Most AI projects don't fail at the model layer. They fail because nobody asked three questions early enough.

I've reviewed ~40 AI initiatives across healthcare, fintech, and SaaS in the last two years. The ones that shipped — and kept running in production — had one thing in common: the team could answer these three questions in the first week, not the first quarter.

1. "What decision does this replace?"

If the answer is "it helps people make better decisions," you don't have a spec — you have a hope. Every AI system that actually ships replaces a specific decision that a specific person currently makes manually.

Bad: "We want AI to help our sales team prioritize leads."

Good: "When a new inbound lead hits the CRM, Sarah currently spends 4 minutes reading the company's website and LinkedIn to decide hot/warm/cold. We want to automate that classification."

The good version tells you exactly what data you need, what the output schema looks like, and what "wrong" means. The bad version gets you a 6-month discovery phase.

2. "What happens when it's wrong?"

This is the question that separates systems that run in production from demos that get screenshots in a board deck.

Every AI system is wrong sometimes. The question is: what's the blast radius?

• If a lead scoring model miscategorizes a $500 MRR lead → a rep wastes 4 minutes. Acceptable.

• If a clinical triage model miscategorizes a symptom cluster → someone gets the wrong care pathway. Not acceptable without a human checkpoint.

The answer to this question determines your architecture. It tells you whether you need a human-in-the-loop, a confidence threshold, a fallback path, or just a good logging pipeline. If your team hasn't answered it, they're building a demo, not a system.

3. "Where does the data already live?"

The most expensive sentence in AI is "we'll need to build a data pipeline."

Before you spec a single agent or pick a model, inventory what you already have:

• Is it in a database you control, or locked in a SaaS tool's API?

• Is it structured or do you need extraction?

• How stale is it — real-time, daily batch, or "someone updates a spreadsheet on Fridays"?

If 80% of the data you need already exists in systems you own, you have a 90-day project. If it doesn't, you have a data engineering project first and an AI project second. Knowing which one you're signing up for saves you six figures.

---

These three questions aren't clever. They're boring. That's the point — boring questions asked early beat brilliant architecture imposed late.

If you're scoping an AI initiative right now and want a second opinion, the Diagnostic Intensive is a half-day deep dive where I'll tell you what's real, what's not, and hand you a strategy memo in 72 hours.