AI Readiness Before Adoption

Discover, evaluate, prioritise, design, implement, govern, optimise. The sequence matters more than the tool selection, because an organisation that implements before it evaluates cannot say afterwards what it gained or what it risked.

The single most useful constraint is to prove value on one narrow use case with a measured baseline. Narrow scope makes the result legible. A broad pilot produces enthusiasm and no evidence.

Under the Protection of Personal Information Act, whether an organisation acts as a responsible party or as an operator changes its obligations materially, and the distinction should be established and recorded at the start of the engagement rather than negotiated afterwards.

Practically, that means the supplier register records the tool, the provider, the data location and the no-training commitment before a single document is indexed. It also means client confidential material is never indexed into a system with wider access than the engagement it belongs to.

For institutional use, retrieval over an approved corpus is almost always the correct architecture. It keeps the source of truth editable, it allows every answer to cite where it came from, and it fails visibly rather than silently.

Configure the assistant to refuse rather than guess when retrieval returns nothing relevant, and then test that it does. This is a functional requirement, not a preference, and it belongs in the acceptance criteria.

Log every query. Queries reveal what people cannot find, which is a more honest content audit than any workshop, and it costs nothing beyond the discipline of reading it monthly.

A committee is not asking whether AI is interesting. It is asking who is accountable, what data leaves the institution, what happens when the system is wrong in public, and how the decision will be reviewed.