AUTOGOVERN FINANCE · FREE PUBLIC VISION BOARD

Faster decisions.
Defensible AI.

Make every AI-assisted financial decision explainable to the customer, defensible to the supervisor and reversible by an accountable person.

Explore the opportunities ↓Download detailed board (.md)

Free to read and download. No account, email or sign-in required.

THE NON-NEGOTIABLE FOUNDATION

Design fairness, model risk and resilience into the decision.

Suggested controls to validate in the actual product and jurisdiction. Customer outcomes, model soundness, explainability, resilience and accountability need evidence throughout the lifecycle.

Customer outcomes and fairness

Test for disparate treatment and impact across prohibited bases and proxies. Design for good outcomes, not just accurate scores; a model that is right on average can still harm a segment.

Model risk discipline

Inventory every model and AI system, tier by materiality, validate independently with effective challenge, monitor continuously and revalidate on change — the spine of SR 26-2, PRA SS1/23, OSFI E-23 and MAS expectations.

Explainability and consumer rights

Every adverse decision needs specific, accurate reasons (Regulation B, FCRA), a dispute route and a human who can override. Automated decisions in the EU and UK carry GDPR Article 22 duties.

Security, resilience and controlled actions

Authenticate people and services; enforce entitlements before retrieval; keep kill switches and pre-trade limits outside the model; register and test ICT third parties (DORA); rehearse recovery.

Data lineage and quality

Trace every feature to its golden source with BCBS 239 discipline. Consumer-report and special-category data carry their own permissible-purpose and retention rules.

Evidence and accountability

Three lines of defence with named owners. Record versions, approvals, findings, monitoring freshness and human decisions so a supervisor can reconstruct why a decision was made.

Model risk references: SR 26-2, PRA SS1/23; fair lending: Regulation B.

THE VISION BOARD

Eight opportunities. Three adoption horizons.

These are proposed uses, not live services or readiness ratings. Expand each card for its data, human authority, controls, owner, evidence, monitoring and fallback.

01 / Establish the foundation

Begin with bounded assistance in operations, servicing and governance work. Lower autonomy still needs an inventory entry, conduct review and monitoring.

Customer service & servicing assistantAnswer product and account questions, draft responses for agents and route customers to the right person faster.Explore controls & delivery details +

Data involved: Approved product terms, fee schedules, policy documents and the minimum account context the channel already shows. No card numbers, credentials or full statements in prompts.
Human authority and limits: No advice, no fee waivers, no account changes and no collections decisions without a person. Route vulnerability signals and complaints to staff.
Controls to implement: Bound topics and retrieval sources; disclose that the customer is speaking to AI; test out-of-scope, urgent and vulnerable-customer conversations; Consumer Duty / UDAAP foreseeable-harm review; record interactions.
Accountable team: Head of customer operations with compliance and conduct reviewers.
Evidence before use: Accuracy and harm test results, escalation drills, injection and leakage tests, accessibility review.
What to monitor: Unsupported answers / reviewed answers; escalations attempted / completed; complaints mentioning the assistant; abandonment.
When to pause and fallback: Pause affected intents on misleading answers or failed escalation; keep phone, chat-to-agent and branch service staffed.

Onboarding & document operationsExtract and check documents for KYC/KYB, reconcile records and summarise cases so analysts spend time on judgement.Explore controls & delivery details +

Data involved: Identity documents, corporate records and transaction references already collected under the onboarding programme; retain per the existing schedule.
Human authority and limits: Analysts make the onboarding, risk-rating and exit decisions. AI proposes extractions and matches; it does not approve or close a case.
Controls to implement: Confidence thresholds with human review; duplicate and wrong-entity checks; sanctions and PEP screening remain in the approved engine; audit trail of proposals versus decisions.
Accountable team: Financial-crime operations lead with compliance and data owners.
Evidence before use: Extraction accuracy by document type, false-match tests, analyst override analysis and access-control tests.
What to monitor: Corrections / proposals; wrong-entity events; queue age; override rate by analyst and document type.
When to pause and fallback: Disable affected extraction or matching on wrong-entity or quality failures; analysts continue from source documents.

