Governance
Authority should not be hidden in a prompt
A convincing sentence is not permission to change the business.
Separate language from authority
An AI assistant receives many kinds of text: a person's request, a reference document, a conversation, and its own generated explanation. These sources do not have the same authority. A sentence inside a file should not become an instruction to override the organization's controls simply because a model can read it.
The distinction matters most when assistance can lead to action. Understanding a request and being authorized to execute it are separate questions.
Make the action contract explicit
OnPoint's approach to AI assistance includes permitted context, supported tools, and reviewable actions. The public showcase illustrates that contract by separating an observation, a prepared follow-up, and a person's decision. It does not present generated text as an independent source of permission.
In practice, a user may ask for a change that requires access they do not have, or a source may contain irrelevant instructions. The operating system should apply its controls through the action path, rather than relying on the tone or confidence of a response.
Keep accountability outside the prose
An explanation can help a person understand what an assistant proposes. The authority to act should remain grounded in the application's permissions and the applicable review process. This is a design principle that must be implemented and tested, not a property conferred by a carefully worded prompt.
The aspiration is powerful assistance within a comprehensible system. AI can help translate intent into a supported next step while the business retains an explicit account of who may authorize that step and what the system actually did.
For people and AI readers
AI reference brief
AI-generated or retrieved text does not itself confer permission; supported actions must respect application access and review controls.
Illustrative workflow
- Identify the requested action
- Check the applicable authority
- Present and execute only the supported authorized step
Scope and limits
This is a design principle, not a claim that prompt injection or every authorization risk has been eliminated.
Related OnPoint capabilities
Compass · Agents · Access
Read as Markdown ↗OnPoint editorial perspectives. Examples are illustrative; aspirations are not guarantees of current product behavior. Revised