Adverse action notices, explained.

Why a lender that says no owes the applicant a letter

The life of one declined application, in five steps

1. The applicant appliesDana Whitlock asks for $24,000 online on a Monday night and uploads her pay stubs.
2. The application becomes completeTuesday at 10:04 am the stubs are verified. The bank has everything it normally needs, so the 30-day clock starts.
3. The decision is madeThe score is under the cutoff, the debt-to-income limit fails, a referral to underwriting ends in a decline on day 24.
4. The notice is written and sentReasons chosen, turned into words, checked, printed or e-delivered by day 30. This is where "Insufficient credit" is caught, or mailed.
5. The record is kept and testedThe letter, the reasons and the information used are kept 25 months; samples are re-checked; examiners ask for them.

Step 4 is the subject of this guide. The people, the bank and the figures in this example are fictional.

Early clocks, edited sample forms and withdrawn AI guidance: what surprises lenders

Adverse action, principal reasons, reason codes and key factors

Thirteen terms are defined in the one-page summary.

Which credit decisions count as adverse action, and which owe no notice

In shortA plain denial is obvious. Less obvious: a counteroffer the applicant does not take, a cut to an existing account that does not hit a whole class of accounts, and a refused increase all count. Actions on delinquent accounts, point-of-sale declines, and refusals because the lender does not offer the product or the law forbids it do not.

Regulation B defines adverse action in three parts.8 The first is "a refusal to grant credit in substantially the amount or on substantially the terms requested in an application", unless the lender makes a counteroffer and the applicant "uses or expressly accepts the credit offered". The second is "a termination of an account or an unfavorable change in the terms of an account that does not affect all or substantially all of a class of the creditor's accounts". The third is a refusal to increase the credit available after the applicant asked for an increase.

The same section lists what is excluded: a change the applicant agreed to; any action "taken in connection with inactivity, default, or delinquency as to that account"; a decline at the point of sale (with exceptions); a refusal because the law forbids the credit; and a refusal because the lender "does not offer the type of credit or credit plan requested".8 When an action falls in both lists, the exclusion wins.

The counteroffer: one letter or two

A counteroffer is an approval on different terms: a smaller amount, a higher rate, a shorter term, a co-signer. If the applicant takes it, there was no adverse action. If not, the lender owes a notice within 90 days of the counteroffer.2

The commentary lets the lender send one combined letter: "A creditor that gives the applicant a combined counteroffer and adverse action notice that complies with § 1002.9(a)(2) need not send a second adverse action notice if the applicant does not accept the counteroffer." Appendix C's Form C-4 is the sample.3,6 A combined letter closes the 90-day clock the day the offer goes out.

Line decreases and closed accounts for existing customers

A line decrease, a closed account or a worse rate on one customer's account, decided on that customer's behavior, is adverse action with its own 30-day clock.2 A change applied to a whole class of accounts, such as a repricing of every card in a portfolio, is not. In 2022 the CFPB issued an advisory opinion stressing that ECOA protects people after they receive credit; that opinion was on the May 2025 withdrawal list,7 but the regulation's definition and its 30-day rule for existing accounts are unchanged.2,8

ECOA and FCRA notice rules since 1976, and the guidance and enforcement shifts of 2025–2026

In shortThe core rules have barely moved in decades: ECOA's specific-reasons duty (1976), Regulation B's notice section, the FCRA notice (expanded in 2003 and 2010). What changed recently is the guidance and the enforcement around them: the CFPB withdrew its circulars on reasons from complex models in May 2025, two bank regulators stopped examining for disparate impact (a neutral-looking practice that falls harder on a protected group, whatever the intent), and in July 2026 a CFPB rule took that "effects test" out of Regulation B. The notice duties themselves are unchanged.

What the 2025 withdrawal changed, and what stayed in the regulation

The CFPB's May 12, 2025 notice withdrew 67 guidance documents, including both circulars, and said the withdrawal "is not necessarily final" and that it "will deprioritize enforcement" of the withdrawn guidance while it reviews it.7 The circulars were short. Their force came from quoting the regulation: Circular 2022-03 said a creditor's "lack of understanding of its own methods is therefore not a cognizable defense",11 and Circular 2023-03 said a creditor using the sample checklist must still give the actual reason, giving the example that a denial driven by an applicant's profession is poorly explained as "insufficient projected income".12

The sentences the circulars rested on are in the regulation and its commentary today.

