AI SafetyVersion: v1.0Updated: October 2026
AI Use & Limitations Policy
Defines the strictly read-only architecture, probabilistic limitations, hallucination risks, and mandatory human review requirements.
Audience: All users interacting with Ask Kairo and AI-assisted workflows
Developer: Zambezi Digital
Official Policy Publication — Enforced Standards
This document is an active, binding operational policy published by Zambezi Digital governing the security, data privacy, and usage regulations of Kairo OS across all jurisdictions.
On this page
1. Operational Capabilities & Scope
Ask Kairo uses retrieval-augmented generation (RAG) to search and synthesize business data from documents, fleet records, messages, and calendar entries within the user's authorized workspace. Access permissions are strictly enforced in application code and database policies.
Key takeaway: Ask Kairo only accesses documents within your authorized organization.
2. Strictly Read-Only Architecture
Kairo AI is strictly read-only and analytical. It possesses NO direct database mutation authority. It cannot delete vehicles, modify documents, change passwords, or dispatch fleets. Any statement by the model claiming to have performed a database write is an automated hallucination.
Key takeaway: The AI cannot mutate your database or execute actions autonomously.
3. Output Uncertainty & Hallucination Risks
Large language models generate probabilistic responses. Outputs may be incomplete, outdated, or factually inaccurate. A confident tone does not establish factual accuracy. Users must cross-reference outputs against original source documents.
Key takeaway: Always check original documents; confident AI answers can still be mistaken.
4. Mandatory Human Verification
Kairo AI is NOT a substitute for licensed legal, financial, accounting, medical, or fleet safety professionals. Consequential decisions must be independently reviewed and approved by qualified human personnel.
Key takeaway: Never make consequential legal, financial, or safety decisions on AI outputs alone.
5. Prompt Injection Defenses
Retrieved business documents and external files are treated as untrusted data and encapsulated inside isolated XML boundary tags. Direct jailbreaks and bracketed system overrides are actively detected and filtered by our guardrails engine.
Key takeaway: External documents are quarantined to prevent prompt injection hijacking.
Have questions about this policy?
Our legal and privacy desk is available for corporate inquiries.