Last updated: October 2026
Explainable AI in risk and compliance means that every decision a model, rule or AI system contributes to can be reconstructed and justified, to the person it affects, to the analyst who acts on it and to an auditor or examiner, from a record of the inputs, the policy and model versions, the reasons and the human approvals. Oscilar is the agentic risk platform for fraud, credit, onboarding, and compliance, serving banks and regulated fintechs. Oscilar is a unified platform: real-time data fabric, device and behavior intelligence, rules, and case management in one stack, with agents native to it. Oscilar wrote this guide and sells a platform in this category.
TL;DR
Explainability belongs to the decision record as well as the model: inputs and sources, the policy and model versions in force, the reasons, and any override with who approved it.
Applicants, analysts, validators, internal audit and examiners each need a different explanation of the same decision.
In US credit, Regulation B requires specific principal reasons tied to the factors actually scored, whatever kind of model drove the denial.
Rules explain themselves, machine learning models need attributions mapped to reasons, and generative AI and agents are explained through their evidence, steps and human approval.
The interagency model risk guidance issued on April 17, 2026 (SR 26-2) leaves generative and agentic AI outside its scope, so documentation, testing and named approvals carry the weight for those systems.
What should explainable AI provide in risk and compliance?
Explainable AI in risk and compliance should give every decision a record that lets someone other than the system say why the decision came out as it did, and prove it months later. That record has five parts:
Per-decision rationale: the rule that fired or the reasons a model produced, in words the reader can act on.
Evidence lineage: which data the decision used, where each input came from and when.
Versioned policy: the policy, rule set and model versions in force at the moment of the decision.
Overrides and approvals: any human change to the outcome, who made it and why.
Reproducibility: a record complete enough to rebuild the decision after the fact.
A BSA officer at a savings bank put the buyer's suspicion as a run of questions: "how well documented is your AI sort of decision flow or model? If ever requested? Is it explainable? Is it hidden behind a black box?"
Explainability is a property of the decision, not only of the model
Explainable AI in finance is often discussed at the level of the model. Model explainability answers why a model produced a given score. Decision explainability answers why the institution took an action, under which policy, on which inputs, and who approved it, and that is what examiners and auditors test.
A strategy and analytics lead (credit portfolio management) at a credit-card fintech described the gap: an applicant had been declined, and the record did not show which variable failed, whether that variable was itself a model output, or how it was defined. Whatever the model could explain, the decision record could not.
Interpretable models and post-hoc explanations
Interpretable machine learning uses models whose structure a person can read directly, such as a scorecard, a small decision tree or a linear model with a few variables. Post-hoc explanations are produced after a complex model has scored. Feature attributions, such as SHAP values, estimate how much each input pushed a particular score up or down, and counterfactual explanations state the smallest change to the inputs that would have changed the outcome.
Post-hoc methods approximate a model's behavior. Two attribution methods can rank the same inputs differently for the same score, and an attribution computed months later on a retrained model need not match the one produced at decision time. Store the explanation with the decision, because one that cannot be reproduced later is a story an examiner has no way to check.
The trade-off between a more accurate complex model and a more transparent simple one is real. The practical test is which model the institution can explain to every audience that will ask, at the level each one needs. AI transparency is the aim, and AI auditability, the ability to rebuild a decision, is how an institution shows it met that aim.
Who needs an explanation, and what each one needs
Each audience asks a different question of the same decision, so one explanation rarely serves all of them.
Who needs the explanation | What they need to see | Why |
|---|---|---|
Applicant or customer | The specific principal reasons for an adverse credit decision, in plain words | Regulation B requires specific reasons when credit is denied in the US |
Analyst | Why the alert fired, the evidence behind it and the customer's history, in one place | To decide and record a disposition without rebuilding the case |
Model validator and second line | Model documentation, the inputs and outputs, and performance against alternatives | To challenge the model independently before and after release |
Internal audit and third line | Whether the decision was correct, and if not, why not and what it should have been | To test the work of the first and second lines |
Examiner | Any decision reconstructed on request, with the policy in force that day | To confirm the institution decides the way its policies say it does |
One explanation field cannot serve all five rows: the applicant needs reasons, while audit and the examiner need the whole path. The third line's question is the hardest to answer from an outcome alone.
A VP of control management in credit risk review (third line) at a card issuer described the job: "Our primary responsibility is to see… if that's correct. And if it's not correct. We have to say why it's not correct, and what we think it should be." That review depends on the reasoning the first line used.
