Fair lending is the one corner of responsible AI where the regulated enterprise cannot plead novelty. US fair-lending compliance is, per the most recent field study of these programs, "likely the longest-running example of algorithmic fairness on the ground", nearly fifty years of non-discrimination law applied to algorithmic credit decisions (arxiv.org/abs/2606.02957). The same study's central finding frames this whole Part: standing supervisory authority "has played a unique role in successfully fostering fair lending practices — a regulatory design feature that is distinct from other areas of civil rights law and almost completely absent from recent policy proposals for dealing with algorithmic discrimination." Disclosure artifacts do not keep lenders fair; examiners with the power to compel change do. Build the internal control set as if the examiner is already inside the building, in a supervised institution, she is.
What agentic autonomy adds is not a new form of discrimination but a new way of acquiring one that nobody designed, approved, or documented.
VI.1: How agentic autonomy changes fair-lending and credit risk
A conventional credit model has a fixed feature vector and a validation report describing it. Its discrimination risks are serious but legible: they live in the data, features, and outcome definition, the territory mapped by the foundational disparate-impact literature (lawcat.berkeley.edu/record/1127463) and by credit-market studies showing ML underwriting can widen group outcome gaps without any prohibited input (doi.org/10.1111/jofi.13090; doi.org/10.1016/j.jfineco.2021.05.047). An agentic decision flow breaks that legibility four ways:
- Self-selected features. An agent with tool access chooses at runtime which information to gather, so two applicants can be judged on materially different evidence sets, and the evidence-set choice can itself correlate with protected class (thin files, non-standard documents, names and addresses steering retrieval). The legal literature already holds that auditing approved inputs never sufficed to establish non-discrimination (scholarship.law.umn.edu/minnlrev/vol106/iss3/2); with an agent, the input list is not even stable.
- Self-selected steps. The agent decides how much verification, documentation, and escalation each applicant gets. Procedural burden can distribute unevenly across groups even when approval rates look balanced.
[practice guidance: not directly source-backed] - Emergent, system-level bias. Multi-agent fairness research, using credit scoring and income estimation as its testbed, demonstrates "patterns of emergent bias in financial decision-making that cannot be traced to individual agent components" and concludes such systems "must be evaluated as holistic entities" (arxiv.org/abs/2512.16433). A fairness sign-off on each component is not a fairness sign-off on the pipeline.
- Nondeterminism and drift-by-conduct. Model, prompt, and tool changes alter behavior without a release event, so fairness properties established at validation are perishable. Noise is itself a disparity mechanism: unevenly distributed measurement noise in credit data translates into unequal outcomes (arxiv.org/abs/2105.07554).
The governing risk statement: an agent that chooses its own features and steps can introduce discrimination no one designed, no one approved, and no existing artifact describes. The controls below make that impossible in covered decisions and detectable everywhere else.
VI.2: Disparate-impact testing for agentic decision flows
The classical toolkit survives: adverse-impact ratios and the four-fifths heuristic (eeoc.gov/select-issues-assessing-adverse-impact-software-algorithms-and-artificial-intelligence-used), formal certification and repair of disparate impact (arxiv.org/abs/1412.3756), error-rate-parity metrics (arxiv.org/abs/1610.08452), and the impossibility results that force a documented choice among incompatible criteria (arxiv.org/abs/1703.00056). The bridge from academic metrics to US fair-lending doctrine is mapped in the FAccT alignment work (doi.org/10.1145/3514094.3534154), with open lending-domain tooling available (github.com/solasai/solas-ai-disparity). What changes for agents is the unit of testing:
- Test trajectories, not models. The tested object is the end-to-end flow (retrieval choices, tool calls, escalations, final outcome) against matched cohorts. Component-passing/system-failing is the expected failure mode (arxiv.org/abs/2512.16433).
- Test process, not just outcomes. Join outcome parity with parity of treatment intensity: documentation requests, verification depth, time-to-decision, escalation rate.
[practice guidance: not directly source-backed] - Test under perturbation. Matched-pair testing (identical financials, varied protected-class correlates (name, address, English fluency)) directly probes the self-selected-evidence failure mode; disparate LLM performance against non-native English writers is empirically demonstrated (arxiv.org/abs/2304.02819).
