The AI that keeps working after you shut the laptop.
The demos doing the rounds this month are showing off stamina. A system takes a job across hundreds of small steps over several hours, with nobody at the keyboard, and has something finished at the end of it. That is a real change. For a plumber or a salon it matters because the jobs it now covers are the boring repetitive ones that only ever got done when you were sat at the desk anyway.
So here is the useful version, with the demo taken out.
What actually changed
Until recently, AI was a good conversation. You asked, it answered, you copied the answer somewhere, and the next step was yours. Every chain of work had you in the middle of it.
What is being shown off now is a machine working a shift. Given a goal, it takes a step, looks at what happened, decides the next step, and keeps going for a long time. It came to public attention through a security test rather than a product launch. Researchers watched a pre-release system push through an enormous chain of actions on its own, far past the point anyone was supervising it (Trending Topics, July 2026). That was a lab, not a business. The underlying ability is the same one, though, and it is why the word that keeps coming up is unattended.
The job it can take off your desk
Long work with a clear finish line. Patient rather than clever.
An unpaid invoice is the clean example. Fourteen small steps, none of them one task: a polite reminder, a wait, a check of whether the money landed, a second reminder in a different tone, a note to you if it goes past a line. Nothing there is hard. All of it dies because it takes a fortnight and you have jobs on. Same shape as an enquiry that went quiet on the Tuesday, or a rescheduled booking chain, or a month of paperwork that needs sorting rather than thinking about.
That is the real move. Work that used to need you present now only needs you at the end of it. We wrote a fuller list of the jobs worth handing over first in 10 jobs to hand to automation first.
What it still gets wrong
Two failures, and both get worse the longer it runs.
The first is that it follows a bad brief faithfully. A person who misunderstands you stops after ten minutes and asks. A system working unattended does not. It commits to your misunderstanding and spends three hours being consistently wrong, so a small error at the start turns into a lot of work to undo.
The second is that, left alone, it does not know your business. Raw AI has no idea that you stopped doing bathroom fits in April, that the Tuesday slot is blocked, or that you never say "reach out". So it writes in nobody's voice and occasionally tells a customer something untrue. That is a missing memory rather than a bug in the model, which is a different problem to solve. We took it apart in One Brain, never off the clock.
The fix for both is structural. Give the work a memory of your business so it starts from the truth, and put a gate in front of anything that leaves the building so a wrong turn stops at your desk rather than in a customer's inbox.
What the crew already does
All of it, which is the honest headline: there is nothing here for you to go and set up.
The crew runs on a schedule rather than waiting to be asked, so the chasing and the follow-ups happen on the Wednesday whether or not you open anything. They share one Brain, so what you told the accounts mate about your terms is known by the mate writing the reminder, and a correction you make once sticks. Outward-facing work follows draft, review, send. You approve it before a customer sees it. The activity log shows what ran, when, and what it produced, so "it worked unattended" does not mean "you have no idea what it did".
None of those four things is the model. They are the rails around it. A rail is boring to demonstrate, and it is the entire difference between AI that works a shift and AI that quietly makes a mess for three hours.
The other half of the news you can ignore
The rest of this month's coverage is about which system is fastest, which is cheapest to run, and which one has slipped. Google has told investors to expect new versions at close to a monthly pace, and analysts quoted by InfoWorld are blunt that this pushes real testing and version-management cost onto whoever builds on top. Another maker is now pitching its top system as an everyday driver rather than a special occasion (Gizmodo).
None of that is a decision you should be making. You did not choose an engine when you bought a van. The right arrangement is that whoever runs your crew absorbs those changes, and the only thing you notice is that the work carries on. If a product ever asks you to pick, or goes quiet because something upstream moved, that is the product's problem being handed to you.
What to do with this
Nothing, if you already run a crew that works this way. If you are weighing one up, skip "can it work on its own", because soon everything will claim that. Ask what happens when it works on its own and gets it wrong. Where does the mistake stop, what did it know about your business before it started, and can you see afterwards what it actually did?
Answer those and you can tell a shift from a demo. If you want to run that test properly before paying anyone, we wrote it up in how to test an AI team before you pay.