AI assistants don't know how long their work takes
A new study looked at AI coding assistants and found something simple. They have no sense of time. Their estimate of how long a job takes is structurally wrong, and so is their judgement of their own work. If you're weighing up an AI assistant for your one-person business, that's worth knowing today.
What the researchers found
Researchers asked AI coding assistants like Claude Code and Codex to do something you expect from any co-worker: estimate how long a task will take. They can't. According to The Decoder, both systems are structurally off. With Codex the gap runs up to ten times the actual duration.
There's a second layer. The same assistants rate their own work about 20 percentage points too high. So they tell you two things you can't build on: how long it takes, and how good it is. For tasks that run on their own for a long stretch, that makes oversight hard. You no longer have a gauge you can read.
The study is about programming work. The arithmetic underneath it applies to any work you hand to a machine.
An hour of work you have to check is not an hour saved
This is where time saving with AI usually goes wrong. Someone hears "that'll be done in no time", counts it as profit, and forgets the checking.
The honest sum has three lines:
- The time the work takes off your hands. That's the promise.
- The time you spend reading it back. That line appears in no sales page.
- The chance it has to be redone. On a quote the client has already seen, that costs more than time.
Subtract lines two and three from line one. What's left is your gain. Sometimes that's a lot. Sometimes nothing is left, and then "don't" is the right answer.
Why I never ask you to trust blindly
People ask me whether an AI assistant is reliable. My answer isn't "yes", it's "you have to be able to check it".
That's why I work the way I do. When I make something for you, a quote, an email, a piece of text for your site, you see it before your client does. You read it back and you tell me it needs to be different. That isn't an awkward extra step I tack on. It's how this kind of work stays good.
A system that rates its own work too highly doesn't improve by getting more independence. It improves when someone who knows the trade looks at it. That someone is you.
What you can do with this tomorrow
Three things, whether you shoot photos, run a venue or keep a coaching practice.
Ask how you check it, every time AI is promised to you. If the answer stays vague, so does the time saving. A tool you can't check is a tool you can't trust.
Start with work where you'd spot the mistake. A quote in your own prices, an email to a client you know, a text about your own trade. You'll see within thirty seconds whether it's right. Don't start with your bookkeeping or your contracts.
Count the checking before you sign anything. List what you pay now for tools and outsourced work, and add honestly how much time reviewing costs you. Only then do you know what something really saves you.
You do that sum once. After that you decide with numbers instead of a feeling.
Frequently asked questions
Does this mean I should leave AI alone?
No. It means you have to do the arithmetic on the time saving yourself. Work you can check quickly gives you real time back. Work you have to review for hours gives you nothing, however fast it seems to arrive.
How do I know the work I get back is any good?
By reading it before it goes out the door. So pick tasks where you are the expert first: your own prices, your own clients, your own tone. There you spot a mistake straight away.
Why is Mia telling me this if it argues against her own product?
Because the pretty half is no use to you on its own. If the sum doesn't work out in your favour, you hear that from me before you pay. Sometimes my answer is: don't.
