What problem does it solve?
When a lender declines an application, cuts a credit line or offers worse terms, it often has to tell the customer why: in the US, Regulation B requires the principal reasons for a decline or other adverse action, and EU law gives people a right to an explanation of automated credit assessments. In practice a notice can be little more than a list of checked reasons from a sample form ("limited credit experience", "excessive obligations in relation to income") that says little to the customer and may not reflect what really drove the decision. Complex machine learning models make this harder: a model can weigh many features, and the reasons printed must still be the principal ones and must be accurate.
The rules are explicit. In the US, Regulation B requires a statement of reasons that is specific and indicates the principal reasons, and says that checking the closest reason on a sample form is not enough when it is not the factor actually used. The CFPB's 2022 circular stating that this applies equally to complex algorithms was withdrawn in May 2025 together with many other CFPB guidance documents, but the adverse action notice requirements in 12 CFR 1002.9 and their official interpretation are unchanged. In the EU, the Court of Justice ruled in February 2025 (Dun & Bradstreet Austria) that a person subject to an automated credit assessment is entitled to an explanation of the procedure and principles actually applied, and the EU AI Act adds a right to an explanation for decisions based on high risk systems such as credit scoring. Vague or inaccurate explanations leave customers without a clear next step and the lender exposed on compliance.
How does it work?
- Take the decision record. The assistant receives the decision, the reason codes and their ranking from the decision engine, and the relevant customer and product data.
- Retrieve approved wording. For each reason code it retrieves the approved plain language description and, where allowed, what the customer could do about it.
- Draft within strict limits. A generative model composes the notice and a short internal rationale using only those reasons, in the customer's language and at an agreed reading level.
- Check automatically. A second pass verifies that every reason in the draft maps to a code from the decision, that no code is missing, and that no prohibited content appears.
- Review and send. A reviewer approves the draft, or approved templates go out automatically once quality is proven. Follow up questions go to an assistant limited to the same reasons, with a route to a person.
- Audience
- Employee facing
- Autonomy
- Copilot
- Adoption
- Emerging
- Channels
- Email, Mobile app, Web chat, Agent desktop
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
No public deployment has disclosed a measurable outcome yet.
Value drivers: Compliance quality, Customer experience, Employee productivity.
Indicative value
A consumer lender issuing 100,000 adverse action notices a year
USD 66,667 to USD 400,000
Operations effort saved on decline follow ups per year
How this is calculated
Formula: notices * followUpShare * minutesSaved / 60 * hourlyCost. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Adverse action notices per year notices, notices per year | 100,000 | 100,000 | The reference lender. |
| Share of notices that lead to a question, complaint or reconsideration request followUpShare, fraction of notices | 0.05 | 0.1 | Editorial assumption. Replace with your own contact and complaint data. |
| Minutes saved per follow up case with a clear explanation and a prepared rationale minutesSaved, minutes per case | 20 | 40 | Editorial assumption for looking up the decision and writing a response. Replace with your own time study. |
| Fully loaded cost of a lending operations hour hourlyCost, USD per hour | 40 | 60 | Editorial assumption. |
What it leaves out: Counts handling effort only. The larger value, lower regulatory and fair lending risk and fewer repeat failed applications, is real but hard to price and is left out.
Who already uses it?
2 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Discover Financial Services
United States · Banking · 2024
Discover Financial Services patented a framework that generates the adverse action reason codes a lender must give a declined credit applicant directly from a machine learning model. The system groups correlated input variables, scores each group with partial dependence plots and Shapley Additive Explanations, ranks the groups, and turns the top ranked groups into the reason codes sent to the applicant. The United States Patent and Trademark Office granted the patent in July 2024 on an application Discover filed in May 2020. Discover Financial Services merged into Capital One Financial Corporation in May 2025 (the patent assignment was recorded in July 2025), which is why current patent databases list Capital One as the assignee. No source discloses an error rate, approval volume or other outcome for the system.
No outcome disclosed.
Wells Fargo
United States · Banking · 2022
Wells Fargo Bank patented a computer based credit evaluation system that pairs a machine learning credit risk model with an adverse action methodology: when the model denies an applicant, the system compares the applicant's characteristic values against anchor values taken from a top scoring population, calculates a replacement score for each characteristic, and ranks the characteristics to identify the principal adverse action factors for the denial. The United States Patent and Trademark Office granted the patent in October 2022 on an application Wells Fargo filed in October 2019. Separately, the trade publication Risk.net reported in August 2021 that a team of Wells Fargo researchers had begun deploying an explainability technique for its deep learning credit models, and a paper by six Wells Fargo model risk researchers proposed a related Shapley decomposition method for explaining adverse credit decisions. No outcome metric or notice volume is disclosed by any source.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Ranked reason codes for every adverse decision from the decision engine
- An approved library of plain language descriptions per reason code, per language
- Notice templates and regulatory content requirements per product and market
- Samples of past notices and complaints for testing
Systems to integrate
- Decision engine or loan origination system
- Document generation and correspondence system
- Complaint and case management
- Customer channels for delivery and follow up questions
Complexity: Medium
Drafting text is easy. The hard parts are reliable reason codes from the model, a reason library that compliance owns, and automated checks that make it impossible to state a reason the model did not produce.
- 1
Fix the reason codes first
Confirm with model risk that the model produces specific, ranked principal reasons that are accurate for each decision. The assistant cannot repair weak reasons.
- 2
Build the reason library
For every code, compliance approves a customer description, an internal description and, where appropriate, a note on what the customer could change. Keep owners and review dates.
- 3
Constrain the generation
Pass only the decision's codes and approved descriptions to the model and forbid any other reason. Use templates for the legally required parts of the notice.
