OpenAI just launched dots: always-on AI agents with their own computers, context that builds over time, and the ability to keep working between conversations.
I was part of the early access group. What stood out was being able to keep working on the same project, giving direction and feedback as the work developed.
You can call your dot in ChatGPT, message it through ChatGPT, Slack, or Microsoft Teams, and give it access to the apps it needs for the job.
Think about the people in your organization who have tried AI a few times and never quite found a place for it. An assistant they can talk to, give work to, and come back to later could make the value much easier to see.
The enterprise opportunity starts there. More people able to delegate useful work, with less effort spent managing the AI along the way.
Your dot has a computer and somewhere to do the work
A dot has its own cloud computer and browser, alongside the apps you allow it to use. OpenAI also lets you connect your computer separately, with permission. The setup documentation explains those options.
For a business user, that opens up assignments with several steps: gather information, work through files, run an analysis, create something, and bring it back for review.
Consider a supplier assessment. Someone has to collect the proposals, compare terms, check assumptions, and prepare a recommendation. An agent could help across that sequence, with the right access and assignment. If the question changes halfway through, you can redirect the work.
The appeal is being able to hand over more of that sequence. You spend your attention on the assumptions, tradeoffs, and final decision while the agent helps assemble and revise the work around them.
You can talk to it in the flow of your day
You can start a voice call with your dot in ChatGPT. Supported Slack and Teams connections let you work with it through messages, subject to your organization's setup and permissions.
Imagine a sales leader talking through an account before a meeting. They explain the concern, point to the relevant material, and ask the dot to investigate. Later, they clarify a question in Slack and review what it found.
It feels closer to working with a colleague: explain the situation, let them make progress, and stay in the conversation as questions come up. You still need to check the work and correct misunderstandings.
For a leader trying to expand AI use beyond a small group of power users, that matters. Less setup and less tool switching can make a real assignment a more approachable place to start.
Proactive help is where this gets interesting
Most of us have work we would delegate if we stopped long enough to identify it. A forgotten follow-up. A document that needs another pass after a decision changed.
A persistent agent has an opportunity to notice those things and bring them back into view. Useful context gives it more to work with than the last request you typed.
OpenAI shared an early tester's example: a dot noticed an invoice he had forgotten to issue, prepared it, and sent it after approval. It is a small example with an obvious business consequence. The launch post includes the story.
There is an important boundary here. OpenAI's proactive research is read-only. Its research tools cannot send messages to other people, change app content, or control a computer. Taking action afterward still depends on permissions and the applicable approval rules.
The bigger shift goes beyond OpenAI
Dots are part of an emerging category. Meta describes Muse as an assistant with a cloud computer, persistent context, and proactive suggestions. Grok Bot offers persistent bots that work with browsers, files, and tools. Instinct emphasizes an assistant you can text or call, with proactive follow-through.
Their capabilities and controls differ. The common direction is an assistant that stays involved as your needs develop.
The emerging category
Persistent AI assistants, side by side
Select an assistant to see how its maker describes it.
OpenAI dots
Call in ChatGPT, or message through ChatGPT, Slack, or Microsoft Teams
- Own cloud computer and browser
- Context that builds over time
- Works between conversations
- Read-only proactive research
Meta Muse
Personal AI agent from Meta
- Cloud computer
- Persistent context
- Proactive suggestions
Grok Bot
Persistent bots from xAI
- Persistent bots
- Works with browsers
- Works with files and tools
Instinct
An assistant you can text or call
- Text or call
- Proactive follow-through
The common thread: an assistant that stays involved as your needs develop.
This is starting to feel like Her
I mean the interaction in the film: you talk naturally, return to an earlier topic, and have help available in the flow of your day.
Today, using AI well can still require a lot of work around the work. Which model should I choose? What context does it need? Am I running out of context window? Which tool can actually do the job?
These products are trying to carry more of that burden for the user. You still need to give direction, provide access, and check important results. But more of your attention can stay on what you want to accomplish.
My bet is that smoother interaction will bring more people into regular AI use and surface use cases we are barely considering today. A task that feels too small to set up an AI workflow for might be worth mentioning to an assistant already working with you.
The customer journey can continue past the answer
If you are working on answer engine optimization (AEO) or generative engine optimization (GEO), you are already thinking about how your business appears in AI-generated answers.
Persistent agents add another question: what happens after the recommendation?
A customer could ask an agent to compare suppliers, check availability, configure a service, or prepare a purchase for approval. In that journey, the agent needs to work with the business, carry the customer's constraints through each step, and bring back a result the customer can trust.
