Proactive AI Agents: What to Let an Always-On Assistant See and Do

At 8:00 on a Saturday night, one of our clients had a security issue and emailed their consultant about it, and they CC’d me. I get an incredible amount of email. I would have probably not seen that one until Monday morning, and I might not have completely read it even then. They didn’t call us or open a ticket, either of which would have triggered our on-call response. We always tell people to call us. You wouldn’t email 911 if your house was burning down.

My dot, OpenAI’s new always-on agent, saw it first. It sent me a Slack message: “This client sent an email to their consultant, and this looks like something we might want to look at tonight.” I hadn’t told it to look for anything like that. It went ahead and flagged the email because it figured I’d want to know. My dot is connected to my work email and Slack, and it pings me about the things that get lost in the avalanche of everything else that comes in.

Proactive AI Agents: dots, Grok Bot, and Muse

A chatbot waits for you to ask it something. A proactive agent keeps working when you’re not talking to it: it reads the accounts you’ve connected (email, calendar, Slack, files), decides what you probably need to know, and messages you first. Some will also offer to act, like drafting a reply or proposing a new meeting time.

Three of them launched since mid-August:

• dots, from OpenAI, launched September 29 inside ChatGPT.¹
• Grok Bot, from SpaceXSI (formerly xAI), went into beta August 11.²
• Muse, from Meta, launched September 8 in the US, free for most uses.³

I use all three. Muse handles my personal life through my personal Gmail, Facebook, and Instagram. It tracks inbound packages, and when I got into watches, it went back through a year of email and built me a spreadsheet of every watch I’d bought. It noted when I bought each one and where I bought it, and it saved the receipts as well.

My dot and my Grok Bot are connected to my business accounts, and they work as my chief of staff. They’ll tell me I left someone a voicemail and dropped them an email but never responded to the full thing they sent me. They’ll tell me I’ve got two meetings that overlap tomorrow and ask if I want them to email an alternative time slot. My Grok Bot only ran on a schedule until early October, when SpaceXSI added a main Bot that checks in on its own.¹⁰ For someone like me who’s busy, whose calendar gets crazy, that proactiveness is pretty fantastic.

What the Agent Can See

The first week of October, I sent a client a scope of work that said a piece of their hardware had five open hard drive bays. Then I got a message from my dot: it had noticed the scope of work went out, and it had found a Slack message from June where our engineering department confirmed all eight bays were full. I was able to go back and get that corrected.

I never pointed it at that Slack thread. It found the message because it reads everything I’ve connected it to, and it decided on its own that the two belonged together. That’s why it’s useful at work, and it’s also why you should think hard before you connect one there, because it reads whatever that account can open.

At most companies, that’s more than anyone realizes. Shared folders, old project sites, and “just give everyone access” settings pile up for years. In a 2024 Gartner survey, 40% of IT leaders said oversharing had delayed their Microsoft 365 Copilot rollout by three months or more.⁴ An agent reading in the background will find a salary spreadsheet somebody shared with the whole company years ago.

Muse goes further than the other two. According to a TIME analysis that Meta didn’t dispute, it updates its profiles of you and of the people you mention every hour, including people who have never used Muse.⁶ A client who emails one of your staff can end up in that staffer’s Muse.

Before you connect a work account, find out what it can open. If the answer is “everything on the shared drive,” do the Microsoft 365 access cleanup you’ve been putting off, which makes it safer with or without AI.

What the Agent Is Allowed to Do

All three let the agent take some actions without asking and make it ask before others. dots and Grok Bot also hand some steps back to you entirely. dots lets you set that per action.¹ Here’s where I’d put the line for a company starting out:

• Let it watch. In watch mode, it reads email, calendar, Slack, and files, tells you what needs attention, and does nothing else.
• Let it draft. It writes the reply, the reschedule, or the follow-up and leaves it for you.
• Don’t let it send. A person reads it before it goes to a client, a vendor, or your team. When my agents offer to email someone a new meeting time, I approve that send myself. Grok Bot added a setting on September 30 that lets it send email and Slack messages without showing you a draft.¹⁰ Leave it off.
• Don’t let it buy. It can use the browser, find the best deal, and put it in the cart, then tell you it’s ready for review. You make the purchase.

I don’t give these agents payment information today. I might let one order a pizza from an account where the card’s already saved, and even that I’d have to think about.

