Trust-Earned Autonomy: How Texas Capital Bank demonstrates an AI agent that commits to the core banking ledger, then reverses itself

This article describes Trust-Earned Autonomy, a pattern Texas Capital Bank built to safely grant AI agents autonomous decision-making authority over banking transactions, with automatic reversal when trust degrades.

• AI agents earn autonomy through measured trust scores based on golden-set coverage, replay accuracy, and human reviewer agreement, not self-reported confidence
• Recipes (plain-English decision rules) are versioned and confined to a closed set of ten pre-approved tools, preventing prompt injection and unauthorized actions
• Multi-predicate gates evaluate eight conditions per run before autonomous commits; any failure routes to human review rather than proceeding silently
• Demotion is fast: recipes lose autonomy immediately when trust falls, triggering automatic reversal of prior commits with compensating ledger entries
• System built on AWS using Bedrock AgentCore, DynamoDB, Neptune, EKS-hosted Apache Fineract, and Cedar policy enforcement for amount caps
• Measured on 8-document benchmark: 100% classification accuracy, 92.6% extraction precision, $0.11–$0.61 per document cost

The pattern addresses regulatory expectations for explainability and auditability by providing versioned rules, measurable trust metrics, documented gates, and defined rollback procedures for AI-driven banking decisions.