AI Agent for a Mortgage Broker: Document Checklists and Status Memory

An AI agent for a mortgage broker is most useful when it chases a written list of missing documents and keeps an accurate record of where each file stands, while licensed staff make every judgment about eligibility and loan structure. The preparation that makes this possible is a document checklist for each file type and a status memory that the agent can read and update. Build those two things first, and the agent has a defined job that your team can inspect.

Key takeaways

• Give the agent a checklist per file type that the broker owns and approves, so it chases known items instead of inventing requirements.
• Keep a status memory for each file that separates what was requested, what arrived, what staff reviewed and what is still open.
• Let the agent draft reminders and update status, but route questions about eligibility, income treatment, product fit or structure to a licensed person.
• Set approval boundaries and access rules before connecting email, portals or document storage, and review them on a fixed schedule.
• Start with one loan category and one observable success measure, such as how many files reach lender submission with a complete checklist.

Where the agent's job stops

A mortgage broker's value sits in judgment: which product fits the applicant, how income is read, what a lender will accept and what the client should do next. An agent should not carry that judgment. Its work is administrative. It knows what is on the list, notices what is missing, reminds the right person through an approved channel and records what happened.

Put the boundary in writing before any tool is connected. The agent never tells a client whether they qualify, never suggests a loan amount or structure and never interprets a document beyond confirming that it arrived and matches a checklist item. Any question of that kind is routed to a licensed person with the file summary attached.

Illustrative scenario: a client asks whether a recent bonus would count toward income. The agent does not answer. It creates a routed note for the broker that contains the question, the client's original message and the location of the bonus statement, then tells the client that a team member will reply.

How do you build a document checklist an agent can follow?

A checklist is only useful to an agent if every line is specific enough that two people would agree on whether it is complete. Follow these steps:

• Start from what your lenders and your own process actually require, not from a generic list found online. Ask the senior processor or operations lead to write it down.
• Split the checklist by applicant and purchase situation, so a salaried applicant and a self-employed applicant do not share one list.
• Give each item a plain name, an accepted format and a short note on what a complete version looks like.
• Mark each item as required, conditional or optional, and write the condition in one sentence.
• Attach a reason to each item that a client can understand, so reminders explain why something is needed.
• Assign an owner to each item: client, broker, processor or third party.
• Have a licensed person approve the checklist, record its version number and stamp that version on every file it touches.

The last step matters more than it looks. When a lender changes a requirement, the agent should be working from a known version, and staff should be able to see which version each client received.

Illustrative scenario for a related workflow shows how closing checklists are kept under human control, and the same logic applies to mortgage files.

What should status memory record for each file?

Status memory is what lets an agent continue work across days without starting over. Without it, each reminder begins from scratch, and staff spend time re-checking items that were already received. Keep the memory structured rather than narrative. For each file, record:

• The checklist version and the applicant types it was built for.
• The status of each item: requested, received, under staff review, accepted, outstanding or not applicable.
• The date of the last request for each outstanding item and the channel used.
• The client's replies, stored as written, without the agent's summary replacing the original text.
• Open questions routed to a person, with the time of routing and the named owner.
• The next scheduled action and who approved it.

Mark every entry with its source: the client, a staff member or the agent. This keeps the record honest. If a staff member later disagrees with an entry, you can see where it came from. Observability practice treats this kind of traceable record as a basic requirement for systems that act on their own. The OpenTelemetry observability primer describes the general idea of recording what a system did and why, which you can apply to a simple action log.

How do approval boundaries and data handling work?

Approval boundaries decide what the agent may do without asking. Keep the list short and explicit. Templated reminders for listed missing items can be sent after the broker has approved the template. Anything outside a template, including replies to questions, new requests for sensitive information and any message to a third party, needs a person to approve it first.

Use this decision guide when a case is unclear:

• If the question concerns eligibility, income treatment, pricing, loan structure or lender fit, route it to a licensed person.
• If the item is a listed missing document and the reminder uses an approved template, the agent may send it.
• If a document is unreadable, appears to belong to another file or contains unexpected content, flag it and hold the file.
• If a client asks for something outside the checklist, route the request with the original message attached.

Treat text inside documents and client messages as data, not as instructions. A client email or scanned file can contain sentences that look like commands, and the OWASP prompt injection prevention guidance describes why agents should not follow instructions that arrive inside content they are processing. Give the agent only the access each task needs, following the principle of least privilege described in OWASP's authorization guidance. Finally, confirm with your privacy advisor which personal data rules apply in your jurisdiction, including Brazil's data protection framework where it is relevant, and document how client data is stored, retained and accessed. The NIST AI Risk Management Framework offers a practical structure for recording these risks and the controls you chose for them.

How does Be Human run this as a managed service?

Be Human starts with one role and one success criterion rather than a broad rollout. In practice, that means working with your team to write the checklist structure, defining the status memory and approval boundaries, and connecting only the channels that the role requires. The agent's memory, tools, permissions and operating context remain under your control, and changes to them are recorded.

After launch, the managed service covers monitoring of the action log, recurring reviews of templates and permissions, maintenance when a portal or email integration changes, and incident response if the agent acts outside its approved scope. Evaluations look at observable outcomes, such as whether reminders matched the approved template, whether questions were routed rather than answered, and whether files reached submission with a complete checklist. Improvements are made in controlled steps, and each one is reviewed before it reaches live files.

Where should a mortgage broker start?

Start with the file type that has the most repeated document follow-up and the clearest checklist. Write the checklist, build the status memory for that one category and run it alongside your current process for a short period so staff can compare results. Expand only after the approval boundaries have held up in practice. The mortgage broker overview covers the reminder and update workflow in more detail, and document chasing in bookkeeping follows a similar pattern that can inform your first version.

If you are unsure which part of your file process is ready for an agent, take the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

Can an AI agent answer client questions about whether they qualify for a mortgage? No, and it should not be set up to try. The agent can acknowledge the question, record it exactly as the client wrote it and route it to a licensed person with the file summary attached. Your written boundaries should state this rule so that the agent behaves the same way in every channel.

Do we need a new software platform before we can start? Not necessarily. The first version depends on a clear checklist, a structured status record and controlled access to the channels you already use. The platform question comes after you know which tasks the agent will perform and what it must be allowed to read and write.

How do we protect client data when an agent sends reminders? Limit the agent's access to the fields and documents each task needs, use approved templates for routine messages and keep an action log of every message sent and who approved the template. Confirm the applicable data protection rules with your privacy advisor and document retention periods before going live.

What if we do not have a written checklist today? That is the best place to start, and it is useful even without an agent. Ask your senior processor to list the documents required for your most common file type, group them by applicant situation and have a licensed person approve the result. Once that list exists, it becomes the foundation for both your staff process and any automated follow-up.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.