A lender that relied on the circulars' withdrawal to relax its reason selection would be relying on nothing.

Fair lending exams, as of October 2026

The OCC told examiners on July 14, 2025 to stop examining for disparate impact; it still runs fair lending risk assessments and risk-based examinations and acts on evidence of disparate treatment, meaning an applicant treated differently because of a protected trait such as race, sex or age.13 The FDIC said on August 29, 2025 that it would evaluate discrimination "only through evidence of disparate treatment".14

The CFPB's own rule, effective July 21, 2026, says ECOA does not authorize disparate-impact liability, and its July 2026 exam procedures drop the interagency steps that reference it.15,17 A suit challenging the rule, National Fair Housing Alliance v. CFPB, is pending, with no injunction sought.18 None of this touches the notice section: the CFPB's July 2026 procedures still walk the examiner through every notice element.17

The CFPB itself is smaller: Congress cut its funding cap in July 2025, its supervisory events halved, and it brought one new public enforcement action in 18 months.19 State attorneys general enforce ECOA and their own consumer protection laws too, and Massachusetts did so against a student lender in July 2025.20

The 30-day and 90-day clocks: when each one starts and what stops it

In shortThirty days after a completed application, thirty days after action on an incomplete application or an existing account, ninety days after a counteroffer the applicant does not take. "Complete" means the lender has what it normally considers. Days are calendar days. A referral to manual review does not pause anything.

Dana's clock, day by day

The commentary sets the start: "Once a creditor has obtained all the information it normally considers in making a credit decision, the application is complete and the creditor has 30 days in which to notify the applicant of the credit decision."3 A lender that dates its clock from the decision, or from the day the file leaves underwriting, will send late notices on exactly the files that took longest, which are the files most likely to be declines.

Adverse action notice violations were among the Federal Reserve's top-cited compliance violations in 2023, and a common one was the late notice; at some banks, staff believed a phone call was enough.21

Monday, October 12, 2026 is Columbus Day, a Federal Reserve holiday.22 Regulation B counts calendar days, so Dana's due date does not move. A letter handed to a print vendor that day and mailed on the next postal day still goes inside the clock.

Incomplete applications: decide or ask, within 30 days

When an application is missing something the applicant can provide, the lender has two choices within 30 days. It can deny, giving "incomplete" as the reason, or it can send a notice of incompleteness that names the missing information, sets "a reasonable period of time" to supply it, and says the application will not be considered further without it (Form C-6 is the sample).2,3,6

If the missing item does not stop a decision, the lender must decide on what it has, and "incomplete application" "cannot be given as the reason for the denial".3 A lender may ask by phone first, but if the gap is still open, the written notice is owed inside the 30 days.3

Adverse action notice requirements: the eleven items a consumer decline letter carries when a credit report was used

In shortRegulation B asks for five: the action taken, the lender's name and address, the ECOA anti-discrimination statement, the lender's federal agency, and the specific reasons (or the right to ask for them). The FCRA adds six when a credit report was used: the bureau's contact, that it did not decide, the score with its range, date and key factors, and the rights to a free report and to dispute. Some states add their own statements.

The agency is whichever regulator supervises the lender for ECOA. Bramblewood Bank in the example has over $10 billion of assets, so its letters name the CFPB; a smaller bank names its prudential regulator. A charter change or growth past $10 billion changes the agency on every template at once.

The FCRA content applies whenever the decision rested "in whole or in part" on a consumer report, even if the score was a minor factor: "The credit score does not have to be the primary reason adverse action is taken," the Federal Reserve's compliance staff wrote after the 2022 exam cycle, when FCRA 615(a) failures were among its most cited.24 Information from an outside source other than a bureau triggers a different FCRA paragraph, 615(b), which Form C-1 also covers.6

Paper or electronic delivery of a decline notice

The notice must be in writing, though a lender that received 150 or fewer applications in the preceding calendar year may give it orally. A lender may deliver it electronically only with consent that meets the E-SIGN Act: the applicant "has affirmatively consented", after disclosures about paper copies and withdrawal, in a way that "reasonably demonstrates" access to the electronic form.25 Lenders keep the consent with the record, and send by mail when there is none.

