Days 1-30: discovery
The most expensive mistake in enterprise AI is building the wrong thing fast. Discovery is not a formality. It is the work.
Interview 8 to 12 stakeholders across business units. For each, ask: what manual task takes more than 2 hours per week? What decision do you make with incomplete information? What report do you wish existed but is too expensive to build?
Map the candidates on a 2x2: impact (revenue, cost, risk) vs feasibility (data quality, integration complexity, regulatory constraints). The top-right quadrant is your first sprint.
Days 31-60: prototype and validate
Pick one use case from the top-right. Build a prototype in 3 weeks. The prototype does not have to be production-grade. It has to be testable by domain experts.
Run a structured eval: 100 questions, domain expert scores each answer 1 to 5. Set a threshold (e.g. 4.0 average or higher) before the sprint begins. If the prototype clears it, proceed to production hardening. If it does not, investigate root cause (usually data quality or chunking strategy) before committing engineering resources.
Days 61-90: first production deployment
Scope the first deployment to a single team of 10 to 20 people. This limits blast radius and generates real usage data fast.
Instrument everything: latency, cost per query, thumbs-up/thumbs-down from users, faithfulness score from the automated harness. Review metrics weekly with the business owner. Adjust chunking, retrieval strategy, or model tier based on what the data shows, not intuition.
At day 90, you have a live system, a tuned eval harness, and a clear picture of what the second use case should be. That is the foundation for a credible 12-month roadmap.