[protocol design: practice guidance — not directly source-backed] - Re-test on change. Any change to model, prompt, toolset, or orchestration invalidates prior disparity results until re-tested.
[practice guidance: not directly source-backed]
Doctrinal caution: US disparate-impact law is in motion, HUD has proposed rescinding its Fair Housing Act disparate-impact rule (federalregister.gov/documents/2026/01/14/2026-00590/huds-implementation-of-the-fair-housing-acts-disparate-impact-standard) while New Jersey codifies the doctrine (njoag.gov/wp-content/uploads/2025/12/n.j.a.c.-13-16-disparate-impact-discrimination.pdf). Calibrate legal theory with counsel per jurisdiction, but do not calibrate the testing program down to the weakest current rule: examiners, not doctrine du jour, are what durably shape practice (arxiv.org/abs/2606.02957).
VI.3: Adverse-action explainability when an agent acted
ECOA requires specific reasons for adverse action (law.cornell.edu/uscode/text/15/1691), operationalized by Regulation B (law.cornell.edu/cfr/text/12/1002.9). The CFPB has closed both escape hatches an agent deployer might reach for: Circular 2022-03, complexity of the algorithm does not excuse specific reasons (consumerfinance.gov/compliance/circulars/circular-2022-03-adverse-action-notification-requirements-in-connection-with-credit-decisions-based-on-complex-algorithms), and Circular 2023-03, the sample-form checklist is no safe harbor; reasons must be specific and accurate to the actual decision (consumerfinance.gov/compliance/circulars/circular-2023-03-adverse-action-notification-requirements-and-the-proper-use-of-the-cfpbs-sample-forms-provided-in-regulation-b). Post-hoc explainability for complex credit models has been empirically evaluated in lending (finreglab.org/research/machine-learning-explainability-fairness-insights-from-consumer-lending); the CFPB's early innovation-spotlight commentary on AI/ML adverse-action notices has since been disclaimed and must not be relied on (consumerfinance.gov/about-us/blog/innovation-spotlight-providing-adverse-action-notices-when-using-ai-ml-models).
The agentic twist: the principal reasons for a denial may not be model coefficients at all but trajectory events, a verification the agent chose to run, a document it judged inconsistent. Three consequences, all [practice guidance: not directly source-backed] in their engineering detail: (1) reasons must derive from the recorded trajectory, not generic feature attribution over the final score; (2) a fluent LLM rationale is not evidence of the true cause, reasons must be reconstructed from logged actions and validated counterfactually; (3) if a compliant reason cannot be produced for a trajectory class, that class is prohibited for covered decisions. Explainability is a gate, not a report.
VI.4: Complaints, remediation, and vulnerable customers
Two structural facts harden the customer-outcome backstop. First, volume: an agent generates customer-affecting events at machine rate, so complaint processes sized for human error will saturate. Second, correlated failure: one defective prompt or policy change harms an entire cohort at once, so remediation must be cohort-first, not case-by-case. Both points are [practice guidance: not directly source-backed]; the cited research anchors are the statutory backstops that make remediation non-optional in credit (law.cornell.edu/uscode/text/15/1691; law.cornell.edu/cfr/text/12/1002.9) and the HMDA data that makes lending disparities externally visible whether or not the institution looks (consumerfinance.gov/data-research/hmda).
On vulnerable customers, one blunt paragraph: an autonomous agent is tireless, persuasive, and indifferent by default. Pointed at collections, retention, or cross-sell, it will pursue its objective against a customer in financial distress or cognitive decline with exactly the persistence its objective implies. The controls below treat vulnerability signals as hard interrupts on autonomy, not a tone adjustment.
VI.5: The FCO Control Set
FCO-01: Covered-Decision Inventory and Fairness Scoping
Objective. Ensure every agent action that constitutes or materially influences a covered consumer decision is identified and in scope for fair-lending controls before deployment.
Control. The institution maintains a complete inventory of agentic flows whose actions constitute covered decisions (ECOA/Reg B credit actions, insurance eligibility, pricing, account closure); no agent takes a covered action unless its flow is inventoried with a fairness risk tier, jurisdiction-mapped legal regimes, and a named accountable owner.
