What AI Cannot Do in a Day: The Task I Gave Up On

I gave up on an AI task today, 13 suppliers into a list of 35. The job was to re-check every price we've ever published: 265 dollar figures spread over 26 articles. The fetching part took minutes. What ended it was that I stopped being able to trust the answers coming back.

My position up front: what AI cannot do in a day has almost nothing to do with how hard the task is. It's about whether you can check the work yourself before the day is over.

This is for anyone who keeps reading that AI can do anything in a day and wants a sharper line than "well, it depends." By the end you'll have three questions to ask before you start, which is the difference between a day well spent and a day you write off.

What AI cannot do in a day, in one sentence

AI cannot finish a task in a day when you can't verify the result yourself the same day. Not difficulty. Not creativity. Not some mystical thing about understanding. Verification.

That sounds small until you count how many ordinary jobs fail it. Anything behind a login you don't have. Anything that needs a second reader before it's safe to publish. Anything where the answer comes back in a unit that isn't the one you asked about. All three showed up in a single boring maintenance job today.

The job I quit: re-checking every price we've published

The job was to confirm that every price on our own site is still true, and I got 13 suppliers in before calling it. I picked it because it's the least glamorous task in the building, and those are the ones people assume AI eats for breakfast. We have 32 published articles. Twenty-six of them quote money. A quick count gave 265 dollar figures on the page, 56 of them unique, and across all 32 articles we name 35 different suppliers.

Let's do the math on why that looked like one day of work. There's no writing, no judgment, no taste. Open a pricing page, compare it to what we wrote, flag the difference. Thirty-five pages. A morning, tops.

Here's how the first 13 actually went. Six came back confirmed. Five of them cleanly: Jasper still charges $69 a month, or $59 billed annually. Copy.ai is $29 a month, or $24 billed annually. Suno is still $8 for Pro. Gumloop is still $37. Airtable's Team plan resolved to $20 per user per month on the annual plan, which had been an open question in our notes since 23 July. The sixth was ElevenLabs, and I'll come back to that one.

Then the clean half ran out.

Four pages returned HTTP 403 to anything automated: Canva, Gamma, Midjourney and Perplexity. A 403 is honest, at least. It says no and means it. Those four prices can only be checked by a person sitting in front of a browser, which rhymes with what happened when I handed a day of errands to an AI browser agent: it flew through the finding half of every task, and the finishing half stayed in my own hands.

Three more answered in the wrong currency. Grammarly's plans page served me €12. n8n served €20, €50 and €667. Notion served €9.50 and €19.50 for the two tiers our own notes record as $10 and $20. None of those pages is wrong. They just aren't answering the question I asked, and converting them turns a verified figure into an estimate with an exchange rate baked in.

The moment I stopped trusting the check

The thing that actually ended the job was one page giving me two different answers minutes apart. ElevenLabs, same URL, same session.

The first read came back as this:

Creator - "$11per month" (first month discounted at "$22")

The second read, when I asked for the literal price text instead of a summary, came back as this:

"$22 First month 50% off" and "$11per month"
The regular monthly price is $22, with a promotional first month discounted to $11.

The second one is right. The first one flipped the promotion, and if I'd taken it, we'd have published a sentence pricing ElevenLabs Creator at $11 a month. It would have read exactly as confident as the correct one. That's the part that gets you.

Has the finish line moved before?

The same shape has shown up twice more in the last two weeks, and neither time was the model the problem. Yesterday one of our own runs set up a two-model comparison, wrote the task, built a deliberately messy test file, wrote nine blind checks in advance, and then got this:

Fable 5 requires usage credits. Run /usage-credits to continue or switch models with /model.

The experiment was fine. The gate was shut. Half the comparison simply never ran.

And our Sora 2 test review is called a test review while containing no footage we made, because the consumer app closed on 26 April 2026 and what's left sits behind a paid, logged-in account. That article is good precisely because it says so in its second paragraph, before it reaches its first heading, instead of pretending.

Isn't it just slow, not impossible?

"In a day" is a claim about today. A task that needs a browser I don't have open, a card I'm not putting on file, and a second reader tomorrow is not a one-day task with friction. It's a three-day task with a fast first hour, and the fast first hour is the trap. It feels like progress right up to the point where you notice the finishing half hasn't moved at all.

You could say none of this is a real limit, and you'd have a point. Every one of these jobs finishes eventually. Buy the credits, open a browser, click the currency selector, have a second person read the numbers, and the price audit is done by Thursday.

Three questions to ask before you start

Ask these three before you commit a day to anything, and you'll know which side of the line you're on. This is the checklist I wish I'd run before I started.

1. Can I see the raw source myself? If the answer lives behind a login, a card, or a 403, you're not doing this today. You're waiting for access and calling it work.
2. Does it answer in the unit I need? Euros for a dollar question, eight seconds for a ten-second question, a summary where you needed the literal text. The wrong unit isn't a small gap. It's a different question.
3. Can I check the result without a second person or a second day? If not, the making fits in a day and the finishing doesn't, and only one of those is the job.

Bottom line: one no means pick a different task. Two means you already knew.

My honest take

This is the reason I still believe the one-day thing, not the reason I doubt it. A promise that covers everything covers nothing. The projects worth doing in a day are the ones where you own the whole loop: you can see the source, read the answer, and check it yourself before dinner. That's a smaller set than the internet suggests, and it's a much better one.

So what's on your list this week that fails question three? That's the one to look at first.