Why does this need a method?
Because AI projects that set off in the wrong direction do not fail immediately. They fail two weeks later, once you have put time into them and it is hard to back out. The four questions cost one minute, and those are exactly the two weeks they save.
Take the specific task and work through them.
1. Do I have the material it can work from?
The model does not know your company's data and knows nothing about today's events. Where it has no information, it guesses — and confidently.
- Yes, if you can hand over the document, the email, the spreadsheet, the recording.
- Yes, if it has search and the answer is public information.
- No, if the answer is in your head, in a colleague's head, or in a system it cannot reach.
If no: do not drop the task. Make the material accessible first. Many an "AI project" really stands or falls on this step.
2. Would I recognise a wrong answer?
This is the most important of the four.
- Yes, if you can see it at a glance: whether a summary is faithful, whether an email's tone is right, whether code runs.
- Yes, if there is something to compare against, or a test that tells you.
- No, if the answer is a number, date, regulation or reference you do not know off by heart either.
If no: either build in verification — a citation to the source, a cross-check on a second run — or put it behind human approval. You must not build automation on output you cannot check.
3. What happens if it gets it wrong?
Not all errors are equal.
- Cheap error: a bad draft you rewrite. A bad idea you discard. Go ahead without hesitation.
- Expensive error: a wrong price in a quote, a customer email sent in error, an incorrect accounting entry, deleted data.
If expensive: the question is not whether the model is reliable, but where you can catch the error before it does damage. An approval step before sending is cheaper than any more powerful model.
4. How many times will it run?
This is not a quality question but an economic one — and it decides half the case.
- Once or a handful of times: do not automate. Open the chat, do it, done. The most expensive thing you can build is a forty-hour automation for a task that comes up twice a year.
- Many times a day: here the investment pays — but here the size of the model and the token count matter. Training or Inference covers that.
- Thousands of times a day: at this point the process needs redesigning, not the task solving.
How to read the result
Four yeses → do it, now.
No on question 1 → sort out data access first. That is often not an AI job at all.
No on question 2 → make it checkable, or put a human at the end. If neither is possible, this is not an AI job.
Expensive error on question 3 → go ahead, but with approval. Do not economise on control at the start.
Few runs on question 4 → do it by hand with AI; do not build a system.
Two common traps
"A better model will sort it out." If the answer to question 1 or 2 is no, switching model does not help — most of the limits are structural, see the list in What AI Is Not Good For. Missing data stays missing; unverifiable stays unverifiable.
"Let's do the whole process at once." A long process has many steps, and each one answers the four questions separately. Typically two or three of them will be perfect AI tasks and some will not be. Break it up and automate only the good parts.
A worked example
"AI should categorise and answer the emails coming into customer service."
Let us take it step by step rather than all at once.
- Categorising. Material: the email, present. Would I spot an error: yes, by sampling. Cost of error: small, it lands in the wrong folder. Frequency: many per day. → Four yeses, automate it.
- Drafting a reply. Material: the email and the knowledge base, present. Would I spot it: yes. Cost of error: medium. Frequency: many per day. → Go ahead, but as a draft.
- Sending the reply. Cost of error: expensive and irreversible. → A human approves, at least for the first few months.
Same task, three different answers. Ask about it as one thing and only one wrong answer is possible.
What to take away
Does it have the material? Would I spot the error? What does a mistake cost? How often does it run? Four questions, one minute — and most failed AI projects could have been filtered out right here.