1. Use cases AI projects fail at selection more often than at engineering.
Write down the top 5 candidate use cases in one sentence each If a use case needs a paragraph to explain, it is not ready to scope. Score each on value (revenue, cost, risk) and feasibility (data, latency, error tolerance) A 2x2 of value vs feasibility kills most bad ideas instantly. Check error tolerance honestly Customer support drafting tolerates mistakes; invoice payment execution does not. Start where errors are cheap and reviewable. Identify the human in the loop for the first version Version one should augment a named person’s workflow, not replace an unowned process.
2. Data Locate the data each use case needs and who owns it If the answer is "it is in five systems and two spreadsheets", that is the first project. Check freshness: how stale can the data be before answers are wrong? Verify you have the legal right to process it with an AI provider Customer contracts and privacy policies may restrict sending data to third-party models. Identify PII and what must be redacted or anonymized before it reaches a model
3. Infrastructure and serving Decide API models vs self-hosted, per use case API models (OpenAI, Anthropic) win on speed-to-value for most teams. Self-hosting only pays off at scale or under hard data residency rules. Define latency budgets: chat needs seconds, batch enrichment does not Plan for rate limits and fallbacks Every provider throttles. Decide now what the product does when the model is slow or down. Set up evaluation before launch A test set of 50-100 real examples with expected outputs catches regressions every prompt change.
4. Security and governance Write the list of what the AI is never allowed to do (send money, delete data, sign contracts) Log every prompt and response for audit and debugging Decide who reviews AI output quality, and how often Test prompt injection on anything that touches untrusted input If users or documents can talk to your model, assume someone will try to hijack it.
5. Cost model Estimate tokens per interaction times interactions per month, per use case A support bot at 2,000 tokens per chat and 10,000 chats a month costs very different money on a frontier model vs a small one. Use the cheapest model that passes your evaluation set, not the newest one Set a monthly budget alert on the provider account before launch
Want this mapped to your actual stack? The AI-Readiness Assessment scores your shortlist, audits your data and infrastructure gaps, and delivers a staged roadmap from pilot to production, with a serving cost model for your top use cases.