ZDNET’s key takeaways
- Dot and Duet aren’t game-changers for me – at least not yet.
- Their usefulness is partially limited by their connector options.
- Corporate reluctance to share data could hinder proactive AI agent adoption.
Both OpenAI and Anthropic have recently launched proactive AI personal agents that act as assistants. OpenAI calls its agent Dot and Anthropic calls its Duet. Dot and Duet scan cloud-based services and let you know about important items without waiting for you to ask. These function somewhat differently from traditional chatbots and traditional agent interfaces.
I tested both Dot and Duet over the course of three weeks and found that only one message from either agent was helpful. Duet caught a Chase chargeback that my support filters missed two days earlier. But the AIs mostly missed things I care about and surfaced irrelevant or out-of-date items.
Also: Claude’s new Personal Finance tool aims to help analyze your spending
To use OpenAI Dot, you need the ChatGPT Pro plan, which starts at $100/mo. Anthropic is less expensive. As long as you have any paid plan (starting at $20/mo), you can use Claude Duet.
For my testing, I’ve been using my existing Claude Max plan. OpenAI upgraded me to a Pro plan for the next few months so I could review Dot.
How ChatGPT Dot and Claude Duet work
In a chatbot, you ask a question and the AI answers. With an agentic interface, you give an instruction (like research a report, or organize some files), and the agent goes off and does it. You initiate. The AI does the work. Both of these types of AI are reactive.
Also: I let ChatGPT Work and Claude Cowork loose on my files – only one made me nervous
But Dot and Duet are proactive. They monitor your activities, as well as the connected services you enable. Then, they make suggestions about what you might need, or will need, to do. Think of your Dot or Duet as a chief of staff or a butler who knows what you need, and then either does it, or checks with you and then does it.
Dot and Duet run their own cloud-based virtual computers and allow for dedicated browser use on those VMs. In practice, this is a little creepy. It’s also pretty cool. If you’re the right sort of user, these modes can prove to be very helpful. They can save you a lot of time.
I am probably not the right sort of user, so keep that in mind as I critique these two tools.
Dot and Duet’s usefulness depends on connectors
Because both Dot and Duet are proactive, they rely on knowing about what’s happening in your world to be able to suggest and take action. They do this via ChatGPT and Claude’s connector/plugin interfaces.
When you use a connector with an AI, you’re granting it access to all (or a subset) of that connector’s information. For example, if you connect ChatGPT to Gmail, you’re letting ChatGPT see your email and, potentially, respond. The gotcha to this, of course, is that AIs are somewhat capricious. You can never be fully sure they won’t go rogue.
To be fair, I’ve had both ChatGPT (and its Dot) and Claude (and its Duet) connected to my Gmail for about a month. Other than the AIs both completely misunderstanding my email priorities, neither has done or sent or messed up anything without my permission.
Also: How I fixed a small business’s email overwhelm for $120 – thanks to ChatGPT Plus and Claude
I couldn’t share most of the connections that would have produced more helpful information. If the connectors offered by the AI don’t match the tools you use, they won’t be able to help.
Here’s one personal example: team messaging. I’m involved in seven Slack channels, including a board of directors I serve on. Some channels have non-stop traffic. Others may have one post per week. None of the Slack channels are mission-critical for me. Both Dot and Duet can monitor Slack channels.
The one mission-critical team messaging app I check at least 10 times a day is Google Chat. My main business team uses Google Chat rather than Slack. Even though Gmail has a connector, Google Chat does not.
There’s also one other snag: many companies aren’t exactly comfortable letting AIs have access to their corporate systems.
Here’s another personal example. I have a client for whom 90% of my work is AI-related. We manage our (and our team’s) projects using Asana. But when I asked for an okay to connect my Dot or Duet to Asana, I was told they were very uncomfortable with the idea. I imagine that if Google Chat was available as a connector, they would have vetoed that as well.
