I Gave OpenAI’s Dots Full Control of My Laptop. Here’s What Happened.

I gave OpenAI’s Dots access to my laptop and browser, started handing it work through ChatGPT, and came away impressed. Not “this might be useful someday” impressed. More like “I’m ready to hand over the next thing on my list” impressed.

That is a fairly big admission. It is also not a recommendation to give an AI agent the keys to your accounts. But once I started using it for the work that actually piles up in my day, the appeal became obvious.

Emails, invoices, travel logistics: none of these tasks is necessarily difficult on its own. Together, they eat the time I want to spend creating things. Dots started turning that backlog into a sequence of actions, and I kept giving it more.

Key Takeaways

• Dots drafted emails in my voice and assembled an invoice email across Drive and Gmail.
• Travel planning included hotel recommendations and proactive outreach about rental car access.
• Success with individual work does not establish readiness for complex enterprise requirements.
• Agent access and action boundaries varied, while the legal status of an agent-signed release remained unresolved.

Table of Contents

• The Real Test: Work That Keeps Piling Up
• Email Drafts That Actually Sounded Like Me
• Pulling an Invoice From Drive Into Gmail
• Planning a Weekend in El Salvador
• Useful for Work Does Not Automatically Mean Enterprise-Ready
• As AI Agents Converge, Trust Becomes the Question
• The Contract Twist
• My Verdict: Impressive, With Boundaries That Still Matter

The Real Test: Work That Keeps Piling Up

My job produces a steady stream of small obligations. Meanwhile, content creation requires deep work, which means getting away from the inbox. Those two realities do not coexist particularly well.

If I focus on creative work, the administrative stuff accumulates. Eventually, I need to clear it quickly. That is the situation in which I tested Dots.

The interesting part was not just that it could respond to a request. I could keep messaging it, and it would pick up what I wanted to accomplish and start sequencing the work. Instead of treating every task as a separate conversation, it felt like I could keep feeding it the backlog.

So I went further than I initially expected: I gave it access to the laptop and browser and started asking it to do real things.

Email Drafts That Actually Sounded Like Me

The first request was personal. I asked Dots to read as many of my recent emails and articles as it could and learn my writing style.

Then I began moving through my inbox.

You know the kind of email where you already know what you want to say, but getting the phrasing right takes almost all the energy? You want to be direct without sounding rude. You want to answer the question without creating three more questions. The actual decision is easy. Putting it into acceptable email language is the work.

Dots picked up my voice quickly. I would tell it what I wanted to communicate, and it would put a draft in Gmail. Then I could open the draft, change a little, and send it.

That distinction matters. For this part of the test, I was not asking it to autonomously answer everything in my inbox. The workflow was:

1. Give it context from my existing writing.
2. Tell it the intended response.
3. Have it create the email draft.
4. Review, edit, and send the message myself.

The drafts started piling up in a useful way. Rather than spending my energy composing each response, I was reviewing something that was already close to what I wanted.

That made clearing the inbox much faster. It also showed why an agent can feel different from a chatbot: the output was waiting where I needed to use it, not sitting in a chat window for me to copy elsewhere.

Pulling an Invoice From Drive Into Gmail

Next, I asked it to handle an invoice.

The instruction was straightforward: my invoice was in Google Drive. Go into Drive, download it, attach it to a Gmail draft addressed to the people I usually send it to, and write the accompanying note.

It handled that fairly effectively.

This was more than generating a polite sentence about an attachment. It involved moving between tools, retrieving the file, identifying the usual recipients, and assembling the draft.

The value was in connecting the steps. Each individual action was ordinary. But those ordinary actions are exactly the ones that make administrative work feel endless when there are enough of them.

After the emails and invoice, I was ready to push it further.

Planning a Weekend in El Salvador

I had a weekend trip to El Salvador coming up, so I gave Dots the places we wanted to visit and a hotel budget: an average of about $100 a night across the stay.

That did not mean every hotel had to cost exactly $100. A little more on one night and a little less on another was fine.

I asked it to:

• Find a rental car.
• Find hotels that fit the trip and overall budget.
• Proceed with the hotel bookings.
• Put together an itinerary we could follow once we arrived.

At this point, my attitude was basically: fine, I’m in. I wanted to see how much I could delegate.

There Was Still a Human in the Loop

When I described the experience to Ranjan Roy, he immediately pushed on the booking process. Had I actually let the agent choose everything? How many options had it given me? Was I just accepting its single recommendation?

The answer was less fully automated than my enthusiasm might have suggested. I liked two of the three hotel choices. Dots got into the booking process, but I still needed to complete the final step myself.

