Minimum necessary data
We ask for the smallest useful data set and separate test data from production data.
AI projects handle knowledge, customer conversations and creative assets. We make the boundaries visible before a pilot touches real information, and keep a named human owner in the loop.
We ask for the smallest useful data set and separate test data from production data.
Sources, processors, storage, retention and deletion are documented for each build.
People see what their role requires. Privileged or sensitive actions have stronger controls.
Grounding sources have owners and review dates. The system should know when not to answer.
Escalation, approval and review paths are designed with the teams who operate them.
Certification, hosting and regulatory requirements depend on the project. We define the applicable controls rather than implying universal compliance.
At the start of a project, we make these items concrete enough to review and challenge.
Which model and cloud providers process data, and where.
What is logged, how long it stays and who can inspect it.
Where AI decides, where it only drafts and where a person must approve.
How incidents, corrections and data requests are handled.
Data flow, answer evidence, exception handling and final approval are the real starting point for safe, operable AI.