The AI project feasibility checklist.

Most failed AI projects were never AI problems. Check these first.

• Can you describe the outcome in one sentence without the word "AI"? (e.g. "cut invoice processing from 3 days to 1 hour")
• Does a human currently do this task successfully? If no human can do it, a model probably cannot either.
• Would a 90%-accurate system still be useful? If you need 100%, you need software, not ML.
• Is there a measurable number this project moves — hours saved, revenue, error rate?
• Have you checked whether a plain workflow automation (no AI) solves 80% of it cheaper?

AI projects die in data access more than in modeling.

• Does the data this system needs actually exist today, in a system you control?
• Can you get access to it within two weeks — or is it locked behind another team, a vendor, or compliance?
• Is there enough of it? (For document AI: hundreds of real examples, not five.)
• Do you know what "correct" looks like — labeled examples, historical outcomes, or a human who can judge output?
• Is the data legally usable for this purpose (PII, PHI, customer contracts)?

Honest numbers from real engagements, not vendor decks.

• A production-grade AI pilot realistically starts around $10K+ and 6–8 weeks. Anything cheaper is a demo.
• Do you have budget for month 2+? Models degrade, prompts drift, APIs change — production AI needs ownership.
• Is one named person on your side able to spend 2–3 hours a week with the engineering team?
• Do you need this in under 4 weeks? Then descope to a pilot, or do not start.
• Have you accounted for inference costs at real volume, not demo volume?

The questions that separate demos from systems.

• What happens when the model is wrong — is there a human fallback path?
• Does the output need audit trails or compliance sign-off (HIPAA, SOC 2, legal review)?
• Who owns the system after launch — your team, or are you expecting the vendor to babysit it forever?
• Can the system be tested against a golden set before it touches real users?
• If the AI vendor/API you depend on doubles prices or deprecates a model, what breaks?