- 4
Verify every draft
Add an automated check that maps each stated reason back to a code and blocks the draft on any mismatch, omission or prohibited term.
- 5
Review, measure, then automate
Start with human review of every draft, measure the error rate, and allow automatic sending per product only when the error rate is proven near zero.
Guardrails
- The draft may only contain reasons present in the decision record
- Every principal reason in the decision record must appear in the notice
- No protected characteristic or proxy may appear as a reason
- Legally required elements come from templates, not from generation
- Every sent notice is stored with the decision record and the codes it was built from
KPIs to instrument
- Share of drafts that pass the automated reason check first time
- Error rate found in human review and in sampling after launch
- Time to issue a notice after the decision
- Complaints and reconsideration requests that mention unclear reasons
Human in the loop
Compliance owns the reason library and approves templates. Reviewers approve drafts until the measured error rate allows automatic sending, and a human answers any dispute or reconsideration request.
Common failure modes
- Invented or softened reasons
- A fluent model adds a plausible reason or blurs the real one. Block any reason that does not map to a code.
- Reasons that are accurate but useless
- Codes that describe model internals mean nothing to customers. Invest in the reason library, not only the model.
- Drift between model and library
- A model update adds or renames features and the library falls behind. Tie library review to model change control.
What are the risks and rules?
EU AI Act
Depends on design
The drafting assistant does not assess creditworthiness, so on its own it is not the Annex III point 5(b) credit scoring system. It helps the lender meet the Article 86 right of affected people to a clear and meaningful explanation of decisions based on such a high risk system. If it is built into the scoring system it shares that system's high risk obligations; as a separate drafting tool its tier depends on its design and on how its output is reviewed. The follow up chat assistant must tell customers they are dealing with an AI system (Article 50).
Rules that apply
Guidance
- Regulation B, 12 CFR 1002.9: notifications, with official interpretation (Consumer Financial Protection Bureau, North America). The statement of reasons must be specific and indicate the principal reasons, must describe the factors actually considered or scored, and may not leave out a factor that was a principal reason.
- Regulation B, Appendix C: sample notification forms (Consumer Financial Protection Bureau, North America). Checking the closest reason on a sample form does not satisfy the notice requirement when it is not the factor actually used.
- Consumer Financial Protection Circular 2022-03: adverse action notification requirements for credit decisions based on complex algorithms (Consumer Financial Protection Bureau, North America). Stated that the specific reasons requirement applies equally to complex or opaque models. Withdrawn by the CFPB in May 2025 together with Circular 2023-03 on sample forms; the notice requirements in 12 CFR 1002.9 it interprets are unchanged.
- Interpretive rules, policy statements, and advisory opinions; withdrawal (Consumer Financial Protection Bureau, North America). The May 2025 Federal Register notice that withdrew many CFPB guidance documents, including Circulars 2022-03 and 2023-03 on adverse action notices.
- Innovation spotlight: providing adverse action notices when using AI/ML models (Consumer Financial Protection Bureau, North America). An earlier CFPB blog on how the adverse action rules apply to machine learning models. The page now carries a notice that it describes the requirements incompletely.
- Case C-203/22, Dun & Bradstreet Austria: automated credit assessment and the right to an explanation (Court of Justice of the European Union, Europe). Under the GDPR the controller must describe the procedure and principles actually applied so the person can understand which personal data were used and how; handing over the algorithm is not enough.
- Article 86, right to explanation of individual decision making (European Union, Europe). People affected by a deployer's decision based on an Annex III high risk system (except critical infrastructure, point 2) that has legal or similarly significant adverse effects on them can ask for a clear and meaningful explanation.
- SR 26-2: Revised Guidance on Model Risk Management (Board of Governors of the Federal Reserve System, OCC and FDIC, North America). Interagency model risk management guidance of April 2026 that supersedes and replaces SR 11-7. It applies to the credit model whose reason codes the notice is built from, and is expected to be most relevant to banking organizations with over $30 billion in total assets.
Controls to put in place
- Reason library under compliance ownership with version history
- Automated reason to code reconciliation on every draft
- Sampling of sent notices against decision records, reported to compliance
- Change control that links model releases to reason library review
Frequently asked questions
- Can a generative model write adverse action notices?
- It can draft them, but only from the reason codes the credit model produced and with an automated check that nothing was added or left out. The legally required parts should come from templates.
- Does using AI change the duty to explain credit decisions?
- No. In the US, Regulation B requires specific reasons that describe the factors actually considered or scored, whatever the technology. The CFPB withdrew its 2022 circular on complex algorithms in May 2025, but the notice requirements in 12 CFR 1002.9 and their official interpretation are unchanged. In the EU, the Court of Justice has ruled that a person subject to an automated credit assessment must be told the procedure and principles actually applied, in an intelligible way, and the EU AI Act adds a right to an explanation for decisions based on high risk systems such as credit scoring.
- Which named lenders have this documented?
- Wells Fargo Bank holds a patent, granted in October 2022, for an adverse action methodology that ranks the characteristics of a machine learning credit risk model to identify the principal reasons behind a denial. Trade press reported in August 2021 that a Wells Fargo team had begun deploying an explainability technique for its deep learning credit models. Discover Financial Services holds a related patent, granted in July 2024, that turns Shapley based explanations of a credit model into adverse action reason codes. Both are the lender's own patent filings rather than a disclosed error rate or notice volume, so they show that the methodology is real and built, not that it runs at full production scale.
How to cite this page
Blits.ai AI Use Case Library, "AI drafted explanations for credit declines and adverse actions", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/adverse-action-explanations. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published