Implementation (financial enterprise). Extend the SR 11-7 model inventory with an "agentic decision flow" record keyed to flow version (model + prompt + toolset + orchestration), not model alone. Fair-lending counsel maps each flow to ECOA (law.cornell.edu/uscode/text/15/1691), Reg B (law.cornell.edu/cfr/text/12/1002.9), and diverging state/federal disparate-impact regimes: HUD's proposed rescission (federalregister.gov/documents/2026/01/14/2026-00590/huds-implementation-of-the-fair-housing-acts-disparate-impact-standard) versus New Jersey's codification (njoag.gov/wp-content/uploads/2025/12/n.j.a.c.-13-16-disparate-impact-discrimination.pdf). Reconcile the inventory quarterly against production tool-call logs to catch flows that drifted into covered territory. [cadence: practice guidance — not directly source-backed]
Maturity. Baseline: manual inventory with legal mapping at onboarding. Enhanced: version-keyed inventory with change-triggered re-review. Frontier: automated covered-action detection from trajectory logs reconciled against the inventory.
Ownership. 1st line: business owners register flows and versions. 2nd line: fair-lending compliance assigns tiers and mappings; model risk validates completeness. 3rd line: audits completeness against production logs.
Evidence. Inventory extract with tiers, owners, mappings; reconciliation reports; re-review records.
Mappings. NIST AI RMF MAP; ISO/IEC 42001 AIMS; EU AI Act (creditworthiness high-risk category: article-level not sourced); SR 11-7/OCC 2011-12 (model inventory); DORA n/a.
Sources. law.cornell.edu/uscode/text/15/1691 (T1), https://www.law.cornell.edu/uscode/text/15/1691; law.cornell.edu/cfr/text/12/1002.9 (T1), https://www.law.cornell.edu/cfr/text/12/1002.9; federalregister.gov/documents/2026/01/14/2026-00590/huds-implementation-of-the-fair-housing-acts-disparate-impact-standard (T1), https://www.federalregister.gov/documents/2026/01/14/2026-00590/huds-implementation-of-the-fair-housing-acts-disparate-impact-standard; njoag.gov/wp-content/uploads/2025/12/n.j.a.c.-13-16-disparate-impact-discrimination.pdf (T1), https://www.njoag.gov/wp-content/uploads/2025/12/N.J.A.C.-13-16-Disparate-Impact-Discrimination.pdf
FCO-02: Constrained Decision Policy for Covered Actions
Objective. Prevent an agent from introducing undesigned discrimination by choosing its own features, data sources, or decision steps in covered decisions.
Control. Covered decisions run under an approved decision policy enumerating permissible data sources, features, verification steps, and criteria; the enforcement layer blocks retrieval or use of anything outside the approved set, and the agent cannot expand its own evidence set at runtime.
Implementation (financial enterprise). Implement as a deny-by-default allowlist at the tool gateway, change-controlled like any credit-policy amendment. Treat "the input list was approved" as necessary but never sufficient: input scrutiny alone does not establish non-discrimination (scholarship.law.umn.edu/minnlrev/vol106/iss3/2), and facially neutral inputs still produce group-differential outcomes (doi.org/10.1111/jofi.13090; lawcat.berkeley.edu/record/1127463), so FCO-02 is a precondition for FCO-03, not a substitute. Log and alert on every blocked out-of-policy retrieval: each is a prevented instance of discrimination-by-feature-choice and should be retained as evidence of control operation.
Maturity. Baseline: documented policy; manual trajectory adherence review. Enhanced: gateway enforcement with blocked-attempt alerting. Frontier: policy-as-code with per-trajectory conformance checks and versioning.
Ownership. 1st line: engineering operates the gateway. 2nd line: fair-lending compliance owns policy content; model risk validates enforcement. 3rd line: tests bypass resistance.
Evidence. Approved decision policy per flow version; gateway config exports; blocked-attempt logs; conformance results.
Mappings. NIST AI RMF GOVERN, MANAGE; ISO/IEC 42001 AIMS; EU AI Act (article-level not sourced); SR 11-7/OCC 2011-12 (limitations on model use); DORA n/a.
