- Context
- It needed AI in the daily work of every area without giving up sovereignty over its own data, and without betting blind on a single vendor.
- Mandate
- Define the stack and the architecture, run the market evaluation and bring the board a decision it could stand behind.
- What was done
- Process review with each area and a future-state design signed off by the managers. Proofs of concept with AWS, Azure and GCP, compared layer by layer, and a build-versus-buy analysis carrying total cost for the first year and the ones after. A layered architecture — pluggable, with no vendor lock-in — with private AI governance, zero retention and LGPD compliance from day one. And a six-month roadmap with parallel build and buy paths, backed by an interactive model that simulates the cost of each combination of choices.
- What changed
- The board started deciding with numbers on the table, not a vendor deck. And the design stayed pluggable: changing provider stopped being a fresh start.