Three major companies shipped proactive AI this month, and forty years of automation research says most people will stop reading it by week thirty-six. This article answers the question of whether to adopt a proactive AI assistant, and how to set one up so you are still using it in six months.
The Short Answer
I need to tell you something about myself before I give you advice about this, because it is the whole reason I have an opinion.
I am ADHD-wired. When something genuinely interests me, I go all the way in and the rest of the world goes quiet. That is the good version. The other version is that an interruption at the wrong moment does not cost me thirty seconds. It costs me the thread, and sometimes it costs me the afternoon.
So when three separate companies decided this month that the problem with AI is that you have to remember to ask it for things, my first reaction was not excitement. It was a flinch.
Anthropic put Claude Duet into research preview on October 21, described as a proactive agent app. Microsoft shipped Autopilot, which it frames as a digital coworker with its own identity and configurable permissions. And on Product Hunt this week, a product called Kiwi launched with the tagline "a proactive AI assistant that texts you before you ask."
Three independent bets on the same premise: the prompt box is the bottleneck.
Here is my honest answer to the question in the title. Yes, try one. They are solving something real, and the thing they are solving is probably costing you more than you think.
But set it up on purpose, because the research on what happens when a system starts volunteering information is forty years old and it is not encouraging. Parasuraman and Riley published the framework in 1997, and the finding that should shape your decision is this: automation failures shift your trust and your complacency far more than correct recommendations do. One bad unprompted suggestion costs more than ten good ones build.
Which means proactive AI earns a response rate, and response rates decay. Your job is to decide in advance what it is allowed to interrupt you about, so that the thing is still earning its place in month six instead of sitting in a muted channel with forty-one unread.
Key Takeaways
- Three major proactive AI products shipped in October 2026, all betting that the real bottleneck is remembering to ask rather than AI capability.
- They are solving something real, since most lost value comes from never thinking to use AI for a task rather than AI failing at it.
- Automation research from 1997 found that failures move trust and complacency more than correct recommendations do, so early false alarms are expensive.
- Clinical alert systems are the closest long-running parallel, and override rates there run as high as 96%, with measurable week-by-week decay in response.
- Set three tiers before you turn it on, interrupt now, daily digest, and never, and track your own action rate so you can tell whether it is still working.
Why I Flinched
Let me describe the version of this I already live with, because I suspect it is your version too.
I have notifications from more sources than I can name without checking. Some of them I read. Most of them I have trained myself to ignore, and I did not make that decision consciously. It happened gradually, the way you stop hearing a highway when you live near one.
Now here is the uncomfortable part. Among the things I have trained myself to ignore, some of them occasionally mattered. I do not know which ones, because I stopped looking.
That is the actual risk with proactive AI, and it is not the one people worry about. Everyone frets about an AI acting without permission. The likelier outcome is much more boring: it becomes one more voice in a room that was already too loud, and you solve that by stopping listening.
And I want to be fair to what these companies are building, because the problem they identified is genuine and I have watched it cost people real money.
The single biggest waste I see in AI adoption is not people using AI badly. It is people never thinking to use it at all for things it would have handled in ninety seconds. Someone will spend two hours on a task I know Claude could have drafted, and when I ask why they did not use AI, the answer is almost never "I tried and it failed." The answer is "I didn't think of it."
That is a trigger problem, not a capability problem. And a system that watches and speaks up is, in theory, exactly the right shape of solution.
So I am not dismissing this. I think the premise is correct.
But what if being right about the problem is the easy part?
Forty Years of Somebody Else's Expensive Lessons
Here is what the research says, with the limits clearly marked, because I am drawing an analogy and I want to be honest that it is an analogy.
The foundational framework is from 1997. Parasuraman and Riley named two failure modes for automation. Misuse is over-reliance, which produces monitoring failures and decision bias: you stop checking because you assume it is handled. Disuse is under-reliance, where genuinely useful help gets rejected because of past bad experiences. Both are failures. Most people only guard against one.
Failures cost more than successes earn. The research finding I keep coming back to is that negative experiences, meaning automation failures, have stronger effects on subjective trust and on complacency than positive experiences do. That asymmetry is the whole design problem for proactive AI. A system that interrupts you ten times correctly and once badly does not net out at nine. It nets out worse than that.
