Skip to content

Agentic AI in Mortgage Origination.

The Mortgage Process Already Knows Where AI Belongs. The Industry Just Hasn't Looked How lenders and brokers can use existing process assets to build a structured, auditable, and paced path to agentic AI -- without reinventing what already works

The Australian mortgage industry is not short of AI ambition.

Every major lender has a generative AI programme. Brokers are piloting document tools. Fintechs are promising faster approvals. The conversation is loud.

What is missing is structure. Most of these programmes start from scratch -- new platforms, new workflows, new vendor relationships -- while ignoring the most valuable asset the industry already has: decades of BPMN process diagrams describing exactly how mortgage origination, assessment, and settlement actually work.

Those diagrams are not legacy debt. They are the most complete, compliance-tested description of mortgage process logic in existence. Every serviceability rule encoded. Every APRA CPS 234 requirement mapped. Every exception path documented and tested against real borrowers.

The question is not how to build new AI workflows. The question is how to systematically identify where AI belongs inside the workflows that already exist -- and adopt it at a pace that regulators, risk committees, and borrowers can trust.

What the Mortgage Process Looks Like in BPMN

A standard mortgage origination process in BPMN covers roughly six stages:

  1. Application receipt and initial validation
  2. Identity verification and KYC
  3. Income verification and serviceability assessment
  4. Security valuation
  5. Credit decision
  6. Conditional approval, settlement, and documentation

Each stage contains human tasks -- processors, assessors, and credit officers reviewing documents, making judgments, and advancing the application. At major lenders, thousands of these tasks execute every day. At brokerages, the same steps are done manually for every application.

The volume is high, the variability is high, and the regulatory exposure is significant. That combination makes mortgage origination one of the most valuable targets for AI augmentation in financial services -- and one of the highest-risk places to get it wrong.

BPMN captures the deterministic logic well. But it has no vocabulary for what comes next.

Where BPMN Falls Short for Mortgage AI

Income verification is probabilistic, not binary.

A human assessor reviews payslips, tax returns, and bank statements and makes a judgment about verified income. That judgment has nuance -- a casual employee with variable income is different from a salaried employee, which is different again from a self-employed borrower with complex business structures.

An AI model can do this work. But it returns a confidence score, not a binary answer. BPMN cannot express "if the model is above 90% confident on verified income, proceed to serviceability; if 70-90%, flag for assessor review; if below 70%, request additional documents". Every team that tries this hacks it into service task output variables and exclusive gateways -- obscuring the business logic and making the decision trail hard to audit.

Document intelligence output is not a system event.

BPMN handles document receipt as a message event -- the document arrived, start the next step. AI document intelligence is different. The model reads the document, extracts structured data, assesses completeness and authenticity, and returns a confidence-weighted output. That output can be wrong. It can be confidently wrong. BPMN has no standard error handler for "the AI was confident but the extracted income figure was implausible". There is no boundary event for hallucination.

APRA requires explainability. BPMN does not enforce it.

APRA CPS 234, the NCCP Act, and responsible lending obligations require that credit decisions be explainable and auditable. When AI is embedded inside a BPMN service task, the decision logic is invisible to the process diagram. The diagram shows "credit decision service task" -- it says nothing about which model was used, what version, what confidence it returned, or what the human reviewer was shown before they approved or declined.

This is a governance gap that regulators are increasingly focused on.

APMN: What BPMN Was Missing

APMN -- AI Process Model and Notation -- is an open extension of BPMN 2.0 that makes AI a first-class citizen of the process diagram. It adds the constructs mortgage processes need without changing anything about the existing BPMN foundation.