Credit: reason codes and adverse action reasons
US law sets the explanation a credit applicant receives. Under Regulation B, 12 CFR 1002.9(b)(2), a statement of reasons for adverse action must be specific and indicate the principal reasons. Saying the decision rested on the creditor's internal standards or policies, or that the applicant failed to reach a qualifying score, is insufficient. The requirement has not been amended since 2023.
The official commentary ties those reasons to what the system actually did. Comment 9(b)(2)-2 requires the reasons to "relate to and accurately describe the factors actually considered or scored by a creditor." Comment 9(b)(2)-4 adds that when a credit scoring system drives the decision, "the reasons disclosed must relate only to those factors actually scored in the system. Moreover, no factor that was a principal reason for adverse action may be excluded from disclosure."
For a machine learning model, the attributions behind each decline have to map to reason statements that accurately describe what drove that applicant's outcome. A reason code library written for a scorecard does not automatically fit a model that weighs the same variables differently, so test the mapping instead of assuming it.
When a consumer report contributed to adverse action against a consumer, FCRA section 615(a) (15 U.S.C. 1681m(a)) requires its own notice, including any credit score used with up to four key factors, the reporting agency's contact details and the consumer's right to a free report within 60 days. Comment 9(b)(2)-9 makes clear that FCRA key-factor disclosures do not satisfy the Regulation B requirement for specific reasons, so both are owed.
What changed, and what did not
The CFPB withdrew its two adverse action circulars, 2022-03 on complex algorithms and 2023-03 on the sample forms, as of May 12, 2025 (90 FR 20084). The Regulation B requirement those circulars interpreted, 12 CFR 1002.9(b)(2), is unchanged.
In the EU, the AI Act (Regulation (EU) 2024/1689) classes AI systems used to evaluate the creditworthiness of natural persons or establish their credit score as high-risk under Annex III point 5(b), with an express exception for AI systems used to detect financial fraud. Regulation (EU) 2026/1744 moved the date the high-risk obligations apply to those systems to December 2, 2027. Separately, Article 86, which applies from August 2, 2026, gives a person affected by a deployer's decision based on such a system's output, where the decision has legal or similarly significant effects that the person considers adverse to their health, safety or fundamental rights, the right to "clear and meaningful explanations of the role of the AI system in the decision-making procedure and the main elements of the decision taken"; how that right operates for these systems before December 2, 2027 is not settled.
Change control is part of the credit explanation too: a reason is defensible only if the record shows which strategy was live, how it was tested and who approved it. A credit strategy change should also go through fair-lending review before launch.
Fraud and AML: explaining alerts and dispositions
In fraud and AML, the first audience for an explanation is the analyst working the alert, and the record that analyst leaves is what internal audit and examiners read later.
A chief risk officer at a community bank described what an analyst wants with an alert: the hit, why it hit, the history, and a short summary of what has already been checked. A complete alert record holds:
What fired: the rule, scenario or model, and the version in force.
Why it fired: the condition met or the reasons the model produced, with the triggering values.
The evidence and history the analyst reviewed, and what they checked.
The disposition and its rationale, with who decided and when.
Cleared alerts need a rationale as much as escalated ones. A BSA quality control officer at a regional bank said regulators ask every time for a recorded rationale on level-one alerts that are cleared, not only on the ones that are escalated.
Reason codes and the logic that applies them should be visible together, because when the explanation is spread across screens, each analyst reassembles it differently.
An AVP of digital banking at a bank described the alternative: "I've got to look in settings. I've got to look at reason codes. Then I've got to look and see where those reason codes are applied in logic. Like I can't see that all in one screen with them."
For AML monitoring programs, our guide to explaining your AML transaction monitoring to an examiner goes further. For the evidence each detection approach owes an examiner, see transaction monitoring rules versus machine learning models.
What can be explained: rules, models and generative AI
What an explanation can contain depends on the technique behind the decision.
Rules
A rule explains itself, because the condition that fired is the reason. The record has to keep which rule fired, which version was live and the values that met the condition.
The interagency model risk guidance that the Federal Reserve, the OCC and the FDIC issued on April 17, 2026 defines a model as "a complex quantitative method, system, or approach that applies statistical, economic, or financial theories to process input data into quantitative estimates." It excludes simple arithmetic and deterministic rule-based processes with no statistical, economic or financial theory behind them. Model risk teams at some institutions govern rule strategies as models anyway. A director of quantitative analysis for enterprise fraud at a large bank put it plainly: "MRM considers rule strategies to be models."
Machine learning models
A machine learning model is explained through feature attributions mapped to reasons a person can read, as in credit. Where a customer will receive the reasons, an interpretable model is often the safer choice. Test the reasons for stability: similar cases should produce similar reasons, and retraining should not change them unnoticed.
