AI Credit Decisioning and Risk-Based Pricing — after lender hard-criteria screening
Scope: personal loans through Upstart.com only. Controlled object: the model-dependent portion of an individual personal-loan outcome after lender hard-criteria screening, including borrower qualification and risk-based pricing/terms. The lending partner remains the lender and originator. Hard-policy qualification: "The majority of credit denials on the platform are due to the lending partners' hard criteria from their underwriting policies." Lender hard-rule denials are outside this capability; Upstart's AI does not own all approval or denial authority. Chain: lender establishes and approves its lending program and hard underwriting criteria → lender hard criteria applied before Upstart's underwriting model → applications failing those criteria may be declined without model involvement → surviving application assessed by Upstart's underwriting model for default and prepayment probabilities → model-derived assessment materially affects borrower qualification → underwriting-model output enters a pricing engine together with lender-defined pricing/economic requirements (such as target return objectives and maximum allowable APR limits) → individual risk-based pricing/terms and model-dependent credit outcome → electronic loan offer where applicable → applicant decides whether to accept → lending partner remains lender/originator. The exact mechanism producing a model-dependent decline is not publicly established.
Recorded characteristics
- Function
- Each lender "sets and approves its own underwriting policy that establishes certain 'hard' requirements or criteria, which may include minimum credit scores, minimum and maximum loan amounts, and maximum debt to income ratio. Upstart applies the lenders' hard criteria prior to engaging its underwriting models." "Borrower applications that meet a lender's hard credit criteria are then assessed by Upstart's underwriting model for default and prepayment probabilities, after which a pricing engine takes into account the underwriting model output in addition to pricing requirements set by lenders, such as target return objectives and maximum allowable APR limits." Pricing boundary: underwriting-model output + lender pricing/economic requirements → pricing engine → individual risk-based pricing/terms. Not claimed: that the model alone determines APR; the exact pricing formula; the exact transformation from default/prepayment probability to APR. Approval boundary: Upstart reports that "underwriting model improvements drove higher approval rates in smaller dollar categories of loans" — model changes affect qualification/approval populations; the exact decline mechanism after hard criteria is not publicly established. Actors — Lending partner: establishes and approves underwriting policy; establishes hard criteria; establishes documented economic/pricing requirements; retains "complete authority and control over their lending programs"; remains lender/originator. Upstart: provides model-derived default/prepayment assessment; materially affects qualification on the qualifying path; feeds model output into pricing; operates the documented marketplace process. Applicant: decides whether to accept the resulting offer.
- Data access
- Loan application data collected through the marketplace (Upstart.com application) and the lender's hard criteria inputs (e.g. credit score, loan amount, debt-to-income). The complete production feature set is not publicly established; the Credit Decision API's "2,500+ variables" statement is deliberately not used.
- Actions
- Can take actions
- External actions
- Yes
- Human confirmation
- Not required
- Permission basis
- Not established
- Administrative control
- Lending partners "set and approve" their own underwriting policy and hard criteria; "Because our lending partners have complete authority and control over their lending programs, they predetermine many aspects of their loan offerings, including interest rate and loan size ranges, target returns for various risk profiles, minimum credit score, maximum debt-to-income ratio, fee structures, and disclosures." Human confirmation is recorded as not_required only for the documented qualifying fully automated personal-loan path: "91% of loans on our platform were fully automated", defined as loans "originated end-to-end with no human involvement required by the Company", which for personal loans means "from initial rate request to final funding". This does not establish that lender personnel are necessarily absent; the remaining 9% is not characterised by Upstart.
- Default state
- Not established
- Availability
- Live, current marketplace per the 2025 Form 10-K. This record covers personal loans through Upstart.com only.
- Licensing
- Lending partners pay platform fees for usage of the AI lending marketplace and referral fees for borrowers referred from Upstart.com. Pricing to lenders not publicly established.
- External model or provider
- Upstart's own proprietary AI/underwriting models. No third-party model provider is established.
- Limitations and uncertainty
- Hard-policy qualification: "The majority of credit denials on the platform are due to the lending partners' hard criteria from their underwriting policies." Explainability: the 10-K describes ECOA's requirement that creditors provide notices of adverse action when credit is denied or offered on less favourable terms; this is not evidence of complete model explainability. Upstart warns that regulators "may seek explanations for specific model outputs or outcomes, which may be difficult, costly or impractical given the complexity of our models." Exclusions: Credit Decision API; auto loans; HELOCs; home lending; identity verification; income verification; fraud detection; loan-stacking detection; verification exceptions; consumer acquisition/targeting; servicing; lender hard-rule denials. No Credit Decision API evidence is used. Not publicly established: (1) exact production model architecture; (2) complete production feature set; (3) feature weights; (4) exact model-dependent decline mechanism after lender hard criteria; (5) precise relationship between model probability and qualification threshold; (6) exact pricing-engine formula; (7) exact transformation from risk estimate to APR/terms; (8) lender-specific decision/pricing differences; (9) model-version traceability for an individual decision; (10) complete individual model explanation; (11) complete model-to-decision audit trail; (12) decision-log structure and retention; (13) override mechanisms; (14) lender exceptions; (15) exact model ownership of loan amount; (16) exact permission/authentication mechanism; (17) model update/retraining mechanics; (18) whether lender personnel participate in an otherwise Upstart-automated marketplace path; (19) nature of the remaining non-fully-automated population; (20) documented default state of the automated path per lender; (21) whether adverse-action notices identify model-derived reasons for a specific decision. Monitoring: no monitor. Upstart's website currently returns a Cloudflare challenge (403) to the Registry fetch mechanism and is unsuitable for the normal Registry monitor. The SEC-hosted 10-K is retained as evidence but is not a monitor; SEC monitoring requiring a declared-contact retrieval method has not been approved as part of this record.
Evidence
- Upstart Holdings, Inc. — 2025 Annual Report (Form 10-K), filed 2026
Supports: Function · Actions · Human confirmation · Admin controls · External actions · Limitations · General · Primary source
"Upstart applies the lenders' hard criteria prior to engaging its underwriting models." "Borrower applications that meet a lender's hard credit criteria are then assessed by Upstart's underwriting model for default and prepayment probabilities".
"a pricing engine takes into account the underwriting model output in addition to pricing requirements set by lenders, such as target return objectives and maximum allowable APR limits."
"91% of loans on our platform were fully automated"; defined as loans "originated end-to-end with no human involvement required by the Company"; for personal loans "from initial rate request to final funding".
Each lender "sets and approves its own underwriting policy"; lenders "have complete authority and control over their lending programs" and "predetermine many aspects of their loan offerings".
Marketplace includes "the delivery of electronic loan offers and associated documentation"; "underwriting model improvements drove higher approval rates in smaller dollar categories of loans".
"The majority of credit denials on the platform are due to the lending partners' hard criteria"; explaining specific model outputs "may be difficult, costly or impractical given the complexity of our models."
"proprietary, cloud-based, AI lending marketplace"; "Consumers primarily access Upstart-powered loans through Upstart.com".