--- title: What should an AI agent be allowed to do on a loan? | inflooens url: https://www.inflooens.com/guides/ai-agent-permissions-mortgage/ description: A working policy for AI on a loan file: what it may read, what it may prepare, what needs a person’s approval, what it must never decide, and how to stop it. --- Guide # What should an AI agent be allowed to do *on a loan?* Let it read the file within its user’s permissions, answer with sources, and prepare work for a person. Let it act only on tasks you have named, with a person’s approval where you require it, inside a daily limit, and with one switch that stops it. Never let it approve, deny or price a loan. That is a policy a lender can write down in a page, and it should exist before the tool does. [How Luna AI is governed](https://www.inflooens.com/ai-governance/) By Amit Ghole, President and CTO, inflooens · November 13, 2026 · General information, not legal advice. A ladder ## Five things an AI can do, in order of risk This is our way of thinking about it, not a regulatory standard. The point is that “AI on the loan” is not one decision. It is five, and you can say yes to some and not others. 1 Read ### Read the loan See the parts of the file its user is allowed to see, with identifying details removed or replaced before any outside model is called. Low risk when access is bound to the user. 2 Answer ### Answer and calculate Explain a guideline, calculate income, summarise a file. The risk is a wrong answer, so require sources a person can check. 3 Prepare ### Prepare work for a person Draft the message, fill in the form, line up the next step. A person reviews and sends. This is where most of the time saving is, at little added risk. 4 Act, with approval ### Act once a person approves Send the reminder, request the document, update a field. The lender decides which actions need a person’s approval, and each one leaves a record. 5 Act alone, within limits ### Complete a task end to end Carry out a whole routine task without a person in between. Only for tasks the lender has named, inside volume limits, with a way to stop everything at once. The lines ## Four things it should never do These hold at every step of the ladder, whatever the tool and whoever sells it. Approve, deny or price a loan A credit decision belongs to a person or to an established underwriting system, and the lender must be able to give the borrower its specific reasons. Act outside its user’s permissions If the person using it cannot open a record, neither can it. An AI with wider access than its user is a hole in your access controls. Contact a borrower who has opted out Consent and opt-outs apply to a message whoever, or whatever, wrote it. Leave no trace Every action needs a record of who asked, what was done, when, and on what basis. An action nobody can reconstruct is one nobody can defend. The policy ## Seven lines to write down before you switch it on A lender that can answer these has an AI policy. One that cannot is relying on the vendor’s defaults. **Which actions** A named list of what the AI may do. Anything not on the list is not allowed. **Which need approval** For each action, whether a person must approve it first. Start with approval on everything that reaches a borrower or changes loan data. **How many** A limit on how many actions it may take in a day, so a mistake cannot repeat a thousand times before anyone looks. **Who can stop it** A named person, and one control that halts every AI action across the company. Test it before you need it. **What is recorded** For each action: who asked, what was done, when, and the source it relied on, kept where your own team can read it. **What happens without it** Lending work must carry on if the AI is switched off or unavailable. **Who reviews** A regular look at what it did, what was approved or rejected, and whether the list above still fits. What the policy rests on ## Rules and frameworks to read alongside None of these tells a lender exactly how to configure an AI agent. Each sets an expectation the policy has to meet. Any organization [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) A voluntary framework for governing, mapping, measuring and managing AI risk. A useful structure for the policy above. Every creditor [Regulation B, 12 CFR 1002.9](https://www.ecfr.gov/current/title-12/chapter-X/part-1002/section-1002.9) A creditor that takes adverse action must be able to give the specific reasons, whatever technology helped reach the decision. Banks [Interagency Guidance on Third-Party Relationships: Risk Management (Federal Reserve, FDIC and OCC, June 2023)](https://www.federalreserve.gov/supervisionreg/srletters/SR2304.htm) A bank stays responsible for an activity it hands to a third party, including one carried out by a vendor’s AI. Sources checked on November 13, 2026. This is general information, not legal advice. Rules change, and your own counsel and compliance team decide how they apply to you. Questions to ask any vendor ## To check a tool can carry your policy A policy is only as good as the controls that enforce it. These six tell you whether the tool has them. 1. 