“Proactive” AI is just automation with guesswork
Yesterday I gained access to Claude Duet. Maybe it’s the Baader-Meinhof phenomenon, but 4 times in the course of a 30-hour period, someone asked about or mentioned Claude Duet. Maybe it’s just reaching a wider audience now, but I thought it was strange that something I heard about once when the waitlist was announced suddenly popped up 4 times in short order.
It also seems to me that more companies are starting to release their own version of this, with several reaching out to me asking me if I’d promote their tool to you, dear reader.
I signed up for Duet and will talk more about it in Solo Ops Toolkit Pro, but I have reservations about AI being “proactive” for a few reasons.
Why i’m Not Sold on Proactive AI
One is that AI doesn’t understand intent. It’s just taking a guess based on whatever context it has. It seems the thing Claude Duet is best at right now is reading your emails and concocting a response to them. That’s fine, but it’s also not something I’m keen on outsourcing to AI. It’s not nearly as capable as an actual “teammate” or “assistant,” which is how many of these proactive AIs are branded.
A second is that AI is designed to be sycophantic and appease us. It’s not that hard to think that, even if you tell it never to do something without you giving the OK or asking you about it, it might take the proactive steps a bit too far. When it’s done that in chats, the only recourse is for it to print an apology on screen (note, I didn’t say apologize because it doesn’t have sentiment) and carry on.
Number 3 is something my friend Jessica touched on in her recent newsletter: I’m dubious about the actual productivity gains. As I’ve been saying for what seems like decades now, “doing more stuff” isn’t efficiency. The goal should not be more, it should be better. AI doing things, proactively or not, doesn’t make you more productive unless you’re measuring some actual goal. In baseball, “more” swings lead to more strikeouts. Better swings lead to hits, or at least productive outs (outs that move a runner or score a run).
“Proactive” AI is Really More User-Friendly Automation
In my limited experience, Claude Duet did OK work looking at my inbox. I had it come up with a “coaching proposal” based on a discovery call I had, and it was fine. I mean, I wouldn’t send anything it wrote and I prefer to send videos, but it did manage to map the client’s problems to my appropriate offers. I also very likely mentioned those offers during the call.
Which brings me to my main point (400+ words in…I’m feeling verbose this morning): it seems these “proactive” AIs really just have the same triggers as traditional automation tools. Things like new email, new record in Notion, new task in Todoist, etc. But the action is ambiguous. The AI makes a guess about what you want to do based on stats.
This is perhaps more user-friendly, but it’s also not very predictable; predictability is the entire point of an automation. You should be able to reliably trust that the automation will execute and you will know the end result without having to check it.
Should You Use Proactive AI?
There are too many variables at play for that to be true with “proactive” AIs, at least at this point. And if you cannot trust the output, “proactive” AI becomes a liability.
If you do end up having to approve everything it does for you, then you need to consider if it’s worth having it done for you in the first place, or if more traditional automation (like Zapier) might actually be better.

