How I govern AI in production
Contracts, not text:
structured outputs validated against a schema on every call; the
model fills a locator, the rendering layer owns the document.
Cost and model policy:
prompt caching, token and cost accounting per file, a model chosen
per task, and a tested verifier between the model and the human.
Privacy by construction:
de-identification at the source, no identifiers in prompts, an
immutable audit log on every access to sensitive data.
Policy and program:
I own the AI usage policy and the security program at CLÉ, so the
rules agents and people follow are written down before they are
enforced.
Humans where it counts:
mandatory review before anything leaves the organisation, and two
things I refuse to automate: promotion to production and validation
of the business act (pharmacist, proposal reviewer).
Regulated ground:
AI and data platforms under Law 25 / PIPEDA and provincial standards
for Ontario education (13 school boards under the Anti-Racism Act,
2017), Quebec community pharmacy,
non-profits and international funders: RBAC, envelope encryption,
WCAG 2.1 AA, ISO 27001 practices.
A harness, not trust:
development agents work in an isolated worktree, never see secrets
or production data, can only merge to the dev branch, stop on
command, and every change is a reviewed pull request in the client's
own git history.
Read the guarantees →