WHERE AGENTIC AI BREAKS HERE
Document-tampering inputs the LLM accepts
Altered paystubs, manipulated bank statements, synthetic employment letters. A model reading for fluency does not flag what it cannot verify.
Prompt injection through free-text fields
Hardship narratives and broker comments are injection surfaces. A nudged agent drifts away from policy logic without leaving an obvious footprint.
Disparate treatment at population level
Each decision passes review. The pattern, across protected classes over a sufficient sample, fails an adverse-action analysis.
Agent evaluates the applicant and recommends a decision
Income docs, bureau pull, applicant inputs, and free-text narratives all read inside the agent's autonomous loop.
Disseqt scores it for fair-lending compliance and consistency
Every decision tested against bias signals, document integrity, and decision history across protected classes.
Model risk function sees confidence score and root-cause analysis on flags
Second-line oversight reads what triggered the flag, the population context, and the override path.
CFPB and state-regulator audit pack generated per release
FCRA, ECOA, and state-fair-lending evidence assembled from live decisions, ready for examiner review on demand.
Per-decision fair-lending scoring
Every underwriting decision evaluated against fair-lending policy and bias signals before it leaves the agent's hands.
CFPB- and FCRA-ready evidence pack
Aggregate patterns surfaced across protected classes, before a CFPB examiner runs the same analysis themselves.
One pattern, adjacent workflows
Audit trail generated from live decisions, packaged for second-line review and federal regulator engagement.
Population-level disparate-impact signals
The same assurance shape reused across loss-mitigation, servicing, and adjacent credit-decision workflows.