AI-assisted compliance & governanceMap policies to obligations, draft control tests, triage complaints and prepare validation and board packs.Explore controls & delivery details +

Data involved: Policies, control descriptions, test summaries and minimised evidence references. Customer details stay in complaint and case systems.
Human authority and limits: AI proposes; control owners, validators and compliance decide. AI cannot certify compliance, close findings or accept residual risk.
Controls to implement: Source-linked suggestions, access-scoped retrieval, independent review, retained correction history; treat documents as untrusted input.
Accountable team: Head of compliance / model risk with second- and third-line reviewers.
Evidence before use: Citation-support tests, deliberately missing or conflicting evidence, reviewer disagreement and prompt-injection tests.
What to monitor: Supported citations / reviewed citations; missed obligations; reviewer rejection rate; overdue findings.
When to pause and fallback: Suspend unreliable mappings or summaries; continue manual review from the original evidence.

02 / Validate regulated decisions

Introduce models into credit, fraud, underwriting and pricing only with independent validation, fair-lending evidence, working adverse-action processes and a rehearsed fallback.

Credit decisioning & pricingScore applications, set risk-based pricing and prioritise manual underwriting with consistent, explainable decisions.Explore controls & delivery details +

Data involved: Application data, bureau and consumer-report data with permissible purpose, internal performance history and validated features with lineage.
Human authority and limits: The decision engine applies approved policy; declines near threshold and edge cases route to underwriters. No self-modifying rules or unvalidated challenger models in production.
Controls to implement: Independent validation (conceptual soundness, outcomes analysis, monitoring); fair-lending testing incl. proxies; verified adverse-action reason codes and FCRA notices; decision records with input snapshots; tier-appropriate approval.
Accountable team: Chief credit officer with model risk, fair-lending compliance and data owners.
Evidence before use: Validation report, fairness and explainability tests, reason-code stability, dispute handling and monitoring with denominators.
What to monitor: Approval and pricing outcomes by segment; calibration and drift; reason-code distribution; overrides; complaints and disputes; policy exceptions.
When to pause and fallback: Route to manual underwriting or the last approved version on breach; identify and remediate affected applicants; record the decision.

Fraud & AML transaction monitoringDetect suspicious activity earlier with fewer false positives and clearer analyst prioritisation.Explore controls & delivery details +

Data involved: Transaction, device, session and counterparty signals; investigation outcomes as labels; sanctions and typology references from the approved programme.
Human authority and limits: Analysts decide on SARs, holds and exits. Automated blocking only within approved rules with customer-safe release paths.
Controls to implement: Validation of alerting models (SR 26-2 covers BSA/AML models); threshold governance; typology coverage tests; explainability for investigators; false-positive and false-negative tracking; segregation from customer-facing explanations.
Accountable team: BSA/AML officer and fraud lead with model risk and technology.
Evidence before use: Backtesting against confirmed cases, above-and-below-the-line testing, coverage review and change records.
What to monitor: Alert precision and recall by typology; queue age; blocked-legitimate rate; missed-case reviews; data-feed freshness.
When to pause and fallback: Revert to the prior rule set or model on coverage or precision breach; keep manual review capacity.

Insurance underwriting, pricing & claimsSpeed up quoting, underwriting and claim triage with consistent, testable decisions.Explore controls & delivery details +

Data involved: Application and policy data, approved external data sources with documented provenance, claims history and adjuster outcomes.
Human authority and limits: Actuaries and underwriters own rating factors and exceptions; adjusters decide claims. AI proposes classifications, triage and fraud indicators.
Controls to implement: Written AI Systems Program (NAIC bulletin states); quantitative proxy and bias testing (NY DFS Circular 7, Colorado SB 21-169); ASOP 56 model governance; external-data due diligence; adverse-decision explanations and appeals.
Accountable team: Chief underwriting officer and chief actuary with compliance and data owners.
Evidence before use: Proxy-testing results, actuarial review, external-data quality evidence, filed attestations where required, appeal outcome analysis.
What to monitor: Decision outcomes by protected class and proxy; loss ratio by segment; claim cycle time; appeal reversals; external-data drift.
When to pause and fallback: Suspend a rating factor, data source or triage rule on adverse test results; revert to the approved manual process.