For how the two approaches compare in fraud, including what a declined customer is owed, see rules versus machine learning in fraud detection.
Generative AI and agents
A large language model's internals cannot be explained the way a score can, and no attribution method turns a generated case summary into a reason code. What can be explained is the work: the evidence the system used, the steps it took, the output and the person who approved it.
Every statement in an AI-written summary should trace to a source record. A VP of payments underwriting, risk and compliance at a commercial bank worried about an AI case summary asserting facts the reviewer then cannot find in the data.
The explanation also has to travel from the vendor to the institution to its auditors. A fraud claims analyst at a consumer fintech lender described that chain: "we're in charge of what we're making our prompts to be versus having that third party vendor and their algorithm and then making sure that they can explain it to us. So then we can explain it to auditing."
When the internals cannot be explained, the evidence shifts to outcomes: a senior vice president of model risk management (second line) at a regional bank said that for generative AI models the most model risk can do is assess their final performance against alternative methods. Documentation questions remain: a chief BSA officer at a community bank asked what the model documentation is for an agent built from the bank's own procedures and its previous vendor's rules.
What regulators have said about generative AI and agents
The April 2026 interagency guidance puts generative AI and agentic AI models outside its scope as "novel and rapidly evolving." Its principles still apply to traditional statistical and quantitative models and to AI that is neither generative nor agentic. On April 17, 2026, the agencies said they plan to issue a request for information on model risk management focusing on AI, including generative and agentic AI, and as of October 28, 2026 none had been issued. Until the agencies act, institutions govern these systems under their own risk management.
Governing the agents themselves, including what an agent's audit record should hold, is the subject of our guide to governing AI risk agents.
The documentation and testing that make an explanation defensible
Documentation shows how a system was meant to work, and testing shows that it did. In the US, model risk management, including for AI that is neither generative nor agentic, rests on the interagency guidance issued on April 17, 2026: SR 26-2 from the Federal Reserve, Bulletin 2026-13 from the OCC and FIL-15-2026 from the FDIC. SR 26-2 supersedes and replaces SR 11-7 from 2011 and SR 21-8, the 2021 interagency statement on model risk for BSA/AML systems.
The guidance is expected to be most relevant to banking organizations with over $30 billion in total assets. It sets no enforceable standards, though supervisory action remains available for violations of law or unsafe or unsound practices. Buying a model does not hand off the duty to validate it: SR 26-2 recognizes that banking organizations may not receive a vendor's underlying code, data or methodology, and says that "the principles of model risk management remain applicable" and that "an important element of model risk management is the validation of vendor products, either by internal or outside parties."
What the documentation should hold
Model documentation states the model's purpose, data, methodology and limitations. Around it, the institution needs a versioned policy, a log of every production change and a way for validators to get the inputs and outputs they test.
A BSA manager at a credit union said model validators had struggled to extract the data they needed, and asked how easily a validator can see the data feeding a model and what it outputs. The same BSA manager described the change history a validator should find: "they can see that the change was done, they can see that it was reviewed and approved."
A senior vice president (credit models and decisioning) at a large regional bank described the bank's own expectations: validate the model first, make sure it does not change inadvertently, log and test every change to production policies, and require an approval before anything launches. The same executive wanted policy changes to be easier while keeping an approval process, so business users are not handed the keys to production without guardrails.
What the testing should show
Before release, backtest a new rule, model or strategy on historical decisions and compare its performance with a baseline, usually the version it replaces. After release, check that its reasons stay stable, so the explanation an applicant or analyst receives still describes what the system does.
Credit teams will find validation, release, versioning and rollback covered in our guide to credit model deployment governance. For transaction monitoring models, see AML model validation after April 2026.
Where human accountability has to stay explicit
An explanation names a person as well as a reason. The same reviewer quoted above, a VP of control management in credit risk review (third line) at a card issuer, laid out the three lines of defense: the second line owns policies and procedures and does oversight, while the first line executes them, makes the decisions, owns the risk and self-tests. The third line tests both.
Every disposition should carry the name of the person who made it. An AVP and KYC manager at a community bank described how a missing name reads: "if I don't see a trail, the audit trail with one of my analysts', names on it, then I would question the following day. Was this skipped?"
Show an examiner where a person decided and what the system only recommended. Filing a suspicious activity report, escalating a case, overriding a decline and approving an exception are decisions a named person makes. A model score, an alert or an AI-drafted summary is an input to that decision, and the record should make the boundary visible.
For how review settings and reviewer rights work in practice, see what human-in-the-loop means in fraud and AML. Keeping one record across fraud, AML and credit is easier on fewer systems, a trade-off our guide to consolidating risk point solutions sets out.
