Guide

Judging a mortgage AI vendor: twelve questions to ask.

Ask any AI vendor that will touch a loan file five things, in writing: what borrower data reaches the AI model, who can see and stop what the AI does, whether its work can be traced to a source, where a person decides, and what happens to your data when you leave. A vendor that cannot answer each one specifically is not ready for a regulated lender.

How Luna AI is governed

By Amit Ghole, President and CTO, inflooens · November 13, 2026 · This guide is general information, not legal advice.

The short version

Five areas, in the order a risk team asks

The twelve questions below sit under these five headings. If time is short, get a written answer to one question in each.

  • 1 · Data

    What reaches the model

    Which borrower details are removed or replaced before the AI model is called, where that happens, and whether anything is kept or used for training.

  • 2 · Control

    Who can see, act and stop

    Whether the AI is limited to what its user may see, which actions need a person, and how you halt it.

  • 3 · Evidence

    What it did, and why

    A record on each loan of what the AI did, and answers that point to their source.

  • 4 · Decisions

    Where a person decides

    Whether it makes or shapes credit decisions, how it is tested, and whether you can explain an outcome to a borrower.

  • 5 · Exit

    Whose data it is

    Where your loan data lives, who at the vendor can reach it, and what happens to it when the contract ends.

Why it falls on you

The lender answers for the vendor’s AI

Buying AI from a vendor does not move the responsibility. These are the rules and frameworks a lender’s own examiners and auditors will point to.

Sources checked on November 13, 2026. Rules change, and your own counsel and compliance team decide how they apply to you.

The questions

Twelve questions, with a warning sign and a good answer for each

Send them as written. The value is in comparing what comes back: a specific answer and a vague one look very different side by side.

Data

1. Exactly which borrower data reaches the AI model?

A warning sign“Your data is secure” or “we anonymize it”, with no list of fields.

A good answerA written list: what is removed, what is replaced with a placeholder, and what is still sent, such as loan amounts and ratios.

2. Where is that data removed or replaced?

A warning signAt the vendor or at the model provider, after the full record has already been sent.

A good answerInside your own environment, before the request leaves it. Real values are put back only there.

3. Is our data used to train models, and what does each model provider keep?

A warning signAn opt-out you have to find and switch on, or “industry standard” in place of a list.

A good answerNo training on your data, stated in the contract. A named list of the providers involved and what each retains, in writing.

Control

4. Can the AI see more than the person using it?

A warning signIt runs under one account that can read every loan.

A good answerNo. It works within that user’s own permissions and cannot read or change what the user cannot.

5. Which actions can it take without a person, and who decides that?

A warning signThe vendor decides, or it is all or nothing.

A good answerYour administrators choose which actions need a person’s approval, and there is a limit on how many actions it can take.

6. How do we stop it, and what happens to lending when it stops?

A warning sign“Contact support.” Or the workflow depends on the AI, so an outage stops work.

A good answerOne switch that you control halts the AI’s actions. Your people keep working loans without it.

Evidence

7. Can we see what it did on a loan, and why?

A warning signLogs exist at the vendor and are available on request.

A good answerYes, on the loan itself: who asked, what the AI did, when, and the source it used. You hold that record.

8. Where do its answers come from?

A warning signConfident answers with nothing to check them against.

A good answerGuideline answers cite the section and name the agency, and it says when it does not have a reliable source.

Decisions

9. Does it make or influence credit decisions?

A warning signA score that feeds a decision and that nobody can explain.

A good answerA plain statement of what is an aid and what is a decision. People decide, and you can give a borrower the specific reasons for an outcome.

10. How is it tested, and what happens when the model behind it changes?

A warning signTesting described only in general terms, and model changes you learn about afterwards.

A good answerDocumented testing before each release, including for errors and for uneven outcomes across groups, and notice to you when the model changes.

Exit

11. Where does our loan data live, and what happens when we leave?

A warning signIn the vendor’s shared database, with an export “available”.

A good answerIn an environment you own. At the end of the contract it is returned or deleted as the contract says, and the vendor’s AI keeps no copy.

12. Who at the vendor can reach our data?

A warning signStanding administrator access for the vendor’s staff.

A good answerNobody by default. Access is granted by you, for a stated purpose such as support, and it ends.

How inflooens answers

Our own answers, so you can hold us to them

inflooens makes a Mortgage OS with an AI loan analyst, Luna AI, so these questions apply to us. Each answer below is stated in full on the pages linked underneath.

Data
Before a request leaves the lender’s own org, Luna AI replaces borrower names, phone numbers and email addresses with tokens, and removes Social Security numbers, dates of birth, account numbers and street addresses. The model still receives the loan’s financial details. Real values are restored only inside the lender’s org. inflooens does not use lender or borrower data to train AI models, and the list of model providers and what each retains is given in writing during a vendor review.
Control
Luna AI works within the Salesforce permissions of the person using it. The lender’s administrators choose which actions need a person’s approval, a daily cap limits how many actions Luna AI can take, and one org-wide emergency stop halts them. An AI outage does not stop lending work.
Evidence
Every AI action leaves a receipt on the loan: who asked, what Luna AI did, when, and the source behind it. Guideline answers cite their source and name the agency.
Decisions
Luna AI prepares and people decide. Its approval-likelihood estimate is a planning aid for the loan team. It is not an underwriting decision and is not used to approve, deny or price a loan.
Exit
Each lender runs in its own Salesforce org, provisioned for it and owned by it. inflooens holds no standing access to that org, and Luna AI keeps no copy of loan data.
AI governance Security and data handling Privacy policy

Straight answers

Questions lenders ask about AI vendors

What should a lender ask a mortgage AI vendor?
Ask five things, in writing. What borrower data reaches the AI model. Who can see, approve and stop what the AI does. Whether its actions and answers can be traced to a source. Whether it makes or shapes credit decisions. And where your data lives and what happens to it when you leave.
Who is responsible when a vendor’s AI makes a mistake on a loan?
The lender. Banking regulators expect a bank to manage the risk of an activity it hands to a third party, the FTC Safeguards Rule requires non-bank lenders to oversee their service providers, and Regulation B requires any creditor to give the specific reasons for an adverse action. None of these shifts the duty to the vendor.
Is it safe to send borrower data to an AI model?
It depends on what is sent and where it is removed. The safer pattern is to remove or replace identifying details, such as names, Social Security numbers, dates of birth and account numbers, inside the lender’s own environment before the model is called, and to send only what the task needs.
Can an AI model make a mortgage credit decision?
A creditor must be able to give a borrower the specific reasons for an adverse action, whatever technology was used. In practice lenders use AI to read, calculate, draft and recommend, and keep the decision with a person or with established underwriting systems.
What is a warning sign in an AI vendor’s answers?
Reassurance in place of specifics. “Your data is secure”, “industry standard” and “available on request” are not answers. A vendor ready for a regulated lender gives a list of fields, a named set of providers, and controls that you hold yourself.

More guides

Keep reading

All guides for lenders

Send us the twelve questions.

We answer them in writing, along with your own questionnaire.

security@inflooens.com Or book a session with our architects