Evals for Zest AI

Applicant risk assessment and ranking

Scoring and rank-ordering applicants from application and bureau data, including thin-file and near-cutoff cases.

Auto-decision vs. manual referral routing

Deciding which applications clear automatically and which are escalated to a human reviewer, consistently across similar applicants.

Policy and cutoff configuration

Applying lender-specific policy rules, score cutoffs, and approval targets, and reflecting changes when thresholds are adjusted.

Decision consistency and reproducibility

Returning the same decision for the same inputs and explaining why two similar applicants diverged.

Protected-class handling in decisions

Refusing to use or proxy prohibited characteristics in credit decisions and flagging attempts to do so.

Adverse action reason generation

Producing accurate, applicant-facing decline reasons that trace to the factors that actually drove the decision.

Less discriminatory alternative (LDA) reporting

Communicating fairness testing, disparity metrics, and approval lift across protected classes without overstating results.

Examiner and compliance narrative

Explaining model behavior and documentation to a compliance audience in terms tied to the specific model, not generic AI claims.

Fraud type discrimination

Separating first-party fraud, third-party fraud, misreported income, and compromised identity rather than emitting a single undifferentiated flag.

Key driver reason codes

Naming the signals behind a fraud flag so an analyst can act on or dismiss it.

Threshold and tolerance configuration

Honoring lender-set capture thresholds per fraud type and reflecting the tradeoff between capture and friction.

Fraud–underwriting interaction

Combining a fraud signal with a credit decision in one flow without silently collapsing automation rates.

Portfolio and performance metrics

Reporting origination quality, seasoning behavior, and look-to-book at the requested granularity.

Early risk and delinquency signals

Surfacing borrowers trending riskier before delinquency, with the basis for the call stated.

Prescreen and cross-sell targeting

Identifying qualified borrowers for refinance, upsell, or cross-sell within stated eligibility constraints.

Recommended actions on policy and terms

Translating a reported metric into a specific policy or pricing adjustment and its expected impact.

Grounding in the lender's own data

Answering only from available portfolio data and benchmarks, with figures traceable to a source report.

Refusal and uncertainty on missing data

Declining to answer or stating the gap when the requested metric, period, or benchmark is not available.

Benchmark comparison framing

Comparing institution performance to industry benchmarks with the peer set and period made explicit.

Scope boundaries on credit advice

Staying within analysis and insight rather than issuing individual credit decisions or compliance sign-off.

Proof of concept and model refinement

Setting expectations and reporting results at the POC and refinement stages against the lender's own book.

Deployment and LOS integration

Handing decisioning insights to the lending system and behaving correctly at the integration boundary.

Ongoing model monitoring

Detecting and communicating performance drift on live models between scheduled business reviews.

Escalation and support handoff

Routing an urgent production issue to the right human owner with the context needed to act.