Canonical concept · WikiSure
AI Alignment
The condition where AI agents resolve business terms to the same canonical meaning as the humans and systems they serve.
Definition
AI Alignment, in the semantic governance sense, is the condition where AI agents resolve business terms to the same canonical meaning as the humans and systems they serve. Alignment is not a property of the model — it is a property of the agent + its definition context. The most capable model in the world will produce misaligned decisions if it is grounded on the wrong definition.
Business context
Most enterprise AI failures are not model failures. They are alignment failures: the agent used a plausible but unapproved definition of 'eligible customer,' 'covered loss,' or 'material event.' Semantic governance closes the alignment gap by making the canonical definition resolvable to the agent at runtime.
Insurance example
An LLM-powered claims assistant explains a denial to a customer. If the assistant resolves 'pre-existing condition' to a definition that differs from the policy's canonical definition, the explanation is wrong — and the insurer is exposed. WikiSure makes the canonical definition the only resolvable answer.
Governance example
Every agent action that depends on a definition is logged with the definition's version. If the canonical definition is updated, downstream agent decisions made against the prior version are flagged for re-review — turning alignment from a one-time setup into a continuous governance signal.
Related concepts
Frequently asked
- Is AI alignment the same as model safety?
- No. Model safety is about preventing harmful outputs. Semantic alignment is about ensuring the agent's outputs use the same meaning as the rest of the organization. Both matter; they are not interchangeable.
- How does WikiSure enforce alignment at runtime?
- Agents resolve terms against the WikiSure registry via an API. The registry returns the canonical definition for the agent's namespace and logs the resolution. The agent is free to refuse to act when no canonical definition exists.