Intent provenance is metadata that identifies the origin of an AI product's current intent signal.
In the AON Query contract, user_expressed means the user directly supplied the current request, while inferred_context means the product derived it from bounded available context and must provide a confidence value. Provenance does not prove that an inferred preference is true; it tells downstream systems and users how strongly to treat the signal.
Intent provenance in AON Query
AON's Query contract carries the current intent and its provenance so an AI product can distinguish direct user requirements from bounded inference. This supports clearer relevance behavior without requiring the product to send a complete transcript.
Read the Query intent modelHow intent provenance works
Capture the current intent
The product sends the user's current text or image input rather than a full chat transcript or an invented long-term profile.
Declare the origin
The request marks whether the intent was user_expressed or inferred_context under the active Query contract.
Bound the inference
An inferred signal carries the required confidence and remains limited to context the integration can accurately support.
Use it appropriately
The system can use provenance to interpret relevance and follow-up behavior without treating an uncertain inference as a user-authorized hard constraint.
Implement Intent provenance safely
Keep
user_expressedandinferred_contextas separate provenance states.- The required confidence for
inferred_contextand the distinction between confidence and purchase probability. - Only supported, current intent signals; omit unknown values instead of guessing them.
Do not
- Labeling an AI inference as
user_expressedor presenting it as something the user said. - Using provenance as purchase authorization, consent, or merchant eligibility.
- Treating confidence as a probability that the user will buy or that an Offer will convert.
The two provenance states
The provenance value is simple, but it changes how an AI product should describe, rank, and act on the signal.
| Concept | Meaning |
|---|---|
| `user_expressed` | The user directly expressed the current goal or condition. The product can send it as the request evidence without adding an inference confidence value. |
| `inferred_context` | The product inferred the signal from bounded context. The active contract requires confidence so downstream consumers can distinguish evidence from interpretation. |
| Confidence is not truth | A confidence value is a producer-normalized estimate, not a guarantee that the inferred preference is correct or that the user authorized it. |
| Constraint boundary | Explicit deterministic constraints remain separate from semantic intent and inferred context. Missing information should be omitted rather than guessed. |
- Intent provenance is not a user identity, consent record, or authorization to purchase.
- An inferred signal should not be presented as something the user said or used to override an explicit user condition.
- Provenance metadata does not replace merchant checkout verification or the product's own privacy and consent responsibilities.
Intent provenance vs intent matching
Intent provenance describes where an intent signal came from. Intent matching estimates how well that request fits a candidate Offer. A system can use both, but provenance is about evidence origin while matching is about relevance. See the Intent matching definition.
| Dimension | Intent provenance | Intent matching |
|---|---|---|
| Question answered | Where did the intent signal come from? | How well does the request fit a candidate Offer? |
| What it carries | Whether intent was user-expressed or inferred, plus required confidence for inferred context. | A relevance assessment among candidates that passed applicable eligibility rules. |
| Boundary | Does not prove that an inferred preference is true or authorize a purchase. | Does not prove merchant eligibility, availability, or checkout acceptance. |
Example in practice
A user says, 'Find a lightweight laptop for travel.' That request is user_expressed. If the assistant infers that the user probably values long battery life from the travel context, it can send that as inferred_context with the required confidence, but it should not tell the user that battery life was explicitly requested.
Common questions about intent provenance
What does user-expressed intent mean in an AON Offer Query?
user_expressed means the current intent came directly from the user's request. An AI product should not use it to label an inference or an assumption about the user's long-term preferences.
When is inferred-context intent appropriate in an AON Offer Query?
inferred_context is appropriate only when the integration has a bounded, supportable basis for the inference. The AI product must include the confidence required by the active contract and must not represent the inference as a direct user statement.
Is AON intent confidence a probability that the user will buy?
No. Intent confidence describes the producer's confidence in an inferred intent signal. It is not purchase probability, merchant eligibility, or a commission forecast.
When is confidence required for AON intent provenance?
The active AON Query contract requires confidence for inferred_context, not for user_expressed intent. The value describes confidence in the inference; it is not purchase probability, eligibility, consent, or authorization.
Sources
- AON Offer Query APIAON documentation
- AgentOffer Protocol Query specificationAON specification
- W3C PROV-O provenance ontologyPrimary standard