There's one habit that costs people more than any other when they start using AI seriously, and it isn't timidity. It's handing the AI work that a plain, boring, free tool would have done perfectly.
Here's the whole thing in one sentence:
If you can imagine a tool that would produce the answer, the answer is fixed — and you should be asking for the tool, not the AI.
That's the heuristic. The rest of this is learning to feel it, because the cases where it fires are not the ones you'd expect.
The wrong map most of us carry
Most people hold a simple equation. Unpredictable equals AI; predictable equals not-AI. Randomness on one side, ordinary computers on the other.
It's wrong in both directions, and the two places it's wrong are exactly where the interesting decisions live.
Picture four boxes. Across the top: does the work look like AI, or look like a tool? Down the side: is the output fixed for a given input, or does it legitimately vary?
Two of those boxes are boring. Everyday tools are predictable and look like tools — a spreadsheet adding up a column. AI's real home is work that varies and looks like it needs a mind — drafting an email in a warmer tone. Nobody gets those wrong.
The other two are the surprises. One of them is where the money goes.
Surprise one: it looks like magic, but it's a tool
Work that feels like it needs a brain, and doesn't.
The best example is reading text out of a picture. It feels like magic — you hand over a photo of a page and get the words back. But the same image in gives the same text out, every single time. There's a free tool that does it on your own machine: Tesseract, which is open source and runs offline with no account and no cost per page. (Checked on 3 September 2026: Apache-2.0 licensed, runs locally. That reading hasn't been repeated since, so treat it as true as at that date.)
And here's the nuance that trips people up. Modern text recognition has neural networks inside it. "It uses machine learning" is not the test. The test is whether the output is fixed for the input, and whether a tool exists that produces it.
Once you see the pattern you can't unsee it:
- Reading a barcode or a QR code. Looks like vision; it's a decoder.
- "When and where was this photo taken?" Sounds like a question for an AI. It's a field stored inside the file.
- "What changed between these two versions of a document?" Feels like reading comprehension. A diff tool answers it identically forever.
- "Are these two files really the same?" Compare their checksums.
- "What's 90 days before the hearing?" There is exactly one answer.
- Sorting, counting, de-duplicating, pulling every email address out of a folder. Pattern matching, not reading.
- Resizing images, converting audio, merging PDFs. Tools have done this for decades.
- Rule-based spelling and grammar checking. LanguageTool's core is a set of hand-written pattern rules: same text, same flags. (Checked on 3 September 2026 — the hosted product layers AI on top, which is why this says "core". Also not re-checked since.)
Every one of those is good practice as a tool. Ask an AI to do them and you pay three times over: in money, in time, and in the chance that the answer is different tomorrow.
Surprise two: randomness was always here
Now the opposite box — work that's genuinely unpredictable and has nothing whatever to do with AI. This one is mostly good practice, because the randomness is the entire point.
- Shuffle play. A predictable shuffle is just a playlist.
- Dice, lotteries, dealing cards. The fairness is the randomness.
- Random sampling in a survey or an audit. Check every tenth item and people learn to game the tenth item. Random is the honest version.
- A/B testing. Who sees which version has to be random, or the result is worthless.
- Generating a cryptographic key. The unpredictability is the security; a predictable key is a broken one.
- Fuzz testing — throwing random junk at software to find crashes. Random reaches the cases nobody would think to write.
- Monte Carlo simulation — rolling random numbers thousands of times to estimate something you can't calculate directly, like project risk.
So "unpredictable" was never a synonym for "AI". Computers have been rolling dice on purpose for seventy years.
The expensive box: fixed work handed to AI
This is the one the whole heuristic exists for. Work with exactly one right answer, given to a system that charges you for judgement and may answer slightly differently next time.
- Asking an AI to add up numbers or count words. A calculator's job, done worse.
- Alphabetising a list. One command. Zero variance.
- Converting a data file to a spreadsheet, or reformatting a table. A script's job — and an AI can quietly drop a row, or invent one.
- Transcribing the same screenshot over and over. Read it once with a tool. An AI may read it slightly differently each time.
- Pulling every phone number out of a document. That's a pattern match.
All bad practice. Not because AI can't do them — it mostly can — but because you're paying for judgement where there is nothing to judge.
And now the move that turns this entire box around, which is the most useful thing here:
Have the AI build the tool. Once.
Ask it to write the script, the pattern, the conversion command. Spend its judgement exactly once, on making the thing. Then run the thing forever, for free, with the same answer every time.
That's AI where it belongs — on the one-off act of creating something reliable — and the tool where it belongs, on every run after that.
Where AI genuinely belongs
For balance, the box nobody argues about, so you can feel the heuristic fire in reverse. Try to imagine a tool that produces the answer:
- Draft this email in a warmer tone.
- Summarise this report — the version that decides what matters.
- Name this feature.
- Weigh these two approaches and tell me the trade-off.
No tool you can picture does any of those. That's the tell. The work needs judgement, so it needs AI.
And one more that closes the loop: deciding which tool to use is itself judgement. So the honest division of labour is that AI chooses and builds, and tools run.
So
The heuristic once more, and then go and apply it to the last five things you asked an AI to do.
If you can imagine a tool that would produce the answer, ask for the tool. If you can't, that's the work AI is for. And when the tool doesn't exist yet — ask the AI to build it, once.
The big waste isn't asking AI for too little. It's spending it on work that never needed it.
If you'd like the same kind of structure applied to the building itself rather than to one task, that's what Throughliner is.