Selling lending software to a lender takes seven stages. The chief credit officer or head of lending owns the deal. Loan losses show long after a loan funds, so a back-test on past applications carries the core proof. Launch comes on one loan product.
What are the seven stages?
The order follows the stages of an enterprise B2B sales cycle. Each stage ends at one exit point.
- Discovery. The champion confirms the approval rate, time to decision, loss rate and the end date of the current contract.
- Technical scoping. Both sides agree which past applications the back-test uses.
- Back-test, shadow run, then pilot. The chief credit officer signs the results.
- MOU or LOI. The economic buyer signs.
- Commercials. Both sides agree the pricing unit, volume, term and migration fee, subject to contract.
Then come two more stages:
- Legal, risk and compliance. Model risk, fair lending, vendor risk and security finish their review. The lender signs.
- Go live. One loan product goes live. A review names the next product.
For account data, see an open banking deal with a lender or PSP. For lending through platforms, see an embedded finance deal.
What starts a lending software project now?
A contract end date, a credit model refresh or a rule with a date. Three rules are live on 8 November 2026.
Section 1071 in the US. The CFPB issued a narrower small business lending data rule on 1 May 2026. All covered lenders must comply by 1 January 2028. A lender is covered at 1,000 covered transactions in each of two consecutive years (Mayer Brown, May 2026).
Regulation B in the US. The CFPB amendments took effect on 21 July 2026 (Federal Register, April 2026). They state that the Equal Credit Opportunity Act gives no disparate impact liability. Disparate treatment stays prohibited, including proxies for a protected characteristic (Venable, May 2026).
Two lawsuits ask courts to set the rule aside. No court had paused it by 3 November 2026 (RATA Associates, November 2026). Non-housing lenders can now test new variables, so a model refresh is a good reason to call.
The EU AI Act. AI systems that assess the creditworthiness of individuals are high risk. Fraud detection is carved out (Annex III, point 5(b)). The rules for these systems apply from 2 December 2027 (European Commission, July 2026).
What must the back-test and shadow run prove?
The back-test proves credit quality on old applications. The shadow run and pilot prove the system works today. Agree every threshold before the first run.
- Back-test. Report approval rate at the same loss rate, or loss rate at the same approval rate.
- Shadow run. Score live applications without action. Explain each mismatch.
- Live pilot. Report time to decision, pull-through and the manual exception rate.
- Decline reasons. Report the share of declines with specific principal reasons. In the US, a decline with no usable reason blocks launch.
Pull-through is the share of approved applications that fund.
What goes into the MOU?
The MOU names the loan product, the back-test vintages, the metrics with thresholds, the timeline and the signers. Add the credit committee date, because many lenders need that committee to change credit policy. Name the model validator, the start date and the owner of the migration for loans in progress.
How is lending software priced?
Vendors charge per application, per funded loan, per seat, or a platform fee with a minimum volume. Many add a migration fee. A price per application makes the lender pay for declines. A price per funded loan ties your revenue to the lender's revenue.
What do model risk, fair lending and vendor risk check?
Model status. US guidance of 17 April 2026 replaced SR 11-7. It defines a model as a complex quantitative method based on statistical, economic or financial theory. Fixed rule-based software is outside it. Generative and agentic AI are also outside it (Federal Reserve, FDIC and OCC, April 2026).
A workflow product with fixed rules can argue it is out of scope. A scoring product cannot. The guidance is most relevant above $30 billion in total assets (Federal Reserve, April 2026). It can reach smaller lenders with significant model exposure.
Vendor model validation. A bank may not receive your code. It must still understand the soundness, design, development data and performance of the model, and monitor outcomes (same guidance). Bring a validation file on those four items.
Adverse action. The lender must act within 30 days of a completed application. Denial reasons must be specific (12 CFR 1002.9). Map every model output to a reason the lender can send.
Fair lending. Disparate impact claims remain under the Fair Housing Act and some state laws (Venable, May 2026). Expect proxy tests on your variables.
Third-party risk. On 11 September 2026 the agencies proposed new guidance that scales oversight to vendor risk (OCC, September 2026). Comments close on 16 November 2026. The 2023 guidance applies until then (Skadden, October 2026).
What does launch look like?
Launch one product through one channel, with the old system as a fallback. Review after the first full month of funded loans against the MOU thresholds.
Keep monitoring as repayment data arrives. A second product needs its own credit policy and a validation update.
To start discovery with lenders that have a dated reason to act, get five named buyers, free.