Compare documented processes or financing structures against a reader decision. Provider statements are attributed and illustrations are hypothetical; no paid placement, ranking or personalised recommendation.
Commercial status: There are no affiliate links in this edition. Providers did not pay for inclusion or approve this copy.
“AI underwriting” is a process label, not an explanation of your result. A provider may describe broad data categories, your notice may give specific reasons, and the law may set rules for that notice. Put those three forms of evidence in separate rows before drawing a conclusion.
This US consumer guide uses a lender disclosure as an example, not as a map of every model. The decision in front of you matters more than a claim that automated lending is faster, fairer or more inclusive.
Build an evidence ladder
Start a dated file with three headings: provider disclosure, application record and decision notice. Under the first, save what the provider says it may use. Under the second, keep the information you actually submitted. Under the third, copy the reasons and instructions from the notice.
Do not merge those rows. A category listed on a website does not prove it affected your application. A fact submitted in an application does not prove the model accepted it as accurate. A reason in a notice does not reveal every weight or decision rule.
Record the creditor named in the notice, the platform or technology provider, and any reporting company identified. They may have different roles. Direct a correction or question to the organisation responsible for the record rather than sending the same personal documents to every name on a screen.
Separate vendor claims from decision evidence
Upstart says its system may consider employment and income stability, education and professional background, spending and saving habits, and other financial behaviours. It also says neither Upstart nor its lending partners requires a minimum educational attainment. These are current provider statements; they do not disclose the weight of a factor or establish which item affected an individual result.
The same caution applies to speed and fairness language. The provider says its process can give an instant decision and makes claims about reducing bias. Those are vendor claims, not an audit finding and not proof about your application. Preserve the wording if it matters to your question, then return to the records you can verify.
Create four columns for any disputed input: category, source record, exact discrepancy and requested answer. If you do not know which source supplied a figure, write “unknown”. That is more useful than reverse-engineering a model rule from one result.
Read the notice for specific reasons
Regulation B section 1002.9 says an adverse-action notice must contain specific principal reasons when it provides the reasons. Alternatively, it may explain the right to request specific reasons within 30 days if the request is made within 60 days of notification. A statement that an applicant failed to meet internal standards or a qualifying score is not sufficient as the reason statement.
Those timelines depend on the notice and transaction covered by the rule. Read the document you received rather than applying a deadline copied from a different inquiry or business-credit situation. If the stage is unclear, ask whether the provider treated the interaction as an inquiry, completed application, incomplete application or another event.
A credit-score factor and a creditor's principal reason are not interchangeable. Keep each in its own row. Ask which record supports the stated reason and what procedure exists for correcting an error; do not ask only for a higher score or a different automated outcome.
Check the smallest verifiable discrepancy first
Compare names, account ownership, balances, dates, employment status and the period represented by income evidence. Describe one discrepancy precisely. “The algorithm is wrong” gives a recipient less to investigate than “the notice says income could not be verified, but the uploaded statement covers the requested dates and uses this employer name.”
If a credit-report entry is involved, obtain the relevant report through an official route and use the reporting company's dispute process. If the issue came from information supplied directly to the creditor, use its correction or reconsideration process. Save the original and corrected versions so the change is visible.
A correction does not guarantee approval, restore a quoted rate or prove discrimination. If you suspect unlawful treatment, preserve the notice, provider disclosures and dated records, then use a qualified adviser or appropriate complaint route. Do not diagnose a prohibited practice from the word “AI” alone.
Keep current rules and agency priorities distinct
The CFPB's report published on 1 October 2026 describes its 2025 fair-lending programme and later policy actions. It records the agency's current account of its priorities; it is not a complete statement of every federal or state protection and does not replace the operative rule for a notice.
Use the report to understand the agency's position, the regulation to check notice requirements, and the creditor's documents to understand your transaction. An older policy announcement, a vendor fairness claim and a current legal provision answer different questions. Label the source and date beside each conclusion.
Analyse an approval as an offer, not a verdict
If the application was approved, record the amount received, origination fee, APR, payment schedule and total repayment. Check whether a fee is deducted from proceeds or added to the obligation. Test the payment against a weaker-income month. An automated approval says nothing about where the payment money will come from.
Upstart says checking a rate does not affect a credit score, while proceeding with a full application may involve a hard inquiry. Treat that as one provider's sequence. Confirm the wording for the service you are using before moving from an initial check to an application.
Finish your evidence ladder with one question that a document or named organisation can answer. Correct a specific record, request the stated reasons, compare the actual offer or pause borrowing. The useful outcome is a verified next step, not a theory about a hidden model.
From comparison to calculation
Check the terms in the offer you actually receive.
Credit Over Debt’s Offer X-Ray compares the amount financed, fees, payment schedule and modeled cost using the figures you enter.
Open the free toolkitPrimary sources and product pages
- Consumer Financial Protection Bureau: Regulation B, section 1002.9: Notifications (checked 2026-10-09)
- Upstart: AI-powered lending: how it works and what to expect (checked 2026-10-09)
- Consumer Financial Protection Bureau: Fair Lending Report for 2025 (checked 2026-10-09)
Product details last checked Oct 9, 2026; next scheduled review Nov 4, 2026. This US-focused guide is educational information, not individualized financial, legal or tax advice. Product terms, offers and eligibility can change.