SpaceXSI’s documentation notes that an approval controls the action it’s asking about, and “it does not reverse work already completed.”⁵ By the time an agent asks you to approve an email, it has already read everything it used to write it.

If an agent sends something wrong, your company is responsible for it. In 2024, Air Canada argued that it was not responsible for what its website chatbot told a customer. The British Columbia tribunal that heard the case called that “a remarkable submission” and made the airline pay.⁷ Microsoft Agent 365 requires every agent to have a sponsor, a named person accountable for it.⁸ Use that rule whatever product you pick.

How Your Staff Will React

For me, having an agent watch my inbox and second-guess me is a breath of fresh air. One of the things I struggle with is keeping up with the flow of communication that comes in. It’s just tough.

Not everyone on your team will feel that way. In one study, about 90% of programmers preferred a proactive assistant over a plain chatbot. Another group got a version that chimed in sooner and more often, and fewer than half of them preferred it, even though it still helped them get more done.⁹ Start with the people who are drowning in messages, turn the frequency down, and let everyone else come to it.

Keep a closer eye on the busiest people. One of the testimonials on SpaceXSI’s launch page says the customer trusts the Bot to run “without me verifying and reviewing.”² That’s the person who should stay in watch-and-draft mode the longest.

Which One to Start With

For a company, I’d lean toward OpenAI, on a business plan your company’s admin controls. As of October 2026, it has the controls I’d want: on Enterprise, dots stays off until an admin turns it on, and you set rules for each kind of action.¹ Grok Bot does a good job for me. Its shared computer means you have to plan around which logins live on it, though, and that gets harder across a team. On Cursor’s Teams plan, Grok Bot is on for every member, and admins can’t switch it off.¹¹ Muse started as a consumer app, and Meta is now pitching it to small businesses too.¹² For now, I’d keep it for your personal life, where it has worked well for me.

Trying It Alone or Rolling It Out

I’m hearing almost nothing about always-on agents from clients yet. Most people outside the AI world have no idea what’s been released in the last couple of months. That gives you time to sort out access and rules before your staff start connecting their own accounts to whatever looks helpful.

If you’re one person trying this out, you don’t need us. Go through the vendor’s setup, keep it in watch mode, and see what it catches. If you’re figuring out how a team of fifty or five hundred should use agents, that’s what our thinkAI program is for. We help you decide what AI should and shouldn’t do in your company and train the people who’ll run it. The access cleanup that makes agents risky is Microsoft 365 work, and we do that too. Start a conversation when you get to that point.

Sources

1. OpenAI dots. OpenAI, “Introducing dots,” September 29, 2026, and the dots help center articles on getting started and on privacy, security, and safety. Availability, read-only background research, action rules, and what happens to data after you disconnect an app. openai.com, help.openai.com, help.openai.com

2. Grok Bot launch. SpaceXSI, “Introducing Grok Bot,” August 11, 2026, including the customer testimonial quoted. x.ai

3. Muse launch. Meta, “Introducing Muse,” September 8, 2026, including the training opt-out. about.fb.com

4. Copilot oversharing delays. Matthew Finnegan, “Microsoft moves to stop M365 Copilot from oversharing data,” Computerworld, December 4, 2024, reporting a Gartner survey of IT leaders. computerworld.com

5. Grok Bot shared computer and approvals. SpaceXSI documentation, “Approvals, security, and privacy.” docs.x.ai

6. Muse profiles. Harry Booth, TIME, October 6, 2026. time.com

7. Moffatt v. Air Canada, 2024 BCCRT 149, British Columbia Civil Resolution Tribunal, February 14, 2024. canlii.org

8. Agent sponsors. Microsoft Learn, Microsoft Agent 365 agent identity, “Sponsor and auditing.” learn.microsoft.com

9. How often an assistant interrupts. Chen et al., “Need Help? Designing Proactive AI Assistants for Programming,” CHI 2025. A lab study with 65 participants. doi.org

10. Grok Bot changelog. SpaceXSI, versions 0.64.0 (September 30, 2026, sending without drafts) and 0.66.0 (October 2, 2026, the main Bot that checks in on its own). x.ai

11. Grok Bot for teams. SpaceXSI documentation, “Teams and enterprises.” docs.x.ai

12. Muse for small businesses. CNBC, October 7, 2026, reporting on Meta’s pitch for Muse. cnbc.com