Governed data comes before useful AI

Every organisation wants AI outcomes. Most have a data layer that cannot support them: no catalogue, access granted by copying files, and lineage that lives in one engineer’s memory. The model is rarely the constraint. The platform under it is.

Three capabilities, in order

A governed platform is not a product you buy; it is three capabilities you build and keep. In priority order: a catalogue, access control at the data layer, and lineage.

Govern at the data, not the copy

The anti-pattern that quietly defeats governance is granting access by exporting a copy. Once a dataset has been copied into a notebook, a spreadsheet, and a shared folder, no policy applies to it. Grant access to the governed table and query it in place; never hand out the extract.

-- Access is a grant on the governed object, not a copy of the data.
GRANT SELECT ON gold.customer_360 TO ROLE analyst_marketing;

-- Row-level security keeps the same table safe for a wider audience.
CREATE POLICY region_filter ON gold.customer_360
  USING (region = current_setting('app.user_region'));

Only once those three capabilities exist does the model layer become the interesting problem. Before that, every AI project is quietly paying an unbudgeted tax to work around the platform that should have been there first.