03 / Expand with evidence

Extend to markets, advice and higher-autonomy agents only when monitoring, controls and supervisory expectations demonstrably hold. A maturity sequence, not a release calendar.

Algorithmic execution & market surveillanceImprove execution quality and surveillance coverage with tightly controlled automation.Explore controls & delivery details +

Data involved: Order, execution, market and reference data; surveillance alerts and case outcomes.
Human authority and limits: Pre-trade risk controls and kill switches remain outside the model. No AI-generated order flow without limits, testing and supervisory sign-off.
Controls to implement: SEC Rule 15c3-5 pre-trade controls and annual certification; Reg SCI resilience where applicable; recordkeeping of AI outputs; model validation and change control; FINRA supervision and communications rules.
Accountable team: Head of electronic trading and chief compliance officer with technology risk.
Evidence before use: Control limit tests, kill-switch drills, backtests, surveillance coverage tests and recordkeeping verification.
What to monitor: Limit breaches; execution quality; surveillance alert quality; latency and availability; model changes.
When to pause and fallback: Halt affected order flow through the independent kill switch; investigate before restart.

Investment advice, wealth & marketingPersonalise guidance and marketing while meeting suitability, disclosure and recordkeeping duties.Explore controls & delivery details +

Data involved: Client profiles, suitability records and approved product information under existing consent and privacy rules.
Human authority and limits: Advisers and compliance own suitability and recommendations. No unsupervised advice, no AI claims in marketing that the firm cannot substantiate.
Controls to implement: Advisers Act compliance and marketing rules; FCA Consumer Duty and suitability; conflicts testing; disclosure of AI use; supervision and retention of AI-assisted communications.
Accountable team: Head of wealth and chief compliance officer.
Evidence before use: Suitability tests, conflict and steering analysis, disclosure samples, supervision records and complaint outcomes.
What to monitor: Recommendation dispersion and conflicts; complaint and reversal rates; disclosure coverage; marketing claim substantiation.
When to pause and fallback: Withdraw affected recommendations or campaigns; revert to adviser-only process; notify compliance.

Assess each use against the regimes for its entity type and jurisdiction: SR 26-2 for US bank models, the NAIC bulletin for insurers, DORA for EU entities. A horizon does not establish regulatory status.

FROM REQUEST TO ACCOUNTABLE OUTCOME

A controlled integration path

1 / Authorized request
A customer, employee or scheduled process starts a bounded task. The server resolves identity, entitlements, product context and permissible purpose before any data moves.
2 / Policy and data boundary
Check the approved release, tier and policy version. Retrieve only authorised, lineage-tagged data; consumer-report and special-category data pass through their own permissible-purpose checks.
3 / Controlled model & AI service
Registered model versions, approved endpoints and tools, bounded context and spend. Untrusted content cannot expand entitlements or change recipients.
4 / Review and execution
Present a labelled proposal with reason codes and sources. Bind required human approval to the exact action; execute only through the decision engine or core banking APIs, never by direct database writes.
5 / Evidence and feedback
Store minimised decision records, reason codes, overrides and monitoring observations. Route incidents and complaints to accountable teams; breaches trigger revalidation.

The core system and decision engine remain the system of record. Models return scores and reasons; the decision engine applies approved policy and writes the decision. Governance receives minimized evidence references. Stop controls disable affected AI use while keeping the ordinary process available.

Read the detailed AI action controls →

WHO APPROVES WHAT

Five gates to maintain accountability

Scope, ownership & tiering

Owner: Model owner + model risk management

Define purpose, decision, customers affected, prohibited actions, model type and materiality tier. Name the independent validator and who can stop use.

Evidence: Inventory record, tier rationale and named decision owners.

Decision condition: A bounded use with an accountable owner and tier; unresolved scope is a hold.

Data, third-party & legal approval

Owner: Data owner + third-party risk + compliance + legal

Map lineage and data classes, confirm permissible purpose and consent, review vendor contracts and DORA / interagency third-party requirements, and confirm applicable regimes by jurisdiction and entity type.

Evidence: Lineage register, permissible-purpose rationale, third-party decision and applicability note.

Decision condition: Approved data paths and services; no unapproved provider or data source.