Complacency is an attention problem, not a character problem. Parasuraman and Manzey followed up in 2010 treating automation complacency as attention allocation: humans give less attention to channels they believe are being handled. You do not decide to stop checking. Your attention quietly reallocates and nobody sends you a memo about it.
Medicine already ran this experiment at scale. Clinical decision support systems have been interrupting doctors with alerts for two decades, which makes it the closest long-running parallel we have. A systematic review of 23 studies found average alert override rates between 46.2% and 96.2%, with somewhere between 29.4% and 100% of those overrides judged appropriate depending on the study. One frequently cited estimate puts medication alert overrides around 90% to 95%.
And the decay is measurable. In one study of clinical trial recruiting alerts, response declined roughly 2.7% per week, falling from about 50% to about 35% over 36 weeks. Same system. Same alerts. Nine months later, a third fewer responses.
Pruning is the intervention with evidence behind it. When one institution deactivated clinically irrelevant drug interaction rules, pharmacists' override rates on the remaining unchanged alerts measurably improved. Fewer alerts made the surviving alerts more effective. That is the most actionable finding in this entire body of work.
Now the caveats, and they matter.
Clinical alerts are not AI assistants. The stakes differ, the volume differs, and the alerts in those studies were mostly rule-based rather than judgment-based. A 2026 systematic review in JAMIA explicitly cautions that override rate has never been validated as an actual measure of alert fatigue, so I am not presenting these numbers as proof of anything about Claude Duet.
What transfers is the shape of the thing. Any system that interrupts you earns a response rate, response rates decay with volume and with false alarms, and reducing the number of interruptions improves the ones that remain.
That is enough to design around.
What I Decided to Do About It
Here is how I am approaching this, and it is built around one admission: I cannot be trusted to manage my own attention by willpower. Neither can you. That is not a weakness, it is just true of everyone, and systems that assume otherwise fail.
So the setup has to do the work.
Three tiers, decided before anything is turned on. Interrupt me now. Daily digest. Never. And the conditions for each one have to be mine, written down, not inferred by a tool that has an incentive to seem useful. The "never" list is the one most people skip and it is the most important, because without it everything is eligible.
Almost everything belongs in the digest. When I was honest about the moments I wished something had told me sooner, almost none of them actually needed to reach me within the hour. They needed to reach me before the end of the day. The number of genuine interrupt-now events in a small business is very small: a client about to churn, a payment failure, a deadline that moved. Everything else is important without being urgent, and important-but-not-urgent belongs in one daily block, not in your pocket.
One domain, not everything. Give it the single area where a missed item costs real money. For most of the people I work with, that is a client going quiet or a renewal coming due. Not your whole calendar. Not your whole inbox. One lane where you can tell whether it is working.
Subtract before you add. This is the step I would not skip for anything. If you add a new interruption source to a channel that is already saturated, you have not improved your attention, you have diluted it further. Something has to come off before this goes on.
Track your own action rate, because you will not notice the decay. One line per item: what it surfaced, did I act on it. Roll it up weekly. Check at thirty days and again at ninety. The medicine research says the drift is slow and invisible from inside, so the log is not bureaucracy, it is the only instrument you have.
Write the kill switch while you are still skeptical. Before you get attached. Two or three specific conditions under which you turn it off, and what you go back to doing instead. Parasuraman and Riley's disuse finding is the reason: if you do not retire it on purpose, you will keep paying for something you have quietly started ignoring, and you will also become unreasonably hostile to the next one.
What I like about running it this way is that it turns a vibe into a number. In thirty days I will not be asking myself whether the assistant feels useful. I will be looking at an action rate.
How to Do This Starting Today
1. Write down the moments you wished something had told you sooner. Last thirty days, with specifics. Not categories. Actual moments. This list is the entire specification for what you need.
2. Sort that list honestly, because some of it is free. A reminder fixes more of it than you expect:
[The Job]
Sort my wished-for interruptions into things a reminder would fix and things that genuinely need judgment.
This is for: [ME].
It matters because: I may be about to pay for AI to solve a problem a calendar entry solves for nothing.
[The Background]
Here is what you need to know: moments in the last month when I wished something had told me sooner are [LIST WITH DETAIL]. For each, whether the timing was predictable in advance [YES/NO], and whether knowing required reading or interpreting something [YES/NO].
Do not use: AI as the answer where a scheduled reminder would do the job.
[The Deliverable]
Return: two lists, reminder will fix this and needs interpretation, with a one-line reason each.
