Why is this worth reading?
Because AI projects do not fail because the model is weak. They fail because it was given a task it is structurally unsuited to — and that only became apparent six months later.
The list below is not pessimism. At each point I say whether it is fixable, because the two cases have to be handled very differently.
1. It does not know what it has no information about
The model's training data stops at a given date, and it has never seen your company's internal data. Ask it about a recent event or about your own customers and it guesses.
Fixable: yes. Give it search, or hand over the material. This is the most easily remedied limit of all.
2. It does not know what it does not know
There is no internal signal saying "this bit is uncertain". An invented answer and an accurate one sound equally confident.
Fixable: partly. You can ask for references into the material you supplied, and you can allow it to say no. Why AI Makes Things Up gives five concrete methods. But it never disappears entirely.
3. It does not calculate reliably
Working in its head it makes mistakes, and increasingly so over longer chains of operations.
Fixable: yes, and completely. Given a tool — a calculator, runnable code, a spreadsheet — it stops estimating and starts computing. If numbers matter, always make it work with a tool rather than from its head.
4. It is not consistent
Ask the same question twice and you can get two different answers. That is part of how it works, not a fault.
Fixable: partly. Predictability is adjustable, and asking for a fixed format improves things a lot — that is what Prompt, System Prompt, Temperature is about. Perfect repeatability, though, is not something to expect.
5. It follows long, complex rule sets badly
It will follow five rules reliably. It will not follow forty with exceptions and precedences between them. It drops a few — and does not mention it.
Fixable: partly, by design. Break it up: rather than one prompt handling forty rules, use several steps each handling fewer. Anything that can be written down precisely as a rule is better written as a program.
6. It has no experience
It has never sat with a client, never run a project, never failed at anything. What it "knows" is other people's written-down experience — the average of it, not the best of it.
Fixable: no. This is structural. It is why it gives textbook-ish, cautious, middle-of-the-road suggestions to precisely those questions where the good answer is the unusual one.
7. It bears no responsibility
Nothing happens to it if it is wrong. You cannot cite it in an audit and you cannot sue it.
Fixable: no, and it does not need to be. You just have to know: the responsibility stays where it was. If a quote written by AI contains the wrong number, it is your quote.
8. It does not grasp consequences
It can articulate that a decision is risky. But it does not feel that a sent email is irretrievable, or that losing a customer lasts ten years.
Fixable: no — it has to be managed. That is why we put a human in front of the irreversible steps.
9. It has no taste and no position
Where there is no objectively right answer — tone, sense of proportion, what to leave out — the model returns the average. And the average is rarely distinctive.
Fixable: partly. Show it your own earlier work as a sample and it gets considerably closer. But it does not take the judgement off your shoulders.
What this means for companies
Surveys over the past few years consistently show the same thing: the great majority of companies measure no meaningful effect on results from introducing AI, while individual employees find it useful. The two are not in contradiction.
The explanation is precisely the list above: most individual tasks fall into the "verifiable, I can see immediately if it is wrong" category, where AI is strong. Company processes, by contrast, are long, rule-heavy and consequential — and there all the limits above show up at once.
The bottleneck is not the technology, it is that the process has to be redesigned around it, rather than the tool slotted into the old one.
What to take away
Two questions that fold the nine above into one, and cost ten seconds before you decide: do I have the material it can work from? and would I recognise a wrong answer?
If the answer to either is no, the question is not which model to pick. There, the task has to be reshaped — or solved without AI. That is what Can AI Do This Task? is about.