Most AI you get pitched is rules in a costume. A founder sends me a deck, the word AI is on every slide, and underneath it is a pile of if-then logic that any junior dev could have shipped in a week. The price tag, of course, is not junior-dev money. So before I sign anything with AI on the label, I run it through three questions. They take five minutes and they have saved me from a lot of expensive theater.
Is it machine learning, or rules wearing a costume?
The first question is the cheapest filter you have. If it is real machine learning, it will not fail the way normal software fails. Normal software breaks where you told it to break. ML breaks in different ways, with different failure modes, on inputs nobody thought to test. That is not a flaw you can patch away. It is the nature of the thing.
So the real question is not how clever the model sounds. It is fit. Does this thing fit a problem where the messy, probabilistic behavior of ML is worth the trouble, or is it dressed up to justify a number on a contract. As a founder, that reframes the whole conversation. You stop asking is it AI and start asking does ML actually belong here. If the honest answer is that plain rules would do the job, you are buying hype, and you are paying for the costume.
What is labeled, what is not, and who pays for the labels?
Here is the line everyone gets wrong. People sort AI by the input. Audio, text, images, as if the file type decides the approach. It does not. The line that actually matters is labeled versus unlabeled data, and how the inputs arrive. That single distinction tells you more about cost and risk than any demo.
Put real numbers on it. If you want supervised vision, the rule of thumb is brutal. Around 5,000 labeled examples per category to get decent performance. Read that as a founder, not an engineer. Every category is a labeling bill, paid in human hours, before the model is worth anything. Nobody sells you that line item up front, but it is the line item that decides whether the project is real or a slide.
Unsupervised learning softens the labeling problem, because it can work with more available unlabeled data. It does not make the human disappear. You still need people to validate the outputs, or you are trusting a system nobody checked. So the two questions to ask the vendor are plain. What is labeled here, and who pays for the labels. No labels, no supervised model. If they cannot answer, you have your answer.
Where does the learning happen, in a dataset or in the world?
The third question separates the safe projects from the ones that can hurt you. Some systems learn from a fixed dataset. Reinforcement learning is different. It collects its data by acting. The rewards are delayed and partial, and the good move is not labeled step by step. The system has to try things and find out.
That is exactly why reinforcement learning shines in games and struggles in real life. In a game you can speed time up, run a million rounds overnight, and a mistake costs nothing. In your business, time runs at one second per second, and failure has consequences a customer feels. So the rule writes itself. No safe sandbox, no reinforcement learning in production. If a vendor wants their agent to learn by acting on your live operation, ask where the sandbox is. If there is not one, the answer is not yet.
Why the winners are not the teams with the fanciest slides
The map underneath all this kills the biggest misunderstanding in the room. The data type does not choose your approach. Labeling does. Once that lands, the rest of the market makes sense. Open source, pretrained models, and frameworks already made it possible for non experts to build useful systems. The secret algorithm is not the moat anymore.
So the edge moved. It sits in data collection and fast application, not in a clever algorithm you keep in a vault. The team that wins is the one that gets clean data and ships fast, not the one with the prettiest deck. And the compute reality is unforgiving for anyone who wants to play at the frontier. HD images can mean roughly 40x more compute than classic ImageNet sized inputs. Training a frontier NLP model can run into clusters like 1,024 A100 GPUs and about 25 million dollars just for the hardware in the example used. For most founders, that math is the clearest reason to buy or build on what exists and compete on data and speed, not to chase a moonshot you cannot fund.
The founder takeaway
Steal the three rules and use them as a gate. No labels, no supervised model. No validation loop, no unsupervised output you can trust. No safe sandbox, no reinforcement learning in production. Run any AI pitch through those three lines and most of the hype falls out on its own, before it reaches your bank account. The point is not to be the person who knows the most about machine learning. It is to be the founder who does not pay machine-learning prices for rules in a costume.
If you want an outside read before you write a check, start with your own footing. A free AI and website readiness audit at readiness.ai4.sale shows you where your business actually stands for AI and where a vendor would find soft spots, so you walk into the next pitch asking the three questions from a position of knowing your own numbers. It is the cheapest first move there is, and it costs you nothing.