I’m guessing most people wouldn’t even ask for permission to let the AI have at it, but that’s not how I work with my clients. It’s not clear which systems require some admin-level permission to hook in and which just share whatever a user account can grab. But it’s likely that these AIs will see information that companies probably would prefer stayed private and under strict control.
So that’s a key consideration for these proactive assistants: if you can’t connect, you won’t get relevant information.
Dot and Duet notified me about old and irrelevant emails
Back in January, I tested Gmail’s AI overviews. I found them to be consistently incorrect. In fact, they still are. But ChatGPT’s Dot and Claude’s Duet are no better.
In my case, I think it’s because I get so much email that without fairly deep training, the AIs are simply overwhelmed.
That screenshot is a good illustration. I was making breakfast in the kitchen when my Apple Watch binged one morning. That day, I received more than 1200 email messages. Somehow, this is the one Duet thought needed attention. The problem? Who’s Karlie? I don’t know a Karlie.
Then, there’s this.
This was what my Dot presented to me right after I enabled the agent. I had to do some digging, but it turns out that I did have a customer support email from Inès. In 2023. Yeah, that person’s issues were dealt with years ago.
Notice that item 2 referred to Tuesday’s announcement. When I searched for the vendor, I found it was an announcement for the previous Tuesday.
As for item #3 in that post, that review was completed back in 2025. The video has been live on YouTube for more than 18 months.
I get hundreds to thousands of announcements every day. Lots of pitches come in from companies who want me to discuss their products. The thing is, the AI can’t tell the difference between a PR rep trying to forcibly twist my arm into writing about them, and a truly urgent item that should be pointed out to me.
Over the weeks I’ve been testing Duet and Dot, almost all of the so-called urgent alerts they provided me were about press information that had no relevance to what I was working on.
Duet did flag something my email filters missed
On the other hand, there is some value here. This message showed up from Duet early in my testing.
The first item wasn’t relevant to anything I was working on, but the second caught my attention. I didn’t know about a Chase chargeback. A chargeback is when a customer tells a financial institution they didn’t get a product or don’t want to pay for it, and the financial provider removes the payment from your account.
My policy with my little software business is to always refund a customer if they don’t want an item. So I should never see a chargeback. But more to the point, I have a comprehensive set of filters set up to manage customer support inquiries. So where was this Chase thing coming from?
As it turned out, Chase sent the notice in a new format. It had slipped past my attention two days earlier. I would never have seen it if Claude hadn’t pointed it out. While most of what my Dot and Duet notified me about was entirely irrelevant, this was genuinely important.
I immediately followed up with the customer. The situation ended with everyone satisfied with the result.
Are proactive AI alerts worth the usage consumption?
Both companies say they don’t charge you usage credits for simply turning on Duet or Dot. But any AI processing they do is charged against your usage allocation. Having the AIs constantly spinning through email, and any other service you have connected, could consume your usage credits fairly quickly.
Also: I tried replacing Claude Code with a local AI stack that’s completely free – how it went
For me, the signal-to-noise ratio wasn’t good enough. Yes, Duet caught a financial charge that I didn’t know about. But Duet and Dot missed many other important email messages from clients and constituents that were actually really important. I’m just not sure burning the usage credits is worth it for the small return.
Talking to my Dot feels like making an old-school phone call
I’ve written in previous articles about how I find ChatGPT’s Voice feature valuable when I’m working hands-free. I recently used it to help fix my gaming PC while I had both hands deep into the guts of its motherboard. But, for some reason, OpenAI decided to change up the Dot interface so it looks like you’re talking on your mobile phone. Seriously, who talks on their mobile phone anymore?
On the left is ChatGPT’s normal voice interface. In the middle is the Dot interface. Notice the phone icon in the upper right corner. For those under 40, that icon is what a handset-based wired telephone looked like in the Middle Ages. Yes, telephones were tethered to the wall back then. If you wanted to walk around with your phone, you needed to get a very long cord. It’s a memory most of us born before the turn of the century don’t like to talk about.
Oh, and the third screen? That’s what Dot presents when you call it to talk to it. Seriously.
