Abstract

An intent belief layer system (100) for transactional artificial intelligence assistants is disclosed. The system includes a text encoder configured to receive a user utterance and generate an embedding vector, and an intent classifier configured to generate intent-wise logits for an intent taxonomy in which each intent is associated with risk metadata. A belief estimator applies calibrated probability mapping to generate an observation probability distribution, retrieves a prior probability distribution from a session store, and computes a posterior probability distribution using a Bayesian belief update across conversation turns. A risk engine computes an expected risk score over the posterior probability distribution based on the risk metadata, and derives a corresponding risk band. A decision policy module determines a routing action based on an entropy of the posterior probability distribution and the derived risk band, wherein the routing action comprises one of executing the requested operation, generating a clarification request, escalating the request for human review, or blocking the requested operation. The system thereby enables uncertainty-aware and risk-aware routing of financial service requests while maintaining traceable belief, risk, and routing metadata for auditability.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS