Responsible AI
State of Stick treats AI output as generated material that must earn trust through evidence, validation, and appropriate human review.
Core controls
Section titled “Core controls”- Registered tasks. Every supported task has a defined purpose, input and output shape, model lane, limits, sensitivity, and approval requirement.
- Server-side authority. Browsers and customer inputs cannot choose a provider, model, spending limit, or broader customer scope.
- Grounded claims. Factual output should use approved source material and retain the identity of the supporting sources.
- Structured validation. Generated output is checked against the task’s expected structure before the application accepts it.
- Human approval. Material sales, campaign, proposal, and customer deliverables remain drafts until an authorized person reviews them.
- Operational controls. Feature switches, task limits, rate controls, and budget gates can stop AI work without disabling deterministic platform functions.
Claim classifications
Section titled “Claim classifications”Sales and campaign recommendations distinguish among:
| Classification | Meaning |
|---|---|
| Verified platform fact | An approved statement about State of Stick or a supported platform capability |
| Prospect-specific fact | Information supported by an approved account or prospect source |
| Strategic recommendation | A proposed approach, not a statement of fact or commitment |
| Assumption requiring confirmation | A useful working assumption that must be checked before use |
This distinction helps reviewers identify what is known, what is suggested, and what still needs confirmation.
Data boundaries
Section titled “Data boundaries”Restricted information is not eligible for AI processing by default. Customer-confidential workflows require an authorized internal context and are not treated as public reusable content.
Public documentation, client-side code, and AI prompts must not contain:
- credentials, tokens, or secret configuration;
- private customer records or personal contact information;
- confidential supplier economics, costs, or manufacturing methods;
- secure-device configuration or proprietary verification procedures;
- unpublished legal, contract, or incident material.
Honest limitations
Section titled “Honest limitations”AI may produce incomplete, poorly phrased, or unsupported material. A polished answer is not proof that a claim is true. Reviewers remain responsible for checking sources, customer context, permissions, scope, and commitments before using an output.
There is no automatic assumption that a more expensive model is correct. Premium processing is a controlled quality option, not a substitute for evidence or approval.