For two years, using AI at work meant opening a window and typing a question. That model ended quietly this month. The assistant is now running in the background, reading what you read, and forming opinions about you.
Three signals from the past week point in the same direction. Each one is small on its own. Together, they describe a new layer sitting between people and their own work.
Signal one: OpenAI's dots never stops working
OpenAI released dots, its always-on personal assistant, at its developer day in September. In an interview reported by TechCrunch, Alexander Embiricos, who leads OpenAI's enterprise product team, explained how it works. It is available on plans starting at $100 a month, and once connected to calendar, email, computer and other tools, it works on your behalf without being prompted.
The key feature is what OpenAI calls proactive research. On a cadence, the agent reviews what it knows about you and what it can access, then researches how it could be more helpful. It also refuses read-only access to cloud drives, which tells you how much access the product assumes it should have.
Two details stood out to me.
Embiricos said OpenAI is "deep in lots of conversations with payment processors". And he also said he personally never intends to meet with someone else's agent. The person building the product still wants humans on the other side of the table. That is a useful signal about where agent etiquette is heading.
Signal two: Copilot builds a profile from your inbox
Microsoft started rolling out a Copilot update on October 26 for eligible Microsoft 365 work and school accounts, as reported by Forbes. Copilot now uses Outlook emails, Teams chats and meeting transcripts, including ones you never asked it to analyse, to learn your priorities, relationships and preferences.
Users can ask Copilot what it knows about them and request that specific details be forgotten. Turning the feature off lives under Settings, Chat Settings, Personalization, then Chat History and Work Insights. Inferred memories are deleted 30 days after you disable it, and they can return if you switch it back on within that window.
The detail that matters for companies is this one.
Microsoft says admins can search users' Copilot memory through Microsoft Purview eDiscovery or Content Search. So the profile Copilot infers about an employee is discoverable, which means it is a record. It can be requested in a dispute, reviewed in an investigation and retained under policy.
I wrote last week about Microsoft's view that AI agents should be supervised like insider threats. This is the same company now generating a new kind of employee file.
Signal three: the inbox becomes shared with agents
The third signal is less corporate and more telling. In Ben's Bites this week, Ben Tossell described running his newsletter with a set of agents that sort his saved links and prepare the issue. The same edition flagged Resend Inboxes, a shared inbox that people and their agents work from together, and AgentReady, an open specification for making products usable by AI agents.
None of these is a huge launch. But they show what early adopters assume as the default: the agent has a seat in the inbox, the same way a human assistant would. Small businesses will adopt this first, because the productivity gain is immediate and nobody in a ten-person company writes a data policy before trying a new tool.
What connects them
All three move the same boundary. The question used to be what data you chose to paste into an AI tool. Now it is what data the tool can reach on its own, and what it concludes from it.
That second part is new. An inference about who you work closely with, which clients you prioritise or what you seem worried about is not data you created. It is data the assistant created about you. Most privacy and retention policies were written for documents and messages, not for conclusions drawn by software.
There is also a commercial angle. Persistent agents use far more compute than a chat window, which is part of why vendors are pushing them. The more context the agent holds, the harder it becomes to switch, because the memory does not travel with you. I made a similar point about how your AI stopped being a session, and this is the next step.
What I would put in policy this month
First, decide which assistants can have background access to email, chat and meeting transcripts, and for which roles. Leadership, legal and HR conversations deserve different defaults from a marketing coordinator's inbox.
Second, treat inferred memory as a record. Ask your Microsoft 365 admin how Copilot memory is retained and who can search it. If the answer is unclear, that is your answer.
Third, write down how your company deals with external agents. Will you accept meeting requests from a client's agent? Will your team's agents send email on their behalf? Decide before the first awkward case, not after.
If you want this built properly inside your own systems, with access rules designed in from the start, difrnt.ai builds custom AI solutions end-to-end. Either way, the decision needs an owner.
The assistant has moved into the building. Somebody should decide which rooms it gets keys to.