What "specific principal reasons" means, rule by rule, with the words examiners accept

In shortThe reasons must be the factors that actually decided, named specifically. "Internal standards" and "did not meet our score" are banned by name. More than four reasons is "not likely to be helpful". With a scoring model, only scored factors may appear and no principal one may be left out. A failed knockout rule is always disclosed. The sample checklist must be edited to the lender's real factors.

Dana's reasons, worked through

Dana's file was scored, failed a policy rule, and then went to an underwriter. Under comment 9(b)(2)-7, a file that lands in a gray band (the score range where the system sends the file to a person) and is then declined by that person takes its reasons "from both components of the system".3 The policy rule, debt-to-income after the loan, is a hard limit, and comment -8 says such a factor must be disclosed. The model's own factors rank the rest.

So the four principal reasons are the failed debt-to-income rule, then R14, R03 and R22. R09 and R31 lowered the score too, but by less. The lender's selection method, which comment -5 lets it choose, decides where the line falls; what matters is that the method is written down and applied the same way every time.

Vague words and specific words

"Insufficient credit" does not appear on the sample checklist in Form C-1. Its nearest boxes are "Insufficient number of credit references provided", "Limited credit experience" and "No credit file", each a different factor.6

Adverse action reason codes: where the words on the letter drift from the model's decision

In shortA scoring model returns reason codes; a mapping table turns each code into words; a template prints them. Each of the three links can break. The most fragile is the table, because models are retrained and re-versioned and a code can change meaning. Explanation methods for machine-learning models approximate the model, and different methods can disagree about which factors drove the same decision.

Where the table breaks

Dana's bank retrained its personal loan model and released version 7 on September 1. Its documentation, filed on October 9 to the model inventory (the bank's register of every model it uses), defines code R14 as "length of credit history (age of oldest account, months)". In version 6, R14 meant too little credit information; that meaning moved to a new code, R38. The mapping table in use, version 6.3, still prints R14 as "Insufficient credit".

From September 1 to October 9, 946 letters went out with those words for a factor the new model measures differently, and 214 more sat in the week's queue.

This is the failure the commentary describes: a reason that does not "accurately describe the factors actually considered or scored".3 It is systematic, so every letter the same code and version produced carries it. It is also invisible on the page: "Insufficient credit" reads as a plausible reason to anyone who does not know what R14 now measures.

Choosing reasons from a machine-learning model

Comment 9(b)(2)-5, written for points-based scorecards, gives two ways to pick reasons: compare the applicant's score on each factor with the average score of applicants "whose total score was at or slightly above the minimum passing score", or with the average of all applicants, and take the factors that fall furthest below.3

Modern attribution methods, the techniques that estimate how much each input moved one applicant's score, do something similar: they estimate each input's contribution against a background population. The choice of that background is the part that maps onto the comment, and it should be documented with the model.

Attribution methods are estimates. A 2022 study of six explanation methods across four datasets found that leading methods "often disagree in terms of the explanations they output", and that practitioners "often employ ad hoc heuristics when resolving such disagreements".26 For a lender, disagreement means two defensible methods could print different top-four reasons for the same applicant. Comparing two methods' top factors on a sample is a cheap test, and a persistent disagreement rate is a finding to explain.

Models also have hundreds of engineered inputs that a letter cannot name. Grouping them into factors a person can understand, and deciding that "months since oldest trade opened" and "average age of accounts" both report as "length of credit history", is a modeling decision. It changes which reasons appear, so it belongs in the model documentation and its validation, with the mapping table versioned to match.

SR 26-2 and the models that pick decline reasons

On April 17, 2026 the Federal Reserve, OCC and FDIC replaced SR 11-7, the long-standing model risk guidance, with new guidance (SR 26-2, OCC Bulletin 2026-13, FDIC FIL-15-2026). A credit scoring model is a model under it. Generative AI models, and AI that acts on its own, "are not within the scope of this guidance", and the bank's own governance should guide controls for them.28,29 The notice rules apply whatever the technology: Regulation B has no exemption for complex models, which is why the 2022 circular's point outlived the circular.2,11

States are adding their own duties. Colorado's SB26-189, signed May 14, 2026, covers automated decisions about "financial or lending services": after an adverse outcome, a plain-language description of the automated system's role within 30 days, and human review and reconsideration on request, with the attorney general's rules due by January 1, 2027.16

Credit score disclosure and risk-based pricing notices: which letter goes to whom

In shortA declined applicant whose report was used gets the FCRA content inside the adverse action notice. An approved applicant who got worse terms than most because of a credit report gets a risk-based pricing notice, or a credit score disclosure in its place. A proper adverse action notice satisfies the risk-based pricing duty. Thin-file applicants, who may have no score, need their own handling.

