UNCHORUNCHOR OS Artificial Intelligence and Automation Policy
Version 2026-10-23. Effective 23 October 2026.
This policy explains how UNCHOR OS uses rules, automation and AI models, what they can and cannot do, and the controls that keep you in charge.
1. Rules first, models only where they are better
Most diagnoses and recommendations in UNCHOR OS come from deterministic rules and playbooks whose behaviour is fixed and tested. An AI model serves a task only after it has been evaluated on fixed test cases for that task, has done at least as well as the rules and the current choice, and a person has approved it. If a model fails, times out or gives an answer that does not match the required format, UNCHOR OS falls back to the rules.
2. Which providers, and what data they see
UNCHOR OS does not depend on one AI provider. The organisation that runs this UNCHOR OS service can connect a hosted model (such as Claude by Anthropic or an OpenAI-compatible service) or a model on its own machines; with none connected, every feature runs on rules alone. Each provider is allowed a data class, and a hosted provider never receives more than the second:
- aggregate: totals and counts with nothing that identifies a person or a site
- workspace: information about one workspace's network (measurements, settings, a request typed by its owner), without names, phone numbers, email addresses or device addresses
- personal: may identify a person; allowed only for a model that runs on the service's own machines, so the data does not leave them
A request is refused if it would send data above the class the provider is allowed. Every call is recorded with the task, provider, model, data class, outcome and cost.
3. Agents
Specialist agents (diagnostics, security, operations) investigate problems for you. Each has a fixed role, a list of the tools it may use, and a budget of steps, time and tokens. An agent:
- can only read what its tools allow, and only within your workspace
- must cite the evidence for each conclusion; figures it did not observe are rejected
- cannot change your network, your settings or its own permissions; it can only propose an action, which waits for a person's approval under your workspace's rules
4. What is shown to you
Every diagnosis separates what was observed, what it suggests, what is uncertain, what is proposed and what was actually done and verified. Results produced by a model are labelled with the model's name. Actions report success only after the result has been verified on the gateway.
5. Learning
UNCHOR OS learns which detectors and fixes work from verified outcomes and from labels people give ("it was real" or "a false alarm"). Learning across workspaces uses only de-identified records from workspaces that allow it. Learning never changes permissions, approval rules or security controls.
6. What we do not do
We do not use AI to make decisions that have legal or similarly significant effects on a person without human review, to profile your customers, to read their traffic, or to generate fake reviews, testimonials or claims. We do not let a model run arbitrary commands, database queries or network operations.
7. Questions and objections
The list of AI providers in use, with their regions, is in the app under Settings → Privacy & data → Who processes data. Every investigation shows whether its result came from rules or from a model, and you can always choose not to act on a proposal. Questions or objections: the contact form at https://unchor.co.ke/contact (topic: Support).