Sources. scholarship.law.umn.edu/minnlrev/vol106/iss3/2 (T1), https://scholarship.law.umn.edu/minnlrev/vol106/iss3/2/; doi.org/10.1111/jofi.13090 (T1), https://doi.org/10.1111/jofi.13090; lawcat.berkeley.edu/record/1127463 (T1): https://lawcat.berkeley.edu/record/1127463
FCO-03: Trajectory-Level Disparate-Impact Testing
Objective. Detect group-differential outcomes produced by the end-to-end agentic flow, including bias no component exhibits in isolation.
Control. Before deployment and after any material change, every covered flow is disparity-tested as a whole system across full trajectories against defined protected-class proxies and thresholds; component-level fairness results are inadmissible as evidence of system-level fairness.
Implementation (financial enterprise). Run cohort tests through the production orchestration stack, not a replica; the requirement rests on the demonstrated result that multi-agent systems produce "patterns of emergent bias in financial decision-making that cannot be traced to individual agent components" and "must be evaluated as holistic entities" (arxiv.org/abs/2512.16433). Metric set: adverse-impact ratio (eeoc.gov/select-issues-assessing-adverse-impact-software-algorithms-and-artificial-intelligence-used), disparate-impact certification (arxiv.org/abs/1412.3756), error-rate parity (arxiv.org/abs/1610.08452); document metric selection against fair-lending doctrine (doi.org/10.1145/3514094.3534154) and sign the impossibility trade-off explicitly (arxiv.org/abs/1703.00056). Use lending-domain tooling (github.com/solasai/solas-ai-disparity). Where protected class is unobserved, apply the institution's approved proxy methodology with documented limits. [proxy method: practice guidance — not directly source-backed]
Maturity. Baseline: pre-deployment cohort test on final outcomes. Enhanced: trajectory-level testing with treatment-intensity metrics; re-test per flow-version change. Frontier: continuous synthetic-cohort probing with matched-pair perturbation.
Ownership. 1st line: development runs tests in the release pipeline. 2nd line: model risk owns methodology and thresholds; compliance owns doctrine alignment. 3rd line: re-performs on sampled flows.
Evidence. Disparity reports per flow version; signed metric-selection memo; breach dispositions; re-test trigger log.
Mappings. NIST AI RMF MEASURE; ISO/IEC 42001 AIMS; EU AI Act (article-level not sourced); SR 11-7/OCC 2011-12 (outcomes analysis); DORA n/a.
Sources. arxiv.org/abs/2512.16433 (T2), https://arxiv.org/abs/2512.16433; eeoc.gov/select-issues-assessing-adverse-impact-software-algorithms-and-artificial-intelligence-used (T1), https://www.eeoc.gov/select-issues-assessing-adverse-impact-software-algorithms-and-artificial-intelligence-used; arxiv.org/abs/1412.3756 (T2), https://arxiv.org/abs/1412.3756; doi.org/10.1145/3514094.3534154 (T1), https://doi.org/10.1145/3514094.3534154; arxiv.org/abs/1703.00056 (T2), https://arxiv.org/abs/1703.00056; arxiv.org/abs/1610.08452 (T1), https://arxiv.org/abs/1610.08452; github.com/solasai/solas-ai-disparity (T3): https://github.com/SolasAI/solas-ai-disparity
FCO-04: Less-Discriminatory-Alternative Search and Business-Necessity File
Objective. Ensure that where a covered flow produces measurable disparity, less-discriminatory alternatives are searched and the retained configuration is justified before acceptance.
Control. A covered flow breaching disparity screening thresholds may not deploy until an LDA search across alternative models, prompts, decision policies, and orchestration configurations is performed and a business-necessity justification is documented and approved by the second line.
Implementation (financial enterprise). Agentic systems widen the LDA search space: prompts, tool ordering, escalation thresholds, and orchestration topology are cheap-to-vary alternatives alongside the model, and the emergent-bias result implies interaction topology is itself a fairness lever (arxiv.org/abs/2512.16433). Draw on disparity-repair techniques (arxiv.org/abs/1412.3756) and lending-evaluated explainability-fairness tooling (finreglab.org/research/machine-learning-explainability-fairness-insights-from-consumer-lending). File contents (alternatives tested, disparity/performance deltas, rejection rationale, signatures) written to hand an examiner unedited. [file format: practice guidance — not directly source-backed]
Maturity. Baseline: manual LDA memo on breach. Enhanced: standardized automated sweep over prompt/policy/model alternatives. Frontier: LDA search in the release pipeline with disparity-performance frontier reporting.