1 Can we choose, action by action, which ones need a person’s approval? 2. 2 Is there a limit on how many actions it can take, and who sets it? 3. 3 Is there one control that stops all AI actions, and who holds it? 4. 4 Does it act within the permissions of the person using it? 5. 5 Where is the record of each action kept, and can we read it without asking you? 6. 6 What does it do today, and what is on your roadmap? Show us which is which. [The twelve questions for choosing an AI vendor](https://www.inflooens.com/guides/mortgage-ai-vendor-questions/) Where inflooens stands on the ladder ## Luna AI prepares. People decide. Stated in full on the AI governance page, with what is live and what is planned kept apart. **Steps 1 to 3** Live today Luna AI works within each user’s Salesforce permissions, with identifying data replaced or removed before any model call. It reads, calculates, drafts and recommends, and guideline answers cite their source and name the agency. **Step 4** Live today The lender’s administrators choose which actions need a person’s approval. A daily cap limits how many actions Luna AI can take, one org-wide emergency stop halts them, and every AI action leaves a receipt on the loan. **Step 5** Planned Luna AI completing the tasks a lender allows, end to end, is planned. It will run inside the approvals, daily caps and emergency stop that are live today. **The lines** Luna AI does not approve or deny a loan. Its approval-likelihood estimate is a planning aid, not an underwriting decision, and an AI outage does not stop lending work. [AI governance](https://www.inflooens.com/ai-governance/) Straight answers ## Questions about AI agents on a loan **What should an AI agent be allowed to do on a mortgage loan?** Read the file within its user’s permissions, answer questions with sources, and prepare work for a person to review. Beyond that, it should act only on a named list of tasks, with a person’s approval where the lender requires it, inside volume limits, and with a control that stops it. It should never approve, deny or price a loan. **Can an AI agent make a credit decision?** It should not. A creditor must be able to give a borrower the specific reasons for an adverse action, and credit decisions belong to people and to established underwriting systems. AI is suited to reading, calculating, drafting and recommending. **What is human-in-the-loop for mortgage AI?** It means a person approves an action before it happens. A practical policy does not put a person in front of every keystroke. It names which actions need approval, such as anything that reaches a borrower or changes loan data, and lets low-risk preparation run without it. **How do you stop an AI agent that is making mistakes?** With three controls set up in advance: a daily limit on how many actions it can take, approval on the actions that matter, and one switch that halts all AI actions across the company. Lending work must be able to continue with the AI off. **What can Luna AI do today?** Luna AI, the AI loan analyst in inflooens, reads the file, calculates, answers guideline questions with citations, drafts and recommends. The lender’s administrators choose which actions need a person’s approval, a daily cap limits its actions, and an org-wide emergency stop halts them. Luna AI completing allowed tasks end to end is planned, inside those same controls. More guides ## Keep reading - [Mortgage CRM vs POS vs LOS](https://www.inflooens.com/guides/mortgage-crm-vs-pos-vs-los/) - [How many tools a loan officer uses, and what they cost](https://www.inflooens.com/guides/mortgage-tech-stack-cost/) - [Mortgage pull-through and lead conversion benchmarks](https://www.inflooens.com/guides/mortgage-pull-through-benchmarks/) - [Two-way sync with your LOS: what it should mean](https://www.inflooens.com/guides/encompass-two-way-sync/) - [Loan officers on personal phones: what it costs](https://www.inflooens.com/guides/loan-officers-personal-phones/) - [How to track referral partners and who sent each loan](https://www.inflooens.com/guides/tracking-referral-partners/) - [Twelve questions to ask a mortgage AI vendor](https://www.inflooens.com/guides/mortgage-ai-vendor-questions/) - [Who should hold a lender’s loan data?](https://www.inflooens.com/guides/loan-data-ownership/) - [How to pilot new software while under contract elsewhere](https://www.inflooens.com/guides/pilot-mortgage-software-under-contract/) [All guides for lenders](https://www.inflooens.com/guides/) ## Bring us your AI policy. We will show you, control by control, how Luna AI fits inside it. [security@inflooens.com](mailto:security@inflooens.com) [Or book a session with our architects](https://outlook.office.com/book/inflooens@inflooens.com/)