Every week we talk to a team that’s excited about an AI idea and unsure whether it’s real. The good news: you can usually tell within an afternoon. The trick is to stop asking “can AI do this?” and start asking “do we have what this needs to work?”
Here are the five questions we run before a single line of code.
1. Is there a decision or action at the end of it?
AI that produces an interesting output nobody acts on is a demo, not a product. The strongest projects end in a concrete action: a ticket gets routed, a transaction gets flagged, a form gets filled. If you can’t name the action, you’re not ready to build yet.
2. Do you have the data — and can you actually reach it?
“We have tons of data” and “our data is accessible, labeled, and legal to use” are very different statements. Before committing, we trace the real path from source system to model. That’s where most timelines quietly slip.
3. What does “good enough” look like?
A 95%-accurate model sounds great until you realize the last 5% causes the failures that matter. Define the acceptable error rate and, crucially, what happens when the model is wrong. A human-in-the-loop fallback often turns a risky project into a safe one.
4. Who owns it after launch?
An AI system is not a deliverable you file away. It drifts, data changes, and edge cases surface. If no one owns monitoring and retraining, the value erodes within months.
5. Is AI even the cheapest thing that works?
Sometimes a handful of rules, a better form, or a search index solves 80% of the problem for 5% of the cost. We say this to clients constantly, and it’s why they trust us with the projects that genuinely do need AI.
If you can answer these five clearly, you’re probably ready to build. If you can’t, that’s not a failure — it’s the discovery work that saves you a six-figure mistake.