Start with the decision, not the model
Most AI programmes begin with a platform choice and only later ask what the tool is allowed to decide. That order produces pilots that impress in a demonstration and stall in production, because nobody agreed what a correct answer looks like or who owns an incorrect one.
A better sequence is to name the decision, the evidence required, the consequence of an error, and the person accountable for the outcome. The technology choice follows from those answers.
The content is the hard part
Retrieval-based assistants are only as good as the material they retrieve. In most organizations, guidance, templates, and precedent are spread across document libraries, inboxes, and individual experience. Ownership is unclear and review dates are missing.
- Which sources are approved, and who approved them
- How outdated content is identified and removed
- What access rules apply to each source
- How citations are presented so a person can check the answer
Keep a person in the loop by design
An assistant that drafts and retrieves is a different risk from an assistant that decides. We design the boundary explicitly: what the system may do autonomously, what it may propose, and what must be confirmed by a named role.
Evaluation should be routine rather than ceremonial. A defined test set, a feedback route, and a review cadence tell you more than a one-off accuracy claim.
Governance is not a brake
The organizations that get value from AI are usually the ones that decided early what they would not automate. That clarity lets them move faster on the use cases that remain, because the review path is already agreed.
This publication presents analysis and perspective as of its date and does not constitute investment, legal, or tax advice. See our Risk Disclosures.