If you run a business that sells a physical product, you’ve probably gotten the pitch. Maybe it came by email, or from a consultant at a trade show, or from your cousin who just discovered ChatGPT. The pitch is always some version of: “You need to add AI.”
Add it to what? Nobody says. That’s the problem.
I build AI/ML systems for a living, so I’m supposed to be the guy who tells you AI is the answer. Most of the time it isn’t. When it is, it’s usually a much smaller, duller project than anyone is pitching.
Why the pitches fail
1. They start with the technology instead of the decision.
“We’ll build you an AI platform” is a solution looking for a problem. A good project starts with a sentence like: “We decide X over and over, mostly by gut feel, and we’re probably getting it wrong a lot.” If nobody can finish that sentence, there’s nothing to build yet.
2. They’re scoped like a moonshot.
Six months, a big budget, “transformation.” Then the first usable output shows up at month five, and by then the business has changed, the champion has moved on, and nobody trusts it. Big scopes fail slowly and expensively.
3. They’re black boxes.
A model spits out a score, and the person who has to act on it has no idea why. So they ignore it and go with their gut, which is what they did before. If a salesperson can’t see the reasoning, the tool doesn’t get used. I’ve seen very smart models fail this way.
4. They need data you don’t have yet.
Many pitches quietly assume you have years of clean, labeled history. Most small and mid-size companies don’t, and that’s normal. If the project only works after you’ve collected two more years of data, it’s not a project yet.
5. Nobody defined “better.”
If you can’t say what success looks like in advance (faster, cheaper, more leads, fewer mistakes), you can’t tell whether it worked. It just becomes a thing you paid for.
What a good scoped MVP looks like
Here’s the shape I aim for.
One decision. Not “improve sales.” Something like “which prospects do we call next?” or “which support tickets need a human right now?” A decision you make dozens or hundreds of times is ideal, because small improvements add up.
Public or already-existing data. If I need your internal systems to get started, the project is already too heavy. The best first builds run on data that’s already available, so there’s a real result before anyone signs off on integrations.
A transparent v0. Start with a simple weighted score, not a fancy trained model. Every recommendation comes with its reasons attached: “gap in your coverage,” “customers there are already complaining about the problem you solve,” “looks like your best existing partner.” A non-technical person can read that, argue with it, and tell me where I’m wrong. Those arguments are the most valuable input I get.
Something you can hold in your hands. A ranked list or a one-page report, not a slide deck. A real output makes a better conversation than a concept does.
A clear path to v1. Once people use the v0 and you start recording outcomes (who converted, who didn’t, who was a bad fit), then you train something on real results. The trained model is the second step. Most pitches try to sell it as the first.
- 01One decisionMade dozens or hundreds of times
- 02Existing dataPublic or already on hand
- 03Transparent v0A weighted score with reasons attached
- 04Record outcomesWho converted, who didn’t, who was a bad fit
- 05Trained v1Built on real results
A quick example
Say you sell a specialized product through a network of dealers, and you want more dealers. Today, “who should we approach next?” is answered by inbound interest and instinct.
A scoped MVP might pull public listings for every likely prospect in one or two states and flag where your existing coverage has gaps. It could also read public reviews of those prospects and pick out the ones whose own customers are already describing the problem your product solves. Then it ranks everyone with a plain-English reason attached to each name.
That’s a few weeks of work, not a few quarters. It uses no internal data, so there’s nothing to approve before starting. Even where it’s wrong, it shows you something about your market. And you can judge it quickly, because you probably already know a few of the prospects it ranks and can check whether it agrees with you.
The test I’d use
Before spending anything on AI, ask:
- Is there one specific decision we make repeatedly?
- Can we start with data that already exists?
- Will the output come with reasons a human can check?
- Can we see something real in weeks, not months?
- Is a wrong answer cheap and catchable?
If you can’t say yes to most of these, wait on the AI. A spreadsheet and a phone call may genuinely serve you better, and I’d rather tell you that than build you something shiny.
If you do have a decision you’re making on gut feel and suspect there’s a better way, I’m happy to take a look. No deck, just a conversation about whether it’s worth building.
The trained model is the second step. Most pitches try to sell it as the first.