So, yes, there was still a human in the loop. I reviewed the choices and clicked to complete the bookings. It was substantial delegation, not a completely autonomous vacation purchase.

The Best Moment Was a Follow-Up I Had Not Requested

Then Dots noticed a potential problem: getting to one of the hotels involved some off-road driving.

It thought the rental car might work, but it was not certain. Rather than leave that uncertainty buried in the itinerary, it asked whether it should email Budget, the rental agency, and the hotel to check whether the car could make it there.

I said yes.

When I checked my sent messages, the emails were there, written in my voice and sent to the relevant people.

That was the moment that really got me. The agent was not merely finding a hotel and calling the task finished. It had noticed a dependency between the hotel and the rental car, identified what it did not know, and proposed a way to resolve it.

The actual answer about the road was still something those businesses needed to provide. But catching the issue and handling the outreach was useful work.

Useful for Work Does Not Automatically Mean Enterprise-Ready

I connected the experience to Dots being built for work and using frontier intelligence. Ranjan pushed back on the enterprise label, and I think the distinction is worth keeping.

Email and invoicing are work. That does not mean an agent has demonstrated readiness for a large, complex organization.

In my case, I could decide what to delegate. I had authority over the relevant tasks. An enterprise invoice might involve an approval chain, security requirements, and responsibilities spread across several people.

Helping an individual get work done and operating inside an enterprise are different tests.

Ranjan’s point was that small and medium-sized businesses, individual professionals, and large organizations should not be treated as one market. An agent can have enormous potential for the first two without having solved the requirements of the third.

For broader context on the organizational side of that distinction, the NIST Privacy Framework offers a useful reference for thinking about privacy risk management. My successful inbox session was not a test of those kinds of organizational controls.

As AI Agents Converge, Trust Becomes the Question

The experience was good enough that I briefly thought about deleting Muse.

That surprised me. Muse had been useful for some consumer tasks, but after this session, Dots felt compelling enough that I wondered whether I needed both.

Ranjan had been experimenting with other agents, too, including Q from Manus, which he described as having an inline desktop video editor. His broader observation was that these products were already starting to feel interchangeable. Everyone seemed to be building toward a similar kind of personal agent.

If that continues, the important question may become less “which one has this feature?” and more “which one am I willing to give access to?”

Our own behavior illustrated how inconsistent that answer can be. Ranjan had given Instinct his credit card, but not Muse. He had, however, given Muse his American Express login to check his Platinum benefits, and said it did a good job.

I had given OpenAI broad access to my working environment. When asked whether that meant I trusted it, my answer was not exactly reassuring: I do not really trust any of them. There are some whose access I am willing to tolerate and others whose access I am not.

By “tolerate,” I meant being willing to provide my details. That is not the same as having a settled, consistent theory of trust.

I also tried to give Dots my credit card, and it would not let me. So even within one product, my willingness to delegate and the agent’s willingness to act did not always line up.

The Contract Twist

The most striking difference between agents came with a document signature.

I had recently tried to get Claude Cowork to sign a contract for me. It refused.

Then I asked Dots to go into my email and sign a Web Summit speaker release, the sort of release I had signed many times before.

Dots did it.

That made for a funny exchange about whether I had just created the story of an agent signing a release and somehow getting me thrown out of the event. But underneath the joke was a real unresolved issue: we did not establish whether that agent-signed document was legally binding.

Nor does an agent’s willingness to perform an action prove that it should perform it. One product refusing and another complying tells us something about their different boundaries. It does not, by itself, tell us which boundary is right.

My Verdict: Impressive, With Boundaries That Still Matter

My review is enthusiastic. Dots helped me turn a pile of ordinary obligations into completed work or drafts that were ready for my review. It captured my email voice, assembled an invoice email, advanced my travel planning, and noticed a practical problem with reaching a hotel.

The strongest part was not any single trick. It was the ability to keep handing it tasks and have it carry the work forward across the tools I already use.

But the qualifications are part of the story, too. I still reviewed emails and hotel choices. I completed the final booking steps. Large-scale enterprise requirements were not tested. And the contract example left a legal question unanswered.

I went in with a backlog and came out wanting to delegate more. That is a powerful product experience. It is also exactly the point at which questions about access, approvals, and trust become harder to ignore.

Disclaimer: This is one person’s experience, not a recommendation to give any AI agent access to your accounts, files, or payments. The episode description reports Dots availability for ChatGPT Pro subscribers, with plans from $100 per month, based on a New York Times report discussed during the episode.

Source credit: This article is adapted from the original discussion published by Big Technology. The full discussion is available in “I Gave OpenAI’s Dots Full Control of My Laptop (Here’s What Happened).”