For a business leader, that expands the strategy question from being discoverable to being usable by an agent acting for your customer.
Your product needs to work for agents too
Access will be contested. Amazon blocked Meta's Muse from shopping on Amazon.com in September, citing permission and security concerns.
Shopify is opening a supported path. Its new WebMCP checkout support lets browser agents work with supported checkouts, with buyer confirmation before submission and human handoffs when needed.
That points to a tension worth watching: customers may want to delegate the interaction while platforms want control over how that interaction happens. An agent's technical ability to use a website does not settle whether the business will allow it.
I would start reviewing products for both human and agent users:
Agent-first UX review
Is your product ready for agent users?
Check each question your product can answer "yes" to today.
0 of 3 ready
That is what I mean by agent-first UX: make authorized tasks understandable and reliable for the agent, with the customer still able to inspect and control the outcome. A cleaner checkout flow, clear permission boundaries, and an unambiguous confirmation may matter as much as the page that earned the recommendation.
Pick one customer journey and test it. Measure whether the agent can complete the task correctly, where a person has to intervene, and whether the customer can see exactly what happened.
Where I would put dots to work inside a business
I would start with work that spans several conversations, documents, or systems and keeps changing over time. That gives persistence something useful to do.
These are examples I would test, with suitable integrations and approved access. They are not claims about deployed client results.
Help sales teams follow through on important deals
An account owner could ask a dot to track open questions, commitments, and missing information across a defined set of opportunities. As the account develops, new evidence can inform the next brief or proposed follow-up.
I would look at whether the rep spends less time reconstructing the account and whether important follow-ups happen sooner. Customer commitments stay with the account owner.
Keep a launch current as decisions change
A scope change can affect positioning, sales materials, training, and the launch plan. A dot could help the launch owner find those dependencies and prepare updates. Feedback on one draft could inform revisions across the assignment.
The useful result is a launch team spending less time comparing versions and chasing context. You would still check the actual materials before they go out.
Keep customer feedback connected to product work
A product manager could ask a dot to investigate recurring problems across defined feedback sources. It could gather examples, check whether a problem is already addressed, and develop a brief as new evidence arrives.
I would judge this by whether the team gets to a well-supported product decision sooner. A longer list of suggestions is easy to produce. A useful brief should help someone choose what deserves attention.
Decide where the extra capacity should go
From a strategy perspective, I would want a team to answer two questions. Which work could an agent help carry forward? What would we do with the capacity that creates?
A mid-market company with a small operations team might finally work through recurring problems that never reach the top of the list. Decide which result matters before you start counting agent activity.
I would choose one recurring responsibility and give a named person ownership of the trial. That person should provide the context, define what good looks like, and review the output. Persistent context can help the work develop, but important facts still need source checks.
Compare the entire effort with the current process, including supervision, corrections, and tool costs. Then look for a business consequence: a faster response, a better-prepared decision, or important work completed that would otherwise have waited.
Running a dot pilot
Four steps from one responsibility to real evidence
Step 1. Choose one recurring responsibility and decide which result matters before you count agent activity.
Step 2. Give a named person ownership. They provide the context, define what good looks like, and review the output.
Step 3. Compare the entire effort with the current process, including supervision, corrections, and tool costs.
Step 4. Look for a business consequence: a faster response, a better-prepared decision, or important work completed that would otherwise have waited.
That gives you evidence for expanding the assignment and a reason to revisit work you have treated as too time-consuming to do consistently.
What enterprise teams need to know before connecting work
As of September 30, access is still rolling out. Enterprise dots are in beta and off by default until enabled by an admin. Workspace controls and connected-service permissions determine what is available.
Before bringing in company data, agree on the sources, permitted actions, and work that needs review. Check retention too: OpenAI says disconnecting an app does not erase context already obtained.
Before you connect company data
Agree on these five things first
There is also a separate enterprise direction to watch. OpenAI has previewed specialist dots with defined organizational responsibilities, starting with focused pilots. Its work with Microsoft on Agent 365 integration is a stated plan. Teams should evaluate what is available to them today.
The business question I would take from this launch
Which work could move forward if someone stayed with it between your meetings, decisions, and conversations?
I would also ask what happens when your customers start delegating work to their agents. The opportunity extends to both how your team works and how people use your business.
At HatchWorks AI, we help teams choose the right AI opportunities and turn them into working systems. If dots have you thinking about where agents fit in your business, let's talk through the work you want to move forward.
