ARCHITECTURE
Connect intent, authority and execution.
Business reasoning, real-time voice and independent checks have different jobs. Anson’s architecture brings them into a collections workflow, with the active interfaces, control points and expected behavior established and tested for each project.
Four responsibilities in one workflow.
Business context and decisions
Read authoritative account state, the current objective and approved rules. Propose a response or tool action without inventing facts or permissions.
Real-time voice execution
Handle audio, turns and playback through the configured voice stack. Generated text and audio actually played are separate events.
Independent execution checks
Check the covered response or action at the configured control point. An authorized component applies the decision to allow, reject or pause.
Evidence and review
Connect the policy version, check result, execution receipt and exception owner. Verify the actual result before recording a conclusion.
A control needs an execution point.
A model’s risk flag does not enforce a restriction on its own. A control can constrain only the output or action covered by an integrated, active component. A requested stop, an accepted command and a confirmed final state need separate evidence.
Audio already played cannot be recalled. Pausing new tasks does not establish that an active call has ended.
Anson & Agora
Anson and Agora collaborate on real-time voice for post-loan workflows. Agora contributes the agreed real-time audio and conversational AI capabilities; Anson focuses on business context, approved tasks, execution controls and result verification. Enabled capabilities and end-to-end behavior are confirmed through project configuration and joint acceptance.
Asset onboarding is one part of the chain.
The public V1 API validates signed requests and creates the corresponding asset and case records. It does not launch an outbound call. Confirm task authority, state synchronization, result receipts and enabled interfaces separately.
Test the whole path, including failure and recovery.
Use synthetic cases for interruptions, stale facts, query timeouts, missing permissions, duplicate commands and policy withdrawal. Record the line, network, configuration and component versions. Measure the experience and control behavior separately; a fast response alone does not establish a collections outcome.
Review trust & controls