artificial-intelligence-engineer/references/intake.md

Intake — mandatory questions before starting AI work

Ask (or confirm from context) before producing anything. Do not guess on 1-6.

Problem & value

1. What exact decision or task should the system improve — and what happens today without it (baseline process, baseline numbers)?
2. What does success look like as a measurable target (metric + threshold + evaluation window), and who signs it off?
3. Is ML/AI actually required, or would rules, search or a simple heuristic meet the target? What was already tried?
4. What is the cost of a wrong output (annoyance vs. money vs. safety/legal)? Which error direction is worse (false positive vs. false negative)?

Data

5. Which data exists for this task (sources, volume, label quality, freshness)? Who owns it and may we use it for this purpose?
6. Does the data contain personal or sensitive information? Which legal basis, retention and anonymization rules apply (GDPR, sector rules)?
7. How will the production input distribution differ from the historical data (seasonality, new segments, upstream system changes)?

Solution constraints

8. Latency, throughput and cost budgets per request? Online or batch?
9. Where must it run (cloud/provider constraints, on-prem, edge)? May data leave the environment — is a hosted API (e.g. OpenAI/Anthropic) allowed, or is self-hosting required?
10. Build strategy preference: prompt an existing model, RAG over private data, fine-tune, or train classically? Any hard vetoes?

Operations & governance

11. Who operates the system after launch (on-call, retraining ownership), and which monitoring/alerting stack must it integrate with?
12. Which oversight is required: human-in-the-loop, audit trail, explainability duties, EU AI Act risk class of the use case?
13. How do users get informed they interact with AI, and how do they report bad outputs?

Existing material

14. Which models, pipelines, feature stores, eval sets or prompts already exist in the company and must be reused or replaced?