Ownership. 1st line: development runs the search. 2nd line: compliance approves business-necessity; model risk validates search adequacy. 3rd line: samples files for challenge quality.
Evidence. LDA search files; approvals; trade-off documentation.
Mappings. NIST AI RMF MANAGE; ISO/IEC 42001 AIMS; EU AI Act (article-level not sourced); SR 11-7/OCC 2011-12 (effective challenge); DORA n/a.
Sources. arxiv.org/abs/1412.3756 (T2), https://arxiv.org/abs/1412.3756; arxiv.org/abs/2512.16433 (T2), https://arxiv.org/abs/2512.16433; finreglab.org/research/machine-learning-explainability-fairness-insights-from-consumer-lending (T2): https://finreglab.org/research/machine-learning-explainability-fairness-insights-from-consumer-lending/
FCO-05: Specific and Accurate Adverse-Action Reasons for Agentic Decisions
Objective. Guarantee every adverse action taken or driven by an agent carries the specific, accurate statement of reasons the law requires, regardless of flow complexity.
Control. No covered flow issues an adverse action unless it produces, for that decision, principal reasons specific and accurate to the actual decision basis per ECOA §1691(d) and Regulation B; "the algorithm is too complex" and sample-form checkbox selection are prohibited as reason-generation methods.
Implementation (financial enterprise). Anchor to the statute (law.cornell.edu/uscode/text/15/1691), Reg B (law.cornell.edu/cfr/text/12/1002.9), and the CFPB's closure of both escape hatches: Circular 2022-03 on complex algorithms (consumerfinance.gov/compliance/circulars/circular-2022-03-adverse-action-notification-requirements-in-connection-with-credit-decisions-based-on-complex-algorithms) and Circular 2023-03 on sample forms (consumerfinance.gov/compliance/circulars/circular-2023-03-adverse-action-notification-requirements-and-the-proper-use-of-the-cfpbs-sample-forms-provided-in-regulation-b). Derive reasons from the recorded trajectory (FCO-06), not generic feature attribution: the operative cause may be an agent action, not a coefficient; validate reason fidelity counterfactually on a sample. [fidelity validation: practice guidance — not directly source-backed] Do not rely on the CFPB's disclaimed innovation-spotlight commentary (consumerfinance.gov/about-us/blog/innovation-spotlight-providing-adverse-action-notices-when-using-ai-ml-models). Trajectory classes that cannot meet the bar are blocked from adverse actions.
Maturity. Baseline: compliance-reviewed reason templates mapped to trajectory events. Enhanced: automated trajectory-derived reasons with sampled fidelity checks. Frontier: per-decision fidelity verification gating notice issuance.
Ownership. 1st line: builds reason generation into the flow. 2nd line: compliance tests notices against Reg B; model risk validates fidelity methodology. 3rd line: samples notices against trajectories.
Evidence. Notice-to-trajectory linkage; fidelity test results; compliance sign-offs; blocked-class register.
Mappings. NIST AI RMF MANAGE, GOVERN; ISO/IEC 42001 AIMS; EU AI Act (article-level not sourced); SR 11-7/OCC 2011-12 (use of model output); DORA n/a.