The risk-based pricing rule covers credit granted "on material terms that are materially less favorable than the most favorable material terms available to a substantial proportion of consumers", based in whole or part on a consumer report.9 Lenders rarely compare each borrower one by one; Regulation V, the CFPB's rule under the FCRA, offers shortcuts.

Under the credit score proxy method, a lender finds the score at which about 40% of its borrowers score higher and 60% lower, and sends the notice to everyone below it. Under the tiered pricing method, with four or fewer tiers, everyone outside the top tier gets one.9

Applicants with thin files

Dana is new to credit: her oldest account is 11 months old. People like her are common. The CFPB's corrected estimate, published in June 2025, puts 2.7% of adults with no credit record at all in December 2020 (7.0 million people) and 87.5% with a scored record.30 The rest have a record too thin or stale to score.

For an unscored applicant the letter still owes the ECOA reasons ("No credit file" and "Limited credit experience" are both on the sample checklist), and the FCRA notice still applies if a report was used, even an empty one. A lender's templates need a path for "no score available" as well as for a score.

State law adds little here: in October 2025 the CFPB issued an interpretive rule saying the FCRA "generally preempts State laws that touch on broad areas of credit reporting", replacing a 2022 rule withdrawn in May 2025.31 State statements on the ECOA side of the letter, where a state requires them, are a separate matter for counsel.

Late letters, vague reasons and missing scores: what examiners and enforcers have found

In shortExaminers keep finding late notices, missing credit score content and vague reasons. Enforcement cases show the scale: one fintech lender sent more than 71,800 notices whose reasons were inaccurate. A wrong reason built into a table or template repeats on every letter it touches, which turns one error into a population finding.

What the Federal Reserve's examiners cite

The Federal Reserve's compliance staff publish the violations they cite most, and write AAN for adverse action notice. For 2023, adverse action notice violations were "among the Federal Reserve's top-cited compliance violations", and "a common violation was failing to provide an AAN within 30 days after receiving a completed application".21

Staff at some banks believed a phone call satisfied the rule. The root causes were "inadequate oversight by management and a lack of appropriate employee training" and "weaknesses in the monitoring and audit functions".21 For 2022, the FCRA list included "Not providing an AAN when taking adverse action based, in whole or in part, on information in a consumer report".24

Two adverse action enforcement cases

CFPB v. LendUp (2021). The CFPB alleged that the online lender "failed to provide adverse-action notices within the 30 days required by ECOA for over 7,400 loan applicants" and issued "over 71,800 adverse-action notices that failed to accurately describe the main reasons" for denials.27 The proposed settlement filed in December 2021 barred it from making new loans, with a $100,000 civil penalty "based on LendUp's demonstrated inability to pay".32

Massachusetts and Earnest (2025). On July 10, 2025 the Massachusetts attorney general announced a $2.5 million settlement with a student lender, alleging among other things "inaccurate and non-specific adverse action notices", a knockout rule that denied applicants without at least a green card, and an underwriting variable with disparate outcomes. The lender agreed to AI governance, model testing and an inventory of its models.20 A state enforcer acted where the federal agencies were stepping back.

The common failures, and why they happen

The inadvertent-error defense is narrow

Regulation B forgives "an inadvertent error", such as a "clerical mistake, calculation error, computer malfunction, and printing error", if the lender corrects it "as soon as possible". It adds: "An error of legal judgment is not an inadvertent error under the regulation."34 The CFPB's exam procedures require that the error "occurred despite the fact that the creditor maintains procedures reasonably adapted to avoid such errors".17 A mapping table never checked against a new model version is the absence of such a procedure.

Private suits can claim actual damages and punitive damages up to $10,000 per person, or in a class the lesser of $500,000 or 1% of the lender's net worth, within five years.35

How many credit applications end in a notice: denial rates and volumes in 2025 and 2026

In shortNo one publishes a national count of adverse action notices. Two series frame the scale: about one credit applicant in six was rejected in the twelve months to June 2026, by the New York Fed's survey, and mortgage lenders alone denied 2.1 million applications and closed 695,000 as incomplete in 2025. A lender's own number is in its origination system.

