The Feasibility Checklist

Andrew Ng's famous screen: if a typical human can do a mental task in under a second, today's AI can probably learn it too. It's a decent filter for technical feasibility — but technical feasibility is the cheap half. Plenty of technically-doable AI features lose money.

The commercial checklist is stricter. Five questions:

• Frequent? A task done 10,000 times a day amortizes build cost; one done quarterly never will
• Expensive per instance? Automating a $5 task with a $0.05 model call is a business; automating a $0.02 task isn't worth the engineering
• Does the data exist? You need examples — yours, a vendor's, or what a base model already absorbed. No data, no product
• Are errors tolerable? AI is probabilistic. If one wrong answer in 10,000 is a lawsuit, the accuracy bar may be unreachable
• Is there a feedback loop? Can you detect wrong answers and improve? Products without loops stay at demo quality forever