Must include: the share of my list that a free calendar reminder handles.
Optimize for: accuracy.
[The Questions]
Ask me any questions you have.
3. Count what is already interrupting you. Every source, rough daily volume, and your honest action rate on each. Note which ones you have already stopped reading. That last column is a prediction about the new one.
4. Turn something off to make room. Net interruptions should not go up. If you cannot find anything to remove, that is a finding: your channel is full and a proactive assistant will land in it as noise.
5. Write your three tiers.
[The Job]
Write the rules for what a proactive AI assistant is allowed to interrupt me about.
This is for: [ME].
It matters because: automation research found that failures shift trust and complacency more than correct recommendations do, so one bad interruption costs more than several good ones earn.
[The Background]
Here is what you need to know: the domain I am monitoring is [DOMAIN]. Things that genuinely need my attention within an hour are [LIST]. Things that can wait for a daily summary are [LIST]. Things I never want surfaced are [LIST]. My tolerance for a false alarm is [HIGH/LOW AND WHY].
Do not use: urgency labels the AI assigns itself. The thresholds have to be mine.
[The Deliverable]
Return: three tiers, interrupt now, daily digest, and never, each with the specific conditions that put an item in it.
Must include: the exact test the assistant applies before interrupting me.
Optimize for: clarity.
[The Questions]
Ask me any questions you have.
6. Start the log on day one. One line per item, ten seconds. If it takes longer than that you will abandon it by Thursday, and then you will have no idea whether this worked.
7. Put the thirty and ninety day checks in your calendar now. While you still care about the answer. In ninety days you will either have a number or a feeling, and the feeling will lie to you.
Frequently Asked Questions
Is proactive AI worth trying at all?
Yes, if you give it one domain and measure your own response. The problem it targets is real: most lost AI value comes from never thinking to ask, not from AI failing. The risk is that it becomes background noise, and that risk is managed by volume limits and tracking, not by willpower.
What is the difference between a proactive assistant and a reminder?
A reminder fires on something you already knew was coming. A proactive assistant surfaces something you did not know to look for, which requires interpreting data. When you sort your own wish list, a surprising share turns out to be reminders, and those are free.
How do I know if I have stopped paying attention to it?
Only by tracking. Automation complacency is an attention-allocation effect, so it does not announce itself. Log whether you acted on each item, roll it up weekly, and compare week one to week twelve. A falling action rate is the signal, and you cannot feel it from inside.
How many notifications is too many?
Fewer than you think, and the fix is subtraction. Clinical research found that deactivating irrelevant alerts improved response rates on the alerts that remained. If a proactive assistant surfaces more than a handful of items a day, prune its triggers rather than trying harder to read them.
Should I give a proactive AI access to my email and calendar?
Only after you have decided what it may never do without asking. Microsoft now describes its agent as having its own identity and configurable permissions, which is the right framing and a warning. Write the permitted and forbidden action lists before you connect anything.
The Part That Matters
I want to end on the thing I keep noticing about this week, because I think it says more than any of the product announcements.
In the same seven days that three companies bet on AI that speaks first, five separate teams launched tools to help people manage the AI they already have.
Those two facts belong in the same sentence a lot more often than they get put there. We are being sold more initiative and more oversight simultaneously, by the same industry, in the same news cycle.
I am not cynical about it. I think both products are responses to something real. But it tells you that nobody has figured out the right relationship between a person and a system that acts on its own, and that includes the companies shipping it.
So go in curious and go in small. Give one of these a single lane and thirty days. Count what you acted on.
And if you are like me, if your attention is the scarcest thing you own and interruptions cost you more than they cost other people, do not let anyone talk you into feeling behind for being careful with it. Protecting the ability to go deep on something is not resistance to technology. For some of us it is the entire reason we are able to do good work at all.
The tool should earn its way into your attention. Not the other way around.
P.S. If you try one of these and your action rate is still holding at ninety days, come tell me. That is a data point I genuinely want, and I have not seen anyone publish one yet.
About the author
Jonathan Mast is the founder of White Beard Strategies and writes about using AI to amplify the skill and experience you already have. He runs a 500,000-member community of entrepreneurs learning AI, created the Perfect Prompt Framework, and speaks regularly on practical AI adoption. He is ADHD-wired and protective of his focus, which is why he reads every proactive-AI launch with equal parts interest and suspicion.