Every consultant is now an AI consultant. LinkedIn is one giant “AI transformation” banner. Every SaaS pitch opens with “AI-powered.” Every vendor proposal has a diagram with “AI agent” in a box.
If you're the owner of a $2M–$15M business, you're being pitched something AI-adjacent every week. And you can't tell the real builders from the hype merchants from inside the pitch. The language all sounds the same.
Here are five questions that work as a real-time filter. If the person across the table from you can't answer all five specifically, pass.
1. “Show me the production system.”
Not the demo. Not the slide. Not the video walkthrough. The actual URL of something running in production for a paying customer, right now, that you can click into and see data.
Real builders have links they can drop in a Slack channel in 10 seconds. Hype merchants have “case studies” (read: brochures) and “in a customer deployment” language (read: vapor).
2. “What happens when the AI is wrong?”
Every real AI system has this answer. Every fake one doesn't.
The answer should include: confidence thresholds (below this score, it escalates to human), a human-in-the-loop queue for edge cases, guardrails on dangerous actions (like sending money or deleting data), an audit log that shows which decisions were AI-made vs. human-approved, and a kill switch that turns the whole thing off.
If the answer is “the AI doesn't really get things wrong in practice,” run. They either don't have a production system or don't know how it works.
3. “What's the cost structure — theirs and mine?”
Real AI systems have two cost layers: the platform layer (the vendor's margin) and the inference layer (what the AI models actually charge per call). A real builder can tell you both.
Example of a real answer: “The platform is $500/month. Inference averages $0.004 per patient interaction. For your expected volume of 8,000 interactions/month, total monthly cost is ~$532. The cost-per-result works out to 1.4% of collections.”
Example of a hype answer: “We have flexible pricing based on value delivered. Let's have another meeting to discuss.”
4. “What's the worst result you've ever measured?”
This is the single most revealing question in an AI sales conversation. Real builders have answers. They've seen their system flop on a customer segment, a data mix, a seasonal pattern. They can tell you what they learned and what they changed.
Hype merchants have never measured a bad result because they've never measured anything. Every “customer success story” is a case study of a win. The wins are real — the measurement framework isn't.
5. “What does the handoff look like when I don't need you anymore?”
You want to hear: “Here's how you export your data. Here's how long it takes. Here's what happens to the AI agents — you can either keep licensing them from us or migrate to a different vendor, and we'll help you do either.”
You don't want to hear: “Why would you ever need to leave?”
Lock-in is a strategic choice on their end. If they haven't thought about how a customer leaves, they haven't thought about what happens when their system stops working for you.
The meta-question
Underneath all five of these questions is a single test: is the person across the table a builder who happens to be selling, or a seller who happens to be selling AI?
Builders can answer the five questions because they've shipped systems and watched them run. Sellers can't because there's nothing behind the slide deck.
“AI translator” is the thing I find myself doing in most sales conversations these days. Someone's been pitched an AI solution and wants a second opinion. I run the five questions against the vendor's proposal. Most proposals fail on question 1.
If you're being pitched something and want a second opinion, book a 20-minute call. No pitch on my end — I'll just tell you if what you're being sold is real.