The denial share in mortgage lending is high. In 2025, 2,115,090 applications were denied and 6,827,891 originated, so about 24 of every 100 applications that ended in an origination or a denial were denials (our arithmetic).23 The St. Louis Fed's review of more than 30 million home purchase applications from 2018 to 2024 found denial rates rose from 12.2% to 15.7% during the 2022–2023 tightening, and traced the rise to debt-to-income limits as rates climbed.37 Dana's decline turned on the same kind of rule.

Volumes outside mortgage lending are larger and less visible. A bank with a card program, indirect auto and digital personal loans can send tens of thousands of notices a month; the fictional bank in this guide sends about 14,300. The numbers that matter for a program are its own: notices a month by product and action, the share that are counteroffers and line decreases, and the days from completed application to notice.

How lenders produce adverse action notices today: who touches each letter, and where the days go

In shortUnderwriting decides, credit risk owns the model and its reason codes, compliance owns the words and the testing, and a correspondence team or print vendor sends the letters. Most of the 30 days goes to the decision itself. The notice step is short, but it sits at the end of the clock and inherits every delay before it.

Underwriting, model owner, compliance and print: who touches a notice

A notice crosses four teams; the last has the least time

Codes into words, then release: where the post-decision days go

Origination templates, composition tools or reason tracing

Whatever the tooling, the lender owns the letter. The reasons are its reasons, chosen by its method, and an outsourced print run does not move the duty.

Dana Whitlock's decline letter: one stale reason caught, the mapping table corrected and 1,785 notices released on time

In shortThis is the example from the top of the page, inside the product. Six agents read the decision, trace each reason to the factor or rule behind it and check the letter against 14 rules. Elise Dunmore, the compliance analyst, rewrites the one reason that fails, proposes a corrected mapping table that the model owner and the compliance officer approve, and releases 1,785 notices, every one inside its clock.

1

She opens on the letter itself

On Monday morning, Elise's start screen is Dana's notice, set as the letter Dana will receive, with the 30-day clock above it: day 27 of 30, due Thursday, October 15. Beside each reason, a note traces it to its factor, Dana's own value and the points it cost. Beside the letter, 14 rules from Regulation B, the FCRA and the bank's house and state rules tick as the letter is checked, and stop at 12 of 14.

Reason 2 reads "Insufficient credit", underlined. Its note says the mapping table in use, version 6.3, prints those words for code R14, while model version 7 scored R14 as length of credit history: oldest account 11 months, minus 41 points.

2

The decision was read in under eight seconds

The decision reader took application PL-2609-48213: Personal Loan Score version 7 at 588 against a cutoff of 620, policy rule CP-4.2 failed at 46.2% debt-to-income against 43%, and the clock started at the completed application on September 15. The reason tracer ranked one policy rule and five model factors and picked four principal reasons: CP-4.2, R14, R03 and R22.

The reason-map checker compared each code's words with the model's documentation for version 7 and stopped on R14. The bureau disclosure checker read the credit report used: score 641, range 300 to 850, four key factors with inquiries among them.

3

One click rewrites the reason from the factor

Elise presses Rewrite from the factor. "Insufficient credit" is replaced by "Length of credit history: oldest account 11 months". The note turns green: accurate, specific, reading grade 5.2. The two Regulation B rules that failed, specific principal reasons and only factors actually scored, pass. The letter now meets 14 of 14.

The words come from the factor the model scored and Dana's own value, so the reason says what she can do about it: wait, as her history lengthens.

4

The table is fixed once, with two approvals

The same words sit on 213 more held letters and on 946 already sent between September 1 and October 9. Elise opens R14 in the reason-code map and proposes version 7.1, whose wording fills in each applicant's own months. Jun Park approves as model owner and Wendell Okafor as compliance officer. The 214 held notices are rewritten and checked again; none is left on hold.

For the 946 sent letters, a corrected notice is drafted for each. Whether they go is Wendell's decision.

5

She signs the batch

The day's batch holds 1,785 notices: the 1,571 that passed every check plus the 214 rewritten. Every reason is traced to its factor or rule, every notice is inside its clock, the earliest due on Tuesday, October 13. Elise signs; 1,142 go by e-delivery to applicants with E-SIGN consent on file and 643 go to the print vendor for first-class mail.

