In the space of five days, two of the largest AI suppliers moved the AI agent out of one person’s chat window and into the team’s shared workspace. Neither said who is supposed to check its work.
That second point is the story. It is also the part that falls to a manager rather than to IT.
What Microsoft announced
On 25 September Microsoft announced a redesigned Copilot, the assistant built into the Office software most organisations already pay for. It has three parts.
Home brings together Chat and Cowork. Chat is the familiar assistant that answers questions. Cowork, in Microsoft’s description, is where you delegate entire tasks rather than ask for help with them.
Code lets people build apps and automations by describing them in plain language. Microsoft says it is designed “to enable non-developers to securely create their own solutions within their organization’s environment.”
Autopilot is a persistent agent. Microsoft describes it as able to keep working on tasks independently, including when the person who set it going is away from their computer.
Home and Code are rolling out gradually to Microsoft’s Frontier early-access programme over the coming weeks. Autopilot moves to a private preview at the end of the month. The announcement does not say which business plans will get what, or when.
It also says nothing about admin controls, approval steps or oversight. Not a sentence.
What OpenAI showed four days later
At its DevDay event on 29 September, OpenAI showed Dots: AI agents that work alongside a team in a shared space called ChatGPT Space.
OpenAI’s own pages could not be read for this post, so these details come from others. Simon Willison, who live-blogged the event, recorded OpenAI’s framing: “You also need a place for your team to collaborate with their Dots”. The Rundown reports that Dots “come included with Pro and Business Premium plans”.
Treat both as reported rather than confirmed. Nothing read about the launch describes admin controls, an audit trail, or who inside a business account can assign work to a Dot.
Why the missing detail matters
A personal assistant has an obvious owner: the person typing. If it gets something wrong, they usually see it before anyone else does.
An agent in a shared space is different. Several people can hand it work. It can keep going after the person who started it has gone home. What it produces lands where the whole team can see it and use it. The question of who checks it no longer answers itself.
The word “Business” in a plan name matters too. It means these tools may arrive inside company accounts before the department using them has decided anything at all.
That is not a criticism of either supplier. Launch announcements describe features. The rules for how a particular team uses those features were never the supplier’s to write, and they will not arrive in an update.
The quieter change: anyone can build
Autopilot will get the attention, but Code may matter more to a department head. If anyone on the team can describe an automation and have it built, the team will soon have automations that nobody else knows about.
Each one will be small and sensible: a report that assembles itself, a reminder that sends itself, a spreadsheet that fills itself in. Together they become a set of processes with no documentation, owned by someone who may move to another role next year. That is not a reason to stop people building. It is a reason to keep a list, and to ask the same questions of each automation that you would ask of an agent.
This is delegation, and you already know how to do it
None of this needs a new discipline. Handing work to an agent that runs unattended is delegation, and managers already know what good delegation requires.
You decide which work can be handed over and which cannot. You decide who looks at the result before it goes anywhere that matters. And you are clear about who answers for it if it goes wrong, which is never the person who did the work on their first week.
The difference is speed and volume. A new starter produces a few pieces of work a day and asks questions when they are unsure. An agent can produce a great deal overnight and will not ask. That makes the decisions above more important, not less.
Three lines to write before the agent arrives
Which tasks a person may hand to it unattended. Not “anything routine”. A short list. Drafting a weekly summary from a shared folder might be on it. Sending anything to a customer probably is not, yet.
Who reviews what it produces, and when. Work an agent finishes overnight still needs someone to read it before it is used. Put a name and a time against that, the same way you would for a new starter’s first month.
Who answers for it when it is wrong. Not the supplier, and not “the system”. The person whose name is on the work. If nobody is willing to be that person for a task, the task is not ready to be delegated.
None of these needs technical knowledge, and none needs the tools to have arrived. They are management decisions, and they are far easier to make before the first agent is switched on than after the first mistake.
Being straight about the sources
Microsoft’s description comes from its own announcement of 25 September, read in full. The OpenAI details come from Simon Willison’s live blog of DevDay and from The Rundown’s coverage, because OpenAI’s own pages could not be reached. Both companies sell the products described, and every capability claim here is theirs.
The tools are arriving in the team’s shared space. Who checks their work was never going to be in the announcement. It has to be in your plan.