Most businesses do not need AI yet
A short, honest test for whether AI is the right tool for your problem, from someone who builds it for a living and turns down the work when it isn't.
I build AI systems for a living, so it’s against my interest to say this: most of the businesses that ask me for AI would get more value from a spreadsheet, a properly indexed database, or two hours spent deleting a process nobody remembers approving.
I say it anyway, because the alternative is taking money for something that won’t work, and that’s a worse business than being honest.
Here’s the test I actually use.
1. Can you describe the problem without the word “AI”?
If the request is “we want to use AI,” there’s no problem yet. There’s a board member who read something. If the request is “our team spends nine hours a week reading inbound emails and routing them to the right person,” now we can talk, because that’s a problem with a measurable shape.
The tell is whether removing the technology from the sentence leaves anything behind.
2. Is the task genuinely fuzzy?
This is the real dividing line, and almost nobody applies it.
AI earns its cost on problems where the rules can’t be written down. Reading intent from a badly written message. Judging whether two records describe the same customer. Summarizing something no two people would summarize identically.
If the rules can be written down, if the answer is “when the invoice is over
$5,000, send it to Dana,” then you want an if statement. It’s faster,
cheaper, correct every single time, and it will never invent a person named
Dana. A language model is a poor substitute for a rule you already know.
3. Do you have the data, and is it any good?
Every AI project is a data project wearing a disguise. Before anything gets built I want to know what you have, where it lives, and how much of it is wrong. That last question is the one that quietly ends projects.
The failure mode is predictable: months of model work, a demo that impresses everyone, and then it meets production data and falls apart, because the training set was the clean export and reality is not the clean export.
4. What happens when it’s wrong?
Not if. These systems are wrong sometimes, by construction, and any vendor implying otherwise is selling something.
So the question is what the wrong answer costs. If it drafts a reply a human approves before sending, being wrong costs a few seconds. If it approves a refund, or files something legal, or tells a customer something binding, the cost of wrong is the whole project, and the system needs to be designed around containing that from day one rather than patched with guardrails at the end.
5. Is there a boring version?
Ask it out loud: what would this look like with no AI at all?
Sometimes the boring version is 80% as good for 5% of the cost, and the right call is to build the boring version, run it for a quarter, and see whether the remaining 20% is worth anything. Sometimes the boring version is genuinely impossible, and that’s exactly when AI is worth every dollar.
You can’t tell which without asking.
Where this leaves you
Three yeses and a real answer to question four, and you have a project worth doing: one with a measurable outcome instead of a press release.
Fewer than that, and the most valuable thing I can do is say so early, while it’s still a conversation and not an invoice.
That’s usually how our first call goes, and I’d rather have it that way.
Written by Dylan Trenck, who builds software, automation, and AI systems for businesses through TrenckTech.