
Radar / Personal agents · Everyday AI
Can AI take everyday admin off your hands?
AI assistants want to take over the little jobs that eat your day. We're following how much work they finish, and how much they leave with you.
Imagine asking an assistant to sort out a missing delivery. It finds the order, checks the tracking, contacts the shop and comes back with a replacement date. You answer one question about which day you'll be home.
That's an appealing trade. You still make the decision, but someone else handles the chasing.
It's also a useful way to judge this new crop of AI assistants. Writing a polite complaint is one part of the job. Finding the right order, getting through the shop's website and following up when nobody answers are the rest.
This is what we're putting on our Radar: can you hand over an errand without becoming the assistant's supervisor?
What you can hand over
These products give AI access to tools and, in some cases, a computer of its own. That lets it work across websites and connected accounts. You describe the outcome; the assistant works through the steps.
Grok Bot has a persistent cloud computer, with a browser, files and connected tools. OpenAI's Dots can keep working in the cloud while your computer is off. Hark Pro calls its browser-operated computer Handoff and also offers direct connections to services.
The familiar chat apps are moving into this territory too. Gemini can carry out tasks using connected apps and a local or remote browser. Its remote work can continue after you close your device, though availability depends on your plan and country. Claude offers cloud tasks through what was called Cowork, now gradually becoming part of the main Claude experience. Cloud work can continue with your laptop closed, but access to your local files and browser still needs your connected computer.
Manus offers automations that start when something happens, such as an email arriving or a calendar changing. It also sells a dedicated cloud computer for ongoing work. Meta's Muse describes work across connected apps, with approvals for actions such as sending email and making purchases.
Those differences matter. An assistant that needs your laptop awake may suit a job you oversee during the afternoon. One that can keep working elsewhere is more useful for an errand that stretches into tomorrow. Neither arrangement, by itself, tells us whether the job gets finished.
A schedule can start that work at a particular time. The bigger promise is what happens after it starts. An assistant handling a delivery problem would need to notice a reply, understand what the shop offered and return to the same unfinished errand. A daily tracking summary would only cover part of it.
There are encouraging examples. In a September 14 TechRadar account, Eric Hal Schwartz described Muse finding an unopened utility-company email about a solar-panel inspection. He chose a date and the assistant prepared a reply for approval. That is a small, recognisable piece of admin taken off someone's hands. His account establishes a prepared reply, not a completed appointment.
The work that comes back to you
The awkward parts are often ordinary: a login expires, a website refuses access, an instruction goes missing.
OpenAI's connection guide acknowledges that some sites block cloud browsers or require extra verification. Moving work to your own computer is a possible alternative, but that computer must be available, and its browser session is separate.
Grok Bot's approval guide describes another practical limit. An approval request from unattended work can expire after about ten minutes. The action then does not run; finishing the task requires another attempt. Asking before acting is valuable. Knowing that a job has stopped matters too.
Reusable instructions can also fail to travel. In a Grok Bot support thread, staff acknowledged that instructions visible in a shared template were missing from the imported bot. An October 12 reply said the issue was still being tracked. Later user replies continued to report it. We haven't reproduced the fault or established its present scope, but the consequence is easy to understand: someone thinks they've installed a ready-made assistant, then has to rebuild part of its instructions.
That is the cost we're watching alongside the subscription price. How often do you have to rescue the work?
Where our watch starts
These are documented capabilities and reporting questions, not results from a matched comparison.
Grok Bot
Cloud work, reusable instructions and rules for approving actions.
What we want to see: A repeatable task that keeps its instructions and clearly reports an expired approval.
Dots
Ongoing work using its own computer and permitted app connections.
What we want to see: One errand resumed after a delayed reply, with a clear record of what finished and what needed help.
Hark Pro
A continuing conversation, remembered preferences and browser work through Handoff.
What we want to see: An independently documented errand from request to result, including interruptions.
Gemini
Tasks using connected apps and a local or remote browser.
What we want to see: An errand completed through the remote browser, including any sign-in interruptions and regional restrictions.
Claude
Cloud tasks alongside chat, with a connected computer needed for local tools.
What we want to see: A finished admin task with a clear account of which steps kept running and which needed the owner's computer.
Manus
Work triggered by new emails or other events, with a dedicated cloud computer available.
What we want to see: A recurring job that handles a changed input correctly and leaves a useful record of its actions.
Muse
Everyday and small-business work across connected apps.
What we want to see: A completed task with a receipt or changed record, plus the time its owner spent setting up and correcting it.
Two related approaches belong on the watch too. Perplexity Comet puts an assistant inside a browser for jobs such as shopping, email and research. Lindy focuses on office work across connected tools, including inboxes, calendars and scheduled tasks. We'll follow what people accomplish with them without assuming every assistant offers the same kind of unattended work.
What would move this story forward
We'll follow three ordinary jobs: arranging an appointment, chasing a missing order and collecting receipts for an expense claim. Those are proposed reporting tasks, not tests we've already run.
For each, the useful record includes the result, the elapsed time, the person's own time and any charges. A draft reply, a sent message and a confirmed booking are different outcomes. We'd also want to see what happens on the second and third attempt, when the novelty has worn off.
Our first source review supports the appeal of delegating small chores. It doesn't establish which assistant does them most reliably. We haven't run a matched comparison, and the outside accounts we've read don't provide one either.
The milestone worth coming back for is easy to recognise: you ask once, step away, and return to a finished errand with something to show for it.
Watch opened October 21, 2026. This first edition establishes the baseline. Future updates will name what changed, link the evidence and explain whether it changes what you can reasonably hand over.
Sources
- Grok Botdocs.x.ai
- Dotslearn.chatgpt.com
- Hark Prohark.com
- Geminisupport.google.com
- Claudesupport.claude.com
- Manusmanus.im
- Meta's Musemuse.ai
- TechRadar accountwww.techradar.com
- connection guidelearn.chatgpt.com
- approval guidedocs.x.ai
- Grok Bot support threadforum.cursor.com
- Perplexity Cometwww.perplexity.ai
- Lindywww.lindy.ai