Sources. law.cornell.edu/uscode/text/15/1691 (T1), https://www.law.cornell.edu/uscode/text/15/1691; law.cornell.edu/cfr/text/12/1002.9 (T1), https://www.law.cornell.edu/cfr/text/12/1002.9; consumerfinance.gov/compliance/circulars/circular-2022-03-adverse-action-notification-requirements-in-connection-with-credit-decisions-based-on-complex-algorithms (T1): https://www.consumerfinance.gov/compliance/circulars/circular-2022-03-adverse-action-notification-requirements-in-connection-with-credit-decisions-based-on-complex-algorithms/; consumerfinance.gov/compliance/circulars/circular-2023-03-adverse-action-notification-requirements-and-the-proper-use-of-the-cfpbs-sample-forms-provided-in-regulation-b (T1), https://www.consumerfinance.gov/compliance/circulars/circular-2023-03-adverse-action-notification-requirements-and-the-proper-use-of-the-cfpbs-sample-forms-provided-in-regulation-b/; consumerfinance.gov/about-us/blog/innovation-spotlight-providing-adverse-action-notices-when-using-ai-ml-models (T2), https://www.consumerfinance.gov/about-us/blog/innovation-spotlight-providing-adverse-action-notices-when-using-ai-ml-models/
FCO-06: Decision Reproducibility and Trajectory Evidence Capture
Objective. Preserve a complete, replayable record of every covered agentic decision sufficient to explain, defend, and if necessary reverse it.
Control. Every covered decision trajectory is captured immutably (flow version, all tool calls with inputs/outputs, retrieved data, intermediate judgments, final decision, generated reasons) retained per policy and sufficient to reconstruct the decision basis without the live system. Cross-reference: RUN-18 is the trajectory-capture control of record and MON-01 owns the evidence-store properties; FCO-06 applies them to covered decisions, adding Reg B-grade retention and decision reconstruction.
Implementation (financial enterprise). This is the evidentiary substrate for FCO-03, FCO-05, FCO-08, and FCO-09: adverse-action defensibility under the CFPB circulars presumes the institution can state the actual basis of the decision (consumerfinance.gov/compliance/circulars/circular-2022-03-adverse-action-notification-requirements-in-connection-with-credit-decisions-based-on-complex-algorithms), which no institution without trajectory capture can honestly do for an agentic flow. Store trajectories keyed by applicant and flow version in WORM-capable storage, indexed by event type so analysts can query (e.g., all decisions where the agent demanded extra documentation) without replaying. Where nondeterminism prevents exact replay, capture per-step outputs so the record is deterministic even if the system is not. [storage design: practice guidance — not directly source-backed]
Maturity. Baseline: complete decision logs with retention controls. Enhanced: immutable, indexed, queryable trajectory store. Frontier: deterministic replay harness integrated with fairness analytics.
Ownership. 1st line: platform engineering operates capture. 2nd line: model risk sets completeness standards; compliance sets retention. 3rd line: tests completeness and immutability.
Evidence. Capture-completeness attestations; sampled reconstructions; retention/immutability configs; replay results.
Mappings. NIST AI RMF MEASURE, MANAGE; ISO/IEC 42001 AIMS; EU AI Act (record-keeping: article-level not sourced); SR 11-7/OCC 2011-12 (documentation); DORA (ICT records, aspect only).
Sources. consumerfinance.gov/compliance/circulars/circular-2022-03-adverse-action-notification-requirements-in-connection-with-credit-decisions-based-on-complex-algorithms (T1), https://www.consumerfinance.gov/compliance/circulars/circular-2022-03-adverse-action-notification-requirements-in-connection-with-credit-decisions-based-on-complex-algorithms/; law.cornell.edu/cfr/text/12/1002.9 (T1), https://www.law.cornell.edu/cfr/text/12/1002.9
FCO-07: Production Fairness Monitoring for Agentic Flows
Objective. Detect emerging disparities in live covered flows between formal testing cycles, including disparities from behavioral drift rather than code change.
Control. Covered flows are monitored in production on a defined cadence for outcome and treatment-intensity disparity across protected-class proxies, with alert thresholds, escalation paths, and mandatory disposition of every breach.
Implementation (financial enterprise). Agentic flows drift by conduct, upstream model updates change behavior without a release, so periodic validation alone is insufficient; monitor continuously or near-continuously on high-volume flows. [cadence rationale: practice guidance — not directly source-backed] Minimum metrics: approval/denial and pricing disparity, documentation-request and escalation rates, time-to-decision by group, using lending-domain measures (github.com/solasai/solas-ai-disparity). For mortgage flows, reconcile against the institution's own HMDA submissions: the public dataset regulators and researchers will use to find your disparities for you (consumerfinance.gov/data-research/hmda), as the FinTech-era discrimination literature demonstrates (doi.org/10.1016/j.jfineco.2021.05.047). Breaches route to the fair-lending committee: remediate, justify, or halt.