October 12 is Columbus Day. The letters handed to the print vendor that evening are mailed on the next postal day, Tuesday, October 13: on or before every due date in the batch, since a written notice counts as given when it is mailed.3

The solution's design targets: every notice out within 30 days, and four of four reasons on every letter traced to the factor or rule that decided. They are targets, to be measured on your own declines.

The notice clock queue, the reason-code map before the fix, Dana's score disclosure and the monthly accuracy test

In shortOne letter is the story. These screens show the rest: every notice on its clock, the reason-code map before the fix, what drove Dana's decision and the reason trace, the score disclosure on her letter, the monthly accuracy test, a look-up for any past decision, the clock report and the settings.

Select any screen to open it larger. The bank, applicants, people, models and figures are fictional; the credit bureau shown is invented.

The queue that morning

The compliance analyst releases every notice; a reason-code mapping change needs the model owner and the compliance officer

In shortEach agent has one job and limits it does not cross. None of them changes a decision, changes the mapping table or sends a letter. The analyst releases each notice and each batch; a mapping change needs the model owner and the compliance officer; corrected letters go only on the compliance officer's decision.

The analyst releases; two people change reason wording

The batch Elise signed

What notice software reads from decision records and model documentation, and the reason trace it keeps for every decline

In shortIt reads the decision records your systems already produce, the model documentation already in your inventory, the credit report used and your approved templates, and leaves behind a checked notice for every decline, counteroffer and line decrease, a trace of every reason, a clock report and a sampling record. Your systems of record stay where they are.

Origination, decision engine, model inventory and print vendor it links

For each, check whether it has an API, a file feed, or only screens, and who owns it.

Your decision engine, templates and print vendors keep their jobs

What a late or wrongly worded decline letter costs you: penalties, remediation and exam time

In shortStart from your own costs. Three published figures help: the damages the law allows, one enforcement case that alleged inaccurate reasons on 71,800 notices, and a compliance wage that prices the hours. Nobody publishes how often notices carry the wrong reason, so the number that matters most is the one your own sample produces.

Late notices, vague reasons and remediation: finding your own counts

A pattern finding, a fixable applicant, a delayed model release

Compliance minutes per notice and corrected letters avoided: the payoff from tracing every reason

In shortPut in your own numbers. The calculator starts from a published compliance wage and asks you for the rest: how many notices you send, the minutes each takes now and after, how often a reason turns out wrong, what a corrected letter costs, and what the tool costs a year. Every default marked illustrative is an assumption to replace.

Six returns from tracing reasons before release

Compliance minutes and corrected letters across your yearly notices

A lender sending 170,000 notices a year clears the first two lines on time alone. A credit union sending 3,000 does not, and should buy for accuracy and evidence, or not at all.

How ready your adverse action program is for reason tracing: clocks, mapping tables, overrides, testing and templates

In shortPick the answer closest to your team today. Eight questions score 0 to 2; the last is for reflection. The meter adds up your score as you go, and each question says why it matters and what a strong answer sounds like.

A notice program's score: fix the clock, pilot one model, or roll out

What to ask an adverse action notice vendor about traced reasons, wording control, model changes and evidence

In shortAsk every vendor the same questions, us included. The answers show whether each reason is traced to the decision or only to a template, who controls the words, how model changes are caught, and what evidence you keep.

Trace a past month of decline letters first: a ten-week notice pilot on one product and one scoring model

In shortMeasure first, then trace a past month of letters to find out what your real error rate is, then run in parallel on live declines, then decide with three numbers. None of these steps commits you to a vendor.

1. Pull a baseline: last month's declines for one product, with completed-application dates, decision dates and notice dates.
2. Re-trace 60 sent letters to the decision record and the code documentation for the model version that decided.
3. Gather the inputs: the mapping tables by version, the model documentation, templates by state, and an approval path to share decision data.
4. Run in parallel on live declines for five weeks, with your analysts releasing every batch from the current process.
5. Decide with three numbers: wrong-reason findings caught before release, days from completed application to notice, and staff minutes per batch.

Wrong reasons found, planted errors caught and days to notice

Adverse action notice timing, reasons, credit scores and AI models: the questions lenders search

In shortThe questions lenders and compliance teams search for most, each answered on its own.