" is that idea taken to its natural end for office work. You do not ask it a question and get a reply to check. You hand it a job, "produce this month's customer-feedback summary", and it goes away, reads the sources, drafts the thing, and comes back with a deliverable. Some of these run in the background in the vendor's cloud, so the work continues after you close your laptop.

The names you will hear, as reported and as announced by each vendor, are a good map of the category. Treat the specifics below as a snapshot of a fast-moving field, not fixed fact:

• ChatGPT Work (OpenAI, reported launched 9 July 2026): pitched as an agent for longer, multi-step deliverable tasks, and reported to be part of a new desktop app that merges Chat, Work and Codex into one place.
• Microsoft Copilot Cowork (reported generally available 16 June 2026): give it a task, get back a completed deliverable. Reported to run in a secure cloud environment so work continues with your laptop closed, to be admin-enabled rather than on by default, and to be consumption-billed through "Copilot Credits", with around thirteen built-in skills.
• Claude Cowork (Anthropic; reported desktop first, then web and mobile from 7 July 2026): you hand it a task and it works across your files, calendar, email and connected tools until the job is done.
• Gemini Spark (Google, announced at I/O in May 2026): described as an always-on personal agent that executes goals in the background. Its "Daily Brief" is the same idea in digest form: a standing job that produces a summary on a schedule.

One more is worth noting as direction of travel rather than a product to use today: Microsoft Scout, shown as a preview at Build 2026, points at where this is heading. Do not build a process on a preview.

Coworker, chat, and Agent Mode: three different things

It is easy to lump these together. Keeping them separate is what makes you sound like you know what you are talking about.

A chat gives you an answer. You stay in the loop the whole time, one turn at a time, and nothing happens in the world unless you act on the reply. An Agent-Mode run (the supervised agent behaviour you have seen inside a single tool) works over several steps towards a goal, but you are watching: it is one session, in front of you, and you approve or stop it as it goes. An AI coworker is handed a deliverable and works largely on its own, often in the background and often across many tools, until it has finished. The differences that matter are three: deliverable versus answer, background versus supervised, and, crucially at work, the billing model. Because a coworker can run for a long time and use a lot of compute, several are billed by consumption rather than a flat seat, so an ill-scoped task can quietly cost real money.

The governance questions to ask first

Because a coworker reads widely and acts on its own, the last lesson's risk rule applies with more force: it is only as trustworthy as everything it reads, and now it can act on what it reads without you watching. Before you let one touch real work, ask four plain questions.

• Who enabled it? Several of these are admin-controlled and off by default. If it is on, someone made that decision, and there may be conditions attached. If it is not, that is your answer for now.
• What can it see? A coworker with access to your whole mailbox, drive and calendar has a very large reading surface, which is exactly where hidden instructions and confidential data both live.
• What does a run cost? On consumption billing, "just try it on the big one" is a financial decision, not a free experiment. Know the unit before you start.
• Where do the outputs land? A deliverable that appears in a shared drive, or an email that sits in your outbox ready to send, has different consequences from a draft in a private scratch space. Know where the work comes out before you set it going.

When a boring flow still wins

An AI coworker is not the right tool for every repeatable job, and reaching for one out of novelty is a common mistake. If a task is well defined and predictable, "when a row is added to the feedback sheet, post it to the team channel", a plain automation like a Power Automate flow or a Zap does it more cheaply, more predictably and more transparently than an agent. The flow does the same thing every time, costs a fixed amount, and fails loudly. An agent is worth its unpredictability only when the task really needs judgement across messy, changing inputs. Match the tool to the task: boring and fixed, use a flow; varied and judgement-heavy, consider a coworker, scoped tightly.

Here is the shape of handing over a real deliverable, drawn tightly so a human still owns the irreversible step.

Produce the June customer-feedback summary for Fernway's operations team. Read only these sources: the shared feedback log for June, last month's summary deck, and the three tagged threads in the team inbox. Draft a two-page summary and a five-slide deck in the same structure as last month. Leave both as drafts in my private workspace. Do not send, share, or post anything, and do not change any source file. Flag anything you were unsure about at the end so Maya can check it before it goes anywhere.

Example prompts

Before you begin, estimate how large this task is: which sources you will read, roughly how many steps it will take, and anything that would make it run long or use a lot of compute. List that back to me and wait for my go-ahead before doing any work.

In plain English, list exactly what this AI coworker can currently read (which mailboxes, drives, calendars and connected apps) and what actions it can take on my behalf. Flag separately anything that lets it send, share, delete, or change who has access. I want to check its reach matches the job before I hand it real work.

Try it now

Common mistakes

• Treating "coworker" as just a fancier chat. The difference is that it acts, in the background, across your tools. That is more useful and more consequential. Scope it like it can do real things, because it can.
• Ignoring the meter. On consumption billing, an open-ended task on a large dataset is a spending decision. "Just run it and see" is how a quiet bill appears.
• Handing over a task a plain flow would do. If the job is fixed and predictable, an agent adds cost and unpredictability for no gain. Save the coworker for work that really needs judgement.
• Building on a preview. A feature shown at a conference is a signal of direction, not a tool to put a real process on. Wait for it to be generally available and enabled for you.
• Over-trusting the finished deliverable because it looks done. This is the sharpest version of the over-trust trap. A coworker returns something polished and complete, produced with no one watching the middle steps, which makes it feel more authoritative, not less. But "it came back finished" is not "it is right". You did not see the sources it chose or the steps it took, so anything load-bearing needs checking before it leaves your hands, especially anything you cannot undo. The more autonomous the tool, the more scrutiny the output deserves.

Keeping current

This is the most fast-moving topic on the whole site, and every specific in this lesson (names, dates, billing details, which platform a feature lives on) may have changed by the time you read it. Treat the product details above as reported at a moment in time, not as durable fact, and verify anything load-bearing against the vendor's own current pages before you rely on it. What will not date is the category itself and the four governance questions: who enabled it, what it can see, what a run costs, and where the outputs land. For the current picture, check each vendor's own announcements: OpenAI's news, Microsoft's Copilot blog, Anthropic's news, and the Google blog. Accurate as reported on 14 July 2026.