Somehow, the folks at OpenAI thought it would be more personable if you “called” Dot to have a conversation. This after ChatGPT Voice had been in play for more than a year and worked quite well. Was this some bid to reach out to older customers? Was someone just trying to be cute? This feature doesn’t cause any harm. It works just as well as the Voice interface. But it’s cognitively disturbing to have to “call” your Dot on the phone. It’s just weird.
Five Dot and Duet features I’d like to see
Throughout my testing, I found a bunch of usage cases that would have made a proactive AI assistant insanely great. Unfortunately, they weren’t implemented. So, in the hope that OpenAI and Anthropic will take some of my suggestions, here are features that could really help.
Connect with my programming projects: The very latest release of Duet can connect to Claude Code projects, and can even describe the most recent work. This is cool. You can even initiate sessions on your phone. But what I’d like to see is an integrated view across projects along with proactive reminders of what I need to work on. I suspect we’ll see this soon, but it’s not here yet.
A better Help Scout connector: Help Scout is the support help desk I use to help customers. Both Claude and ChatGPT can read Help Scout posts, but neither allows replies. To be fair, that’s more a limit of Help Scout than it is Duet or Dot, but it shows how the connector’s limits block usefulness. If I could have an alert come in, ask the AI to draft a reply, and then have that reply sent, that could improve customer responsiveness a lot.
An Apple Watch app: This one baffles me. Both Claude Code and Codex can easily write an Apple Watch app. I know, because I did it almost a year ago. Having a phone app is nice, but where I could truly use a continuous connection to my AI assistant is on my wrist. I wear my watch everywhere. Everywhere. Being able to check in, add assignments, and chat with my AI assistant on my Watch would be a huge productivity boost. It should have a complication so I can initiate it right from the Watch face.
Folders and projects in Dot: Dot has no folder or project mechanism. Everything you do with your Dot is done in one long thread. Folders make project organization much more powerful. Folders and Projects are features that are well-baked into ChatGPT already. Digging through one long thread is maddening. Duet does have some organization available. The feature is called Channels.
Better native app integration: Neither Duet nor Dot can update or read Apple Reminders. If they’re going to have connectors in the cloud, it might be helpful for the on-device apps to be able to talk to any of the native apps where Apple has exposed interfaces. EventKit can read and write to Reminders, so that might be the first place to start.
Keep in mind that this is very much a moving target. Both companies are updating and improving these offerings on a nearly daily basis.
How much of your data are you willing to share with an AI assistant?
We’re still in the early days with proactive personal assistant agents. The idea of them is both intriguing and worrisome.
Success will require buy-in from all participants in the ecosystem. SaaS vendors like Help Scout, and corporate clients who allow AI access to their systems (like my request for Asana access), will need to cooperate for the AI to be able to get better at understanding what to look for.
Security issues are a very big concern. Do you want the AI to read all your email or have access to your confidential corporate resources? And worse, everyone you send email to may also be exposing their entire email archive to unspecified and unknown AIs.
This isn’t really new. It’s always a good idea to assume that whoever you email could become a leak. Truly sensitive information should never be sent in an email message. That truth is becoming even more valid today.
Also: Meta Muse is the worst AI agent for privacy – and I’ve tried them all
I am fairly sure that OpenAI, Anthropic, and Google are trustworthy in their own ways. We’ve been, for example, trusting email to Google for years now. But when these proactive agents start being offered by companies who love to hoover information for profit (Meta’s Muse, I’m looking at you), these platforms could become a wildly dangerous foundation for anything from marketing data mining to all kinds of other corporate espionage.
It will be interesting to see what “progress” looks like in the near future.
You can follow my day-to-day project updates on social media. Be sure to subscribe to my weekly update newsletter, and follow me on Twitter/X at @DavidGewirtz, on Facebook at Facebook.com/DavidGewirtz, on Instagram at Instagram.com/DavidGewirtz, on Bluesky at @DavidGewirtz.com, and on YouTube at YouTube.com/DavidGewirtzTV.