A proactive assistant sounds wonderful until you remember that every app on your phone already has an opinion about what needs your attention.
Somewhere between the delivery update and the software asking how your experience was, the word “important” lost a fair amount of credibility.
Claude Duet enters that crowded room with a promising job description: notice what needs you and help move the work along. Anthropic's product description Our interest is in the judgement behind the interruption. A business owner can have plenty of information and still be short of a useful next move.
A warning should arrive with something useful
Imagine you have a customer call this afternoon. An assistant tells you there is an unresolved issue in the account.
Fine. Now you have to find the issue, read the history, work out whether it matters for this call and decide what to do. The message has pointed at the work. Much of the work remains.
A stronger interruption would bring the relevant source, explain the consequence and prepare a next step. You could then spend your attention on the decision itself.
| What arrives | What you still have to do |
|---|---|
| There is an open question on this account. | Find it, establish what changed and decide whether it matters today. |
| The customer asked to change Friday's delivery date. The latest reply is attached; a draft response is ready for review before your 3pm call. | Check the evidence and decide whether the proposed date works. |
Editorial examples of alert quality, not messages captured from Duet.
The second message is longer. It may still cost you less attention because it removes several searches. Brevity alone is a poor measure of helpfulness.
Quiet has to be earned too
There is an awkward counterpoint. An assistant that rarely messages you could have excellent judgement. It could also be missing everything.
You need to be able to check the silence. Pick a few known events from the day and see whether the assistant handled them, deferred them sensibly or overlooked them. Otherwise you are evaluating the comfort of an empty notification tray.
For an owner, different events deserve different treatment. A routine receipt can wait. A customer decision that blocks today's delivery may justify breaking concentration. A vague weekly trend probably belongs in a review you chose to open.
Those are business priorities, not universal notification settings. An agency and a shop may care about different things at exactly the same time of day.
- Something meaningful changes if you leave this until later.
- The alert points to the record behind its claim.
- It is clear why your involvement is needed and what you can do next.
Tap a box as you go. Checks reset when you leave this page.
Approval takes attention as well
Duet's announced approval boundaries are encouraging. Anthropic's launch explanation They also make the quality of the review moment important.
Think about a proposed email. A useful review shows the recipient, the actual message and any commitment it makes. If the date depends on another conversation, you need a way to inspect that context. A friendly “shall I handle it?” leaves too much work hidden behind the button.
This is why we would watch correction effort alongside task completion. Producing ten drafts does not tell you how much of your afternoon they consumed. One accurate, well-supported draft may be the more valuable contribution.
The test we would care about
Launch demos are good at showing an assistant doing something. Everyday usefulness also depends on the occasions when it waits, asks a precise question or admits it cannot establish a fact.
Our view is that proactive assistants will have to earn trust through that ordinary judgement. The evidence will be in the work people keep using after the novelty wears off.
We have not tested Duet, so this is a buying criterion rather than a verdict on its performance. What makes the launch interesting is the possibility of reducing the remembering and gathering that surrounds a task. That would be valuable even when the task itself is quite boring.
If you get access, our evaluation plan gives you a way to record useful catches, missed work and review effort. Start with something you already understand. Your inbox has had enough speculative management.
