Application started by an AI agent

How can an AI agent start a mortgage application, with the sensitive data captured on a secure screen and the file landing in the loan team's queue verified?

Let the agent start it and let the process run it. The AI agent (the lender's own or Callvu AI) calls Callvu over MCP with the borrower's intent; the evaluators assess the request; the borrower is moved to a process-controlled screen for identity, income details, consents and the Social Security number; income and insurance are verified through the document and verification processes; and the application lands in the loan team's queue with everything checked. The agent never holds the sensitive data.

In scopeComplex: multi-party handoffsCallvu moves: prefill, chase and verify, write back

Talk to an expert Try it in the sandbox

Last updated November 7, 2026.

The problem today

The assistant can explain the mortgage process; it cannot take the application without taking the borrower's Social Security number into a chat.

Today, without Callvu

A lender's AI assistant talks to a prospective borrower about rates and products. When the borrower wants to apply, the assistant hands them a link to a form, where they start over, or takes the application in the chat, where their income, their Social Security number and their consents become messages in a model's context.

The application is a regulated process: consents before the credit pull, the demographic questions in the prescribed form, the six pieces that start the TRID clock, data protection under GLBA. None of that belongs in a chat, and an assistant with write access to the LOS is a model risk finding.

The right split is clean: the agent starts and reports; the process captures, verifies, enforces and writes.

How it runs

From the requestto the system of record.

A borrower tells the lender's AI agent they want to apply. The agent calls Callvu's MCP tool to start a mortgage application with what the borrower said. The time is recorded.

Borrower: Talks to the agent, then completes identity, consents, income and assets on a secure screen.
AI agent: Starts the process over MCP and relays status; holds no account data, no identifiers, no consents.
Loan officer: Receives the verified application in the queue and takes it from there.
Processor: Works the file with the verifications already done.
Callvu: Evaluates the request, runs the secure capture, prefills, verifies, enforces consents and HMDA, writes to the LOS.

The agent starts the process
The MCP call carries the borrower's stated intent and any non-sensitive context (the product discussed, the purchase price range). The six evaluators assess the request and the Completion Orchestrator decides: execute, execute with safeguards, escalate, or block. The decision is recorded.
Prefill and secure capture
The borrower receives a link to a process-controlled screen (or is handed into the lender's app). An existing customer is identified and prefilled from the core. Identity, the Social Security number, income, assets, the consents for the credit report, E-SIGN and privacy, and the HMDA questions in the prescribed form are captured there. The agent sees only that the step is in progress or done.
Chase and verify
Income is verified through the employment and asset verification processes, and the homeowners insurance agent is asked for the binder, all through links; the TRID application date is detected when the six pieces exist and the Loan Estimate process opens.
Write back
The application, its consents, the verifications and the application date are written to the LOS, and the file lands in the loan team's queue classified by work type and due date. The agent tells the borrower the reference and what happens next.

Where it ends: The application in the LOS came from a conversation with an AI agent and holds the same consents, verifications and clocks as one taken by a loan officer, with nothing sensitive in the chat. The Decision Trace shows why the AI-started request was allowed and under what safeguards; the Audit Trace shows every step inside and outside the chat.

Runtime enforcement

What is enforcedwhile the work runs.

The regulations preloaded in Callvu Studio and the bank's own policies, applied as steps that run, not lines in a binder. A step that breaks a rule does not run, so a completed run was compliant.

Regulations

ECOA (Reg B, 12 CFR 1002.5, 1002.9 and 1002.13): The application may not request prohibited information, the monitoring information is collected in the prescribed form, and the 30-day notice runs from the completed application; the secure screen enforces the form and the process records the date.
GLBA (Reg P, 12 CFR part 1016, and the Safeguards Guidelines): The borrower's nonpublic personal information is collected and stored under the lender's information security program; a chat session is not a secure channel for it, and the privacy notice is delivered on the process-controlled screen.
TRID application definition (Reg Z 12 CFR 1026.2(a)(3)): The application date is detected when the six pieces exist on the secure screen, which starts the Loan Estimate clock.
Model risk management (SR 11-7 and OCC 2011-12) and FCRA (15 USC 1681b): An AI agent that starts a regulated process is under the lender's model risk program; the evaluator assessment and the Decision Trace evidence each AI-started action, and the credit pull follows the authorization captured on the secure screen.

Company policies

The agent may start the application and read status; it may not capture identity, consents or sensitive data, nor write to the LOS.
Every AI-started request is evaluated and the orchestrator's decision recorded.
The application's consents, HMDA questions and application date are captured and detected on the secure screen exactly as for a loan officer's application.
The LOS is the system of record; the chat holds only the conversation.

Controls

MCP tool with a defined scope.
Evaluator assessment and orchestrator decision.
Hand-off to a process-controlled screen for sensitive data and consents.
Prefill from the core for existing customers.
Verification processes started from the application.
Application date detection and LOS write.

Clocks and deadlines

Loan Estimate (TRID): Three business days from the application date detected on the secure screen.
Notice of action (Reg B): 30 days from the completed application.

Documents and checks

The agent's request and the evaluator assessment with the decision
The application as captured on the secure screen with consents and HMDA data
The verifications started and their results
The LOS write and the queue assignment

In the Decision Trace

Why the AI-started request was allowed and the safeguards applied.
The consents recorded before any pull.
The application date detection.

In the Audit Trace

The agent's call and its time.
Each step on the secure screen.
The verifications and the LOS write.

The handoff

Between the front doorand the back office.

The front door

The AI agent over MCP; sensitive data, consents and signatures on process-controlled screens.

The back office

The loan team receives a verified application in its queue, classified by work type and due date, with no chat transcript to read.

The Decision Desk

A request the Completion Orchestrator escalates goes to a person in the Decision Desk before anything runs; the loan's later decisions run as in the standard application.

Proof

In scope.Proof at scale, not one process at a time.

The kind of process Callvu runs in production for customers in this industry. Ask for a reference on this exact process. The results below are the industry's published proof; each claim code links to its record on agents.callvu.com.

Bank Leumi runs its home lending workflows on Callvu.

MTG-01

Community Bank of Mississippi runs TRID loans on Callvu.

MTG-02

More than 1,000 processes are live across Callvu customers.

Across Callvu customers

Any AI agent, including the customer's own, can start a Callvu process over MCP, and the same evaluation and enforcement apply. Sensitive data never goes into an AI chat; must-not-skip steps stay on a process-controlled screen.

ENF-04 · Self-asserted · verified 2026-11-02

Six evaluator agents assess every invocation (completion risk, compliance evidence, policy interpretation, data trust, audit readiness, exception patterns), and one deterministic orchestrator decides: execute, execute with safeguards, escalate to a human, or block.

ENF-02 · Self-asserted · verified 2026-11-02

Every run produces a Decision Trace (why each AI action was allowed) and every workflow execution an Audit Trace (the chronological governance record).

ENF-03 · Self-asserted · verified 2026-11-02

Questions

What buyers ask next,answered plainly.

Does the agent see the borrower's Social Security number?

No. Identity, the identifier, consents and income details are captured on a process-controlled screen outside the chat. The agent receives a status and a reference.

Which agents can start it?

The lender's own or Callvu AI, over MCP, under the same evaluation and enforcement.

Is the application different from one a loan officer takes?

No. The same consents, HMDA questions, application date detection and LOS write, with the verifications started the same way.

Related processes

Application intake (URLA / 1003)Credit report authorization Verification of employment Try it in the sandbox All home lending processes

Bring this process, or the one that costs you most.

We build it against your policies and your systems, train your team and stay with you through launch. You are live in weeks.