Independent validation & fairness testing

Owner: Independent validation + fair-lending compliance

Set acceptance limits, then test conceptual soundness, outcomes, segments, fairness (including proxies) and explainability. Record findings and conditions.

Evidence: Validation report with effective-challenge conclusion, fairness and reason-code evidence.

Decision condition: Findings closed or accepted by the right authority; unknown coverage is not a pass.

Controlled release with consumer safeguards

Owner: Model risk committee / release authority + conduct reviewers

Approve an exact version, scope and expiry; confirm adverse-action, FCRA, complaint and override processes; complete conduct and jurisdiction reviews (Consumer Duty, NAIC, AI Act high-risk readiness); rehearse the fallback.

Evidence: Release decision, safeguards checklist, fallback drill record.

Decision condition: All required approvals and runtime limits are active for that release.

Monitor, change & retire

Owner: Model owner + monitoring + operational resilience

Review performance, drift, fairness, complaints, incidents and provider changes. Trigger revalidation on breach or material change; report incidents within regulatory clocks; retire cleanly.

Evidence: Fresh metrics with denominators, incident records, revalidation decisions and retirement evidence.

Decision condition: Continue, restrict, suspend or retire through a recorded human decision.

Assign named people and delegates across the three lines of defence. Separate model development from validation and change authorship from required review. AI can prepare evidence; model risk, compliance, legal and conduct owners make the decisions. This operating model is a suggested use of NIST's voluntary AI RMF resource and the FS AI RMF.

TURN THE VISION INTO A PROGRAM

A suggested first 90 days—and beyond

A planning cadence, not a promise that production use will be ready in 90 days. Advance only when evidence and required approvals support it.

First 30 days / discover

Inventory models and AI already in use, assign owners, tier by materiality, map one decision end to end and identify the applicable regimes. Output: inventory baseline and gap register.

Days 31–60 / prove the controls

Complete lineage and third-party reviews for one candidate; set validation acceptance limits; test adverse-action, override, monitoring and fallback processes on synthetic cases. Output: evidence-backed readiness review.

Days 61–90 / decide on a controlled release

Only if validation, fairness and conduct evidence support it, approve a limited release with monitoring and stop criteria; otherwise continue remediation. Output: recorded release or hold decision.

Beyond 90 days / earn expansion

Compare outcomes, fairness, complaints and burden against baseline; revalidate on change; extend to new products, segments or jurisdictions one decision at a time. Output: scoped expansion decisions and attestation-ready evidence.

Fund independent validation, fair-lending testing, data lineage work, third-party reviews, monitoring, training and support. See the detailed assumptions and effort estimates.

MEASURE BENEFIT AND HARM TOGETHER

Success requires more than model accuracy.

Choose decision-specific metrics and acceptance limits before deployment. Record numerator, denominator, sample size, uncertainty, segment coverage, model version, data source and last collection time. Compare approval and pricing outcomes by segment, complaint and dispute rates, override burden, losses, cycle time and customer outcomes against a baseline. Disconnected monitoring is unknown, not green. Model risk and compliance leaders define stop thresholds; there is no universal safe accuracy percentage.

Customers

Fair outcomes by segment, understandable reasons, timely decisions, working dispute routes and access to a person.

Business and risk teams

Losses and cycle time, override burden, validation currency, open findings and confidence in the fallback.

Governance

Coverage of approved uses, fresh evidence, unresolved findings, incident response, reporting on time and verified remediation.

Run a vision-to-release workshop

Choose one opportunity and write its intended benefit, customer segments and forbidden actions.
Name the accountable owners, tier the model and map the exact data and third-party paths.
Set acceptance and pause criteria; identify evidence gaps and assign dates.
Record a release or hold decision after the gates. Keep customer details in authorised systems.

Use the free planning templates →Download every card and gate

Scope and primary sources

A proposed adoption model for banks, lenders, insurers and investment firms, reviewed September 12, 2026. These are planning ideas and suggested controls, not deployed integrations, legal advice, regulatory approval or a guarantee of model performance. AutoGovern currently supports finance governance metadata, Workbench tools and public learning; live model monitoring and decision-system integration are not connected.

In AutoGovern, Finance Governance supports organization metadata records. Use the public how-tos for practical steps and the architecture report for the full technical design and capability gaps.