Maturity. Baseline: quarterly disparity reporting. Enhanced: automated near-real-time monitoring with thresholds. Frontier: monitoring joined to trajectory analytics that localizes which agent behavior drives a moving disparity.
Ownership. 1st line: operates monitoring and first response. 2nd line: model risk owns thresholds; compliance owns breach disposition. 3rd line: audits alert handling.
Evidence. Dashboards and threshold configs; breach tickets with dispositions; HMDA reconciliation memos; committee minutes.
Mappings. NIST AI RMF MEASURE, MANAGE; ISO/IEC 42001 AIMS; EU AI Act (post-market monitoring: article-level not sourced); SR 11-7/OCC 2011-12 (ongoing monitoring); DORA n/a.
Sources. consumerfinance.gov/data-research/hmda (T1), https://www.consumerfinance.gov/data-research/hmda/; doi.org/10.1016/j.jfineco.2021.05.047 (T1), https://doi.org/10.1016/j.jfineco.2021.05.047; github.com/solasai/solas-ai-disparity (T3): https://github.com/SolasAI/solas-ai-disparity
FCO-08: Complaint Handling and Human Reopening of Agentic Decisions
Objective. Ensure customers aggrieved by an agentic decision have an effective route to human review, and that complaints function as a fairness sensor.
Control. Every covered agentic decision is contestable: complaints tag agent-involved cases at intake, guarantee human re-decision by a reviewer with authority and trajectory access under defined SLAs, and feed complaint analytics into fairness monitoring (FCO-07) as a detection input.
Implementation (financial enterprise). [practice guidance: not directly source-backed, except statutory anchors] Tag agent-involved complaints distinctly so agent-caused harm is quantifiable to the board. The reviewer must see the captured trajectory (FCO-06), must not receive the agent's conclusion as a default-accept recommendation, rubber-stamping reintroduces the disparity under a human signature, and must have authority to reverse. Cluster complaints by flow version: a spike on one version is an incident signal, not a service issue. Reversal-rate asymmetry across protected-class proxies is itself a disparity metric; route it to FCO-07. Statutory floor: ECOA/Reg B adverse-action rights attach to the re-decision exactly as to the original (law.cornell.edu/uscode/text/15/1691; law.cornell.edu/cfr/text/12/1002.9).
Maturity. Baseline: agent tagging and human re-decision on request. Enhanced: SLA-managed reopening with trajectory-equipped reviewers and version clustering. Frontier: complaint analytics fused with fairness monitoring; reversal patterns auto-trigger flow review.
Ownership. 1st line: complaints operations own intake and re-decision. 2nd line: compliance oversees SLAs and analytics; model risk consumes reversal data. 3rd line: audits reopening independence and rubber-stamp risk.
Evidence. Complaint taxonomy with agent tags; SLA reports; reversal-rate analyses by group; board reporting.
Mappings. NIST AI RMF MANAGE, GOVERN; ISO/IEC 42001 AIMS; EU AI Act (article-level not sourced); SR 11-7/OCC 2011-12 (governance); DORA n/a.
Sources. law.cornell.edu/uscode/text/15/1691 (T1), https://www.law.cornell.edu/uscode/text/15/1691; law.cornell.edu/cfr/text/12/1002.9 (T1), https://www.law.cornell.edu/cfr/text/12/1002.9
FCO-09: Cohort Remediation and Fairness Lookbacks
Objective. Ensure that when a fairness defect is found in an agentic flow, every affected customer, not only complainants, is identified and made whole.
Control. Upon a confirmed fairness defect (testing or monitoring breach, complaint cluster, exam finding), the institution identifies the full cohort decided by the defective flow version, re-decisions them under the corrected flow, remediates adverse deltas including corrected notices, and reports material events per regulatory obligations.
Implementation (financial enterprise). [practice guidance: not directly source-backed, except statutory anchors] Trajectory capture (FCO-06) pays for itself here: cohort identification is a query over flow-version-keyed trajectories, and re-decisioning replays the cohort through the corrected flow rather than adjudicating case-by-case. Size remediation for correlated failure, not idiosyncratic error. Where re-decision changes the outcome or basis, Reg B notification duties attach to the corrected decision (law.cornell.edu/cfr/text/12/1002.9). Pre-agree lookback triggers and remediation standards with legal so speed is not gated on deliberation; track completion as a board-visible metric.