How to reconstruct a decision for an examiner
To reconstruct a decision, pull the record for one decision and check that it holds each of these:
The inputs and their sources, with timestamps.
The data and feature versions used.
The policy version and the rule set in force that day.
The model version and its output.
The reasons produced, in the form the applicant, analyst or examiner would read them.
The action taken.
Any override, with who made it and why.
The approvals.
The notice or disposition sent.
The strategy and analytics lead (credit portfolio management) at a credit-card fintech mentioned earlier asked the question every reconstruction has to answer: "would I also have the rule set used on that day to make that output?" If the rule set was edited in place, the answer is no.
Test the record before an examiner does. Pick a decision from last quarter and rebuild it from the record alone, without asking the person who made it.
Where reconstruction usually breaks
The outcome is stored without the path, so a decline shows no variable, reason or policy version.
The policy or rule set was edited in place, and the version in force on the decision date no longer exists.
Model explanations are regenerated on request from a retrained model instead of stored at decision time.
Escalations carry a rationale and cleared alerts do not.
Overrides carry no name.
AI-written summaries state facts with no link back to the source records.
Where the decision and its rationale live in your stack, as opposed to where data is routed, is covered in our guide to risk orchestration versus risk decisioning.
Where Oscilar fits
Oscilar is a unified platform: real-time data fabric, device and behavior intelligence, rules, and case management in one stack, with agents native to it. For explainability, the relevant part is the record. Oscilar keeps a full audit trail per decision with stored reasoning, human-in-the-loop by design, and the platform is designed to fit the interagency model risk management guidance, which institutions operate under as an expectation rather than a certification.
The reconstruction test above applies to any platform, Oscilar included. Explainable AI in risk and compliance comes down to whether a past decision can be rebuilt from its record and defended to each audience that asks.
Frequently asked questions
What must be explainable at the individual-decision level?
For a single decision, an institution should be able to show the inputs and their sources, the policy, rule set and model versions in force, the rule that fired or the reasons the model produced, the action taken, and any override with who approved it. If that record cannot rebuild the decision later, the explanation will not hold up with an examiner.
What reasons must a lender give when a model contributes to a credit denial?
Under Regulation B, 12 CFR 1002.9(b)(2), a lender must give specific principal reasons for adverse action, and the official commentary requires those reasons to relate to the factors actually scored, with no principal factor left out. That holds whatever kind of model drove the denial, and FCRA key-factor disclosures do not replace the specific reasons.
Can a generative AI or agent output be explained the way a model score is?
No. A generated summary or an agent's action has no reason code behind it, so the explanation comes from the record of the work: the evidence used, the steps taken, the output and the person who approved it. Model risk teams can assess such systems by testing their outcomes against alternative methods.
What is the difference between explainable and interpretable AI?
An interpretable model is one a person can read directly, such as a scorecard or a small decision tree. Explainable AI is the broader term: it covers interpretable models and post-hoc methods, such as feature attributions and counterfactuals, that approximate why a complex model produced a given output.
How does the EU AI Act treat AI used for credit scoring?
The EU AI Act classes AI systems that evaluate the creditworthiness of natural persons or establish their credit score as high-risk under Annex III point 5(b), and it expressly excepts AI systems used to detect financial fraud. Regulation (EU) 2026/1744 moved the date those high-risk obligations apply to December 2, 2027.
Does SR 26-2 apply to generative AI and AI agents?
No. SR 26-2, issued with OCC Bulletin 2026-13 and FDIC FIL-15-2026 on April 17, 2026, places generative AI and agentic AI models outside its scope. The agencies said they plan to issue a request for information on model risk management focused on AI, including generative and agentic AI, and none had been issued as of October 28, 2026. Until then, institutions govern these systems under their own risk management.

Oscilar Team
The Oscilar Team is comprised of experts from many domains of risk operations. These articles express viewpoints and knowledge from a variety of sources and contributors across the organization.
DISCLAIMER
The content on this website is provided for informational purposes only and does not constitute legal, tax, financial, investment, or other professional advice. Any views or opinions expressed by quoted individuals, contributors, or third parties are solely their own and do not necessarily reflect the views of our organization.
Nothing herein should be construed as an endorsement, recommendation, or approval of any particular strategy, product, service, or viewpoint. Readers should consult their own qualified advisors before making any financial or investment decisions.
Oscilar makes no representations or warranties as to the accuracy, completeness, or timeliness of the information provided and disclaims any liability for any loss or damage arising from reliance on this content. This website may contain links to third-party websites, which Oscilar does not control or endorse.



