Who OpenAI's always-on agents work for
OpenAI launched dots at DevDay on September 29. OpenAI calls them “remarkably capable, always-on agents,” and says the magic is when a dot brings you finished work “sometimes before you even think to ask.”
Snowflake, Databricks, Google and Meta have all made the case that an AI agent is only as useful as what it knows about your company. OpenAI makes it too, and adds the detail we care about most: an agent shouldn’t wait for someone to ask.
What OpenAI built
A dot has its own cloud computer and browser, connects to “over 4,000 apps” through plugins, and OpenAI says it “can work towards your goals 24/7.” You can message your dot in ChatGPT, Slack or Teams. When you aren’t working with it, your dot “looks for ways to help in the background” by reading the apps you’ve connected with read-only tools, which OpenAI calls proactive research. Inside OpenAI, “A bug appears in Slack, and dots immediately start investigating.” Outside it, an early tester’s dot “noticed he’d forgotten to invoice a publication, prepared the invoice, and sent it after his approval.” Dots are rolling out on Pro and Business Premium plans, with a beta for Enterprise.
OpenAI describes a dot as “an extension of you,” and the first dot each person gets works on that person’s behalf. OpenAI is also piloting specialist dots, which a company sets up “with its own identity, credentials, and access to the systems it needs” to take on a defined responsibility. OpenAI is starting with focused enterprise pilots, drawing on its own internal testing in areas like procurement, invoice processing and customer support.
Workspace agents, launched in April as “an evolution of GPTs,” belong to a team rather than one person. Someone describes a workflow, ChatGPT turns it into an agent, and the agent can run on a schedule or join Slack to “pick up requests as they come in.” One of OpenAI’s examples, the Product Feedback Router, “Monitors Slack, support channels, and public forums” and turns what it finds into prioritized tickets. And in ChatGPT Work, launched in July, Scheduled Tasks can run on a schedule, “when an event occurs,” or to “monitor for changes over time.”
So a dot watches its owner’s apps on its owner’s behalf, a workspace agent watches what its builder pointed it at, and a scheduled task watches what someone asked it to.
A new renewal date in a vendor’s email
Imagine an agency with 40 employees. A software vendor emails the office manager: “Your annual plan will now renew on November 1 instead of December 1, at the new rate.” The operations lead had planned to cut seats before December 1, and the finance lead has the renewal in December’s budget.
The office manager’s dot may well notice the vendor’s email during proactive research and flag it. But the dot flags the email to the office manager, because the dot is an extension of the office manager. The operations lead’s dot and the finance lead’s dot each work from the apps their own owners connected. A specialist dot for procurement could take the email on, once the pilots open up and if the vendor’s email reaches it. A workspace agent could handle it too if someone had built one for vendor contract changes, but few teams build agents for atypical messages, choosing to handle them manually instead.
So the operations lead learns the new date only if the office manager passes it on, and some of the people who have to act on the change may never see it.
Who an always-on agent works for
In May we wrote that most AI helps people act on goals they already have, and that we built this+that for the work that arrives from somewhere else. Most of the work in a company arrives from someone else: customers, partners, vendors, colleagues. An agent that starts from its owner’s goals misses most of it.
Dots show that OpenAI agrees an agent should start before anyone asks. Agents that keep working between conversations and answer in Slack and Teams are the right design.
The open question is who the work is for. A personal agent starts from its owner’s goals and its owner’s apps, and so does a scheduled task its owner sets up in ChatGPT Work. A workspace agent or a specialist dot starts from a job someone defined when they set it up, and OpenAI describes specialist dots as taking on “well-defined responsibilities.” The vendor’s email falls between the two. The office manager’s dot may catch it, but on the office manager’s behalf. Nobody at the agency built an agent for vendor emails like it. The people who have to act on the new date aren’t the person who received it, so an agent that starts on its own has to notify everyone affected, not only its owner.
How to automate it
AI reads the agency’s messages on the team’s behalf, notices that the vendor’s email says the renewal date has moved, and proposes the updates in the system the team works from: a task for the operations lead to cut seats before November 1, and the new date on the vendor’s page where the finance lead looks for it, each linked to the email it came from. Whoever on the team is closest to the vendor confirms them, and from then on the people and AI assistants that read from those places get the same date.
That’s what we’re building at this+that. DoBox reads the messages a company already gets across Gmail, Outlook, Slack, Microsoft Teams, and Google Chat, and turns them into tasks visible to everyone who needs to see them, then hands each to a workflow that handles it or to the person who should handle it. Our brain holds what the company knows, kept current today by workflows that write back to it as they run, with a person approving or editing each update.
Try it on last month’s email. Find a message from a vendor or a customer that announced a new date, and check whether everyone who had to act on the new date knew about it. If not, see how DoBox can help.