Maturity. Baseline: documented lookback procedure executed manually. Enhanced: automated cohort identification and batch re-decisioning from the trajectory store. Frontier: pre-authorized remediation playbooks with tested execution paths and regulator-ready templates.
Ownership. 1st line: business and engineering execute lookbacks. 2nd line: compliance sets remediation standards and reporting; model risk certifies the corrected flow. 3rd line: verifies cohort completeness and adequacy.
Evidence. Trigger log; cohort queries and counts; re-decision results; remediation records; corrected notices.
Mappings. NIST AI RMF MANAGE; ISO/IEC 42001 AIMS; EU AI Act (corrective actions: article-level not sourced); SR 11-7/OCC 2011-12 (issue management); DORA (incident response, aspect only).
Sources. law.cornell.edu/cfr/text/12/1002.9 (T1), https://www.law.cornell.edu/cfr/text/12/1002.9; law.cornell.edu/uscode/text/15/1691 (T1), https://www.law.cornell.edu/uscode/text/15/1691
FCO-10: Vulnerable-Customer Safeguards in Agent Interactions
Objective. Prevent autonomous agents from exploiting, pressuring, or disadvantaging customers with reduced capacity to protect their own interests.
Control. Customer-facing agents in covered or conduct-sensitive contexts (collections, retention, sales, hardship) operate under enforced vulnerability rules: defined signals (financial-distress indicators, disclosed hardship, indicia of cognitive impairment, limited English proficiency, third-party representation) act as hard interrupts that constrain agent objectives, cap contact frequency and persuasion tactics, and route to designated human handling.
Implementation (financial enterprise). [practice guidance: not directly source-backed, except as noted] Encode triggers in the runtime policy layer, not the prompt: prompt-level instructions yield to objective pressure; policy-layer interrupts do not. On trigger: freeze objective-driven persuasion, disable urgency/scarcity tactics, cap contact frequency, offer human handoff, log for conduct review. Limited-English handling has an evidentiary basis beyond courtesy, LLM systems demonstrably perform worse on non-native English writers (arxiv.org/abs/2304.02819), so include it both as an interaction safeguard and as a tested dimension in FCO-03 matched-pair testing. Red-team the agent with simulated vulnerable-customer personas before release and periodically after; report trigger volumes and vulnerable-cohort outcomes versus baseline to the conduct-risk committee.
Maturity. Baseline: documented vulnerability policy with prompt-level rules and human handoff. Enhanced: policy-layer triggers with logging and persona-based release testing. Frontier: vulnerable-cohort outcome-parity monitoring integrated with FCO-07 and board conduct reporting.
Ownership. 1st line: business and engineering implement triggers and handoffs. 2nd line: conduct risk owns the vulnerability taxonomy and outcome review. 3rd line: audits trigger efficacy and tests circumvention.
Evidence. Vulnerability taxonomy and trigger configs; interrupt logs; persona red-team reports; cohort outcome analyses; committee minutes.
Mappings. NIST AI RMF GOVERN, MANAGE; ISO/IEC 42001 AIMS; EU AI Act (article-level not sourced); SR 11-7/OCC 2011-12 (governance: aspect); DORA n/a.
Sources. arxiv.org/abs/2304.02819 (T2): https://arxiv.org/abs/2304.02819
VI.6: Closing note for the Chief AI Risk Officer
The fair-lending field study's lesson frames the whole set: five decades of tolerably fair algorithmic credit decisioning came not from artifacts but from a standing supervisory function with authority to compel change (arxiv.org/abs/2606.02957). The FCO controls give your internal supervisory function that same standing over agents: complete visibility (FCO-01, FCO-06), enforced constraints (FCO-02, FCO-10), system-level measurement (FCO-03, FCO-07), and the authority to force alternatives, reversals, and remediation (FCO-04, FCO-05, FCO-08, FCO-09). An agent that can choose its own features and steps must answer to a function that can choose to stop it.