AI can drive a computer by itself now. It still flunks more than half the tasks — and that's the useful part.

There's a demo doing the rounds. You type something like "book the team's flights and file the receipt," and an AI takes over the screen — moves the cursor, clicks through the browser, fills the forms, does the whole thing while you watch. Eighteen months ago that was a party trick that fell over on the first pop-up. Now it mostly works. Which is exactly when it gets dangerous.

Because "mostly" is doing a lot of lifting in that sentence, and the demo never shows you the other half.

The number nobody puts on the slide

There's a benchmark called OSWorld that scores these screen-driving agents on real, messy computer tasks — the kind with dialog boxes, logins and things in slightly the wrong place. The best agents now finish somewhere around 44% of those tasks, up from about 14% a couple of years back. In the same window the cost per task has dropped roughly tenfold. Both of those are genuinely big deals; the capability curve is steep and it's still bending up.

But 44% done is also 56% not done. A tool that completes fewer than half of full, multi-step jobs is a remarkable research result and a terrible thing to leave alone with your business. That gap — between "impressive on a benchmark" and "safe to trust unsupervised" — is where almost all the disappointment with AI agents is going to come from this year.

The fix is scope, not faith

Here's what the teams getting real value out of this have quietly worked out: the same agent that face-plants on a sprawling open-ended task nails a narrow, bounded one. "Go run my admin" is a coin toss. "Take this one form, copy these five fields into it, then stop" is close to reliable. The model didn't get smarter between those two sentences — the size of the ask got smaller.

That's the whole trick, and it's unglamorous. Reliability isn't a fixed property of the AI; it climbs fast as you shrink and pin down the job. So the winning move isn't to wait for the agent to get good enough to run everything. It's to carve off the parts that are already in its wheelhouse and leave the rest with a human.

• An agent's success rate isn't one number — it rockets up as the task gets narrower and better defined.
• A boring, well-specified job it repeats a hundred times beats a clever open-ended one it botches on the first try.
• Anywhere it can act on its own, you need to see what it did and undo it. At a 44% hit rate on the big stuff, the undo button isn't a nicety — it's the whole safety model.

What that means if you book jobs and send invoices

You don't need an AI that runs your business. You need one that eats the specific, repetitive, soul-sapping tasks that are the same every time and don't need judgement. The first reply to a 9pm enquiry. The standard quote off a known price list. The "are you free Tuesday" you answer fifty times a week. Bounded, checkable, repeatable — that's the sweet spot the benchmark is quietly pointing at.

The businesses that get burned this year are the ones who believe the demo and hand the agent the open-ended stuff — then spend more time checking and unpicking its work than they ever saved. The ones who come out ahead treat it like a fast, tireless junior who's brilliant at the narrow task and should never be left to run the place.

Where we land on it

This is the exact line we build to, and the benchmark is a good reason we're glad we did. Our agents handle the repetitive first response and the boring 80%; you approve what matters, and you can always see and undo what they did. Not caution for its own sake — just the obvious read of the numbers. A tool that's right less than half the time on the big tasks and nearly always right on the small ones should be pointed at the small ones.

So when the next "AI runs your whole business" demo lands in your feed: believe the demo, distrust the conclusion. The capability is real. The autonomy isn't there yet. Point it at the narrow, boring, repeatable work, keep a hand on the undo button, and you get most of the upside with almost none of the risk. It's not the exciting version. It's the one that actually saves you time this year.