How we work
The firm sells AI governance, so its own AI build standard is public. Each principle names where it was applied. Client work is anonymised.
01
Model output is validated against real records on the server before it reaches a user. Prompt-level grounding reduces a failure rate; it is not a control.
Where it was applied: Carnival Concierge, where a recommended event must resolve to a real event record.
02
Anything with a compliance or safety consequence is a versioned rule with an owner and an effective date. AI is confined to conversation, extraction and drafting.
Where it was applied: CareAI Assistant's pattern-detection engine; the safety rule engine in the women's health specification; compliance-owned rule sets in the insurance onboarding blueprint.
03
Every product states what the system may do and what requires an authorised human, before the first screen is designed.
Where it was applied: The insurance advisor onboarding blueprint; the women's health specification's launch boundary.
04
Model calls sit behind a provider interface with a second provider configured as failover, so a provider outage is not a product outage.
Where it was applied: The Caribbean AI Tutor and the bank engagement prototype.
05
Generated content that reaches a user passes a review queue. The queue is a product feature, not an operational afterthought.
Where it was applied: Clinician-reviewed content in the women's health specification; briefs in the legal case-brief platform.
06
Retrieved content is data, never an instruction. Instruction-like sequences are neutralised, and an injection suite runs in the release gate.
Where it was applied: The firm's AI QA agent operating brief, and this site's own assistant when it is switched on.
07
A configured model, a spend cap and a static fallback, so that cost is bounded and failure degrades rather than breaks.
Where it was applied: This site's assistant design, which degrades to a searchable FAQ at its spend cap.