The constructs most relevant to mortgage:

ragTask -- retrieve financial context before making a decision. Credit bureau data, previous application history, property title records, ATO income data. The retrieval step is explicit in the diagram, not hidden inside a service call.

agentTask -- AI performs the assessment. Specifies the model, version, prompt context, and expected output schema. When Shaun the credit officer opens the application file next week and asks "what did the AI assess and why", the answer is in the process record.

confidenceGate -- route based on AI confidence score. Above threshold: straight-through processing. Mid-range: assessor review of AI output. Below threshold: full manual assessment. The thresholds are configurable by loan type, borrower segment, and regulatory requirement.

humanInLoopTask -- structured assessor review of AI output. The assessor sees the AI recommendation, the confidence score, the documents the AI referenced, and the key factors driving the output. They approve, override, or escalate. Every decision is recorded against the process instance.

escapeGate -- automatic fallback. If AI confidence drops below a minimum floor, if a model times out, or if the output fails structural validation, the escapeGate routes to the manual track automatically. The assessor queue. No borrower application stalls because an AI component failed.

modelVersionGate -- run a new model version on a percentage of applications alongside the current model, compare outputs, validate before full deployment. This is how you upgrade your income verification model without a big-bang release and the risk that comes with it.

The TwinTrack Principle for Mortgage

Mortgage origination APMN diagram

The core idea is foundational separation between AI infrastructure and deterministic infrastructure, joined by lightweight orchestration ramps.

Your existing mortgage process -- the one that passes APRA audits, that your assessors know, that your compliance team has signed off -- continues to run on the reliable track. It does not change. It does not stop.

AI runs on the innovation track in parallel. For each application that meets the routing criteria, both tracks process simultaneously. The ramps control which track delivers the outcome:

On-ramp criteria might be: application type is standard residential, borrower is PAYG, documentation is complete. These are the low-risk, high-volume applications where AI confidence is highest and regulatory exposure is lowest. Start here.

Off-ramp triggers: confidence below threshold, document anomaly detected, borrower segment outside training distribution, escapeGate fires. Any of these returns the application to the assessor queue on the reliable track. The assessor never knows the AI track ran -- they just see a normal application.

As evidence accumulates -- comparing AI outcomes to assessor decisions on the same applications -- the routing criteria widen. Self-employed borrowers are added when confidence on that segment is validated. Complex income structures are added later. The pace is determined by your risk committee and your regulatory dialogue, not by the AI vendor's deployment schedule.

This is paced AI adoption with a reliable fallback. The governance surface area is reduced because AI decisions are isolated from deterministic decisions -- each has its own audit trail, its own compliance framework, its own escalation path. Frictionless innovation in the AI track. Unconditional reliability in the deterministic track.

A Real Example: Income Verification

Income verification is where most mortgage AI programmes focus first -- high volume, time-consuming, document-intensive, and amenable to pattern recognition. It is also one of the highest regulatory risk points in the process.

In BPMN, income verification is typically a human task: assessor reviews payslips and tax returns, calculates verified income, records the figure in the origination system.

In APMN with TwinTrack:

ragTask retrieves all income documents from the document management system, plus the borrower's ATO tax return via CDR integration where available, and any previous income assessments from prior applications.

agentTask extracts structured income figures from each document, reconciles across sources, identifies discrepancies, flags anomalies, and calculates verified income per APRA serviceability guidelines. Output includes confidence score, extracted figures, and a structured explanation of the calculation.

confidenceGate routes: above 92% confidence on a standard PAYG borrower proceeds to serviceability calculation automatically; 75-92% presents the AI assessment to the assessor for review and confirmation; below 75% routes to full manual assessment with AI output available as a reference only.

humanInLoopTask (for mid-confidence cases) presents the assessor with the AI-calculated income, the documents it referenced, the specific figures it extracted, and a plain-language explanation of any discrepancies or flags. The assessor confirms, adjusts, or overrides. Every decision and its reason is recorded.

escapeGate catches document read failures, model timeouts, output validation failures, and cases where extracted figures are outside plausible range for the declared income. All return to the manual track.

The compliance record shows: which model assessed the income, what version, what confidence it returned, what documents it referenced, what the assessor saw, and what decision they made. This is the audit trail APRA expects.

Starting Point for Mortgage Leaders

If you are running a mortgage origination process today -- in Appian, IBM BPM, Pega, or any BPMN-compatible platform -- you have everything you need to start this analysis.

Upload your BPMN to TwinTrack. In minutes, you get:

  • A map of every human task in the process with an AI readiness score
  • APMN output showing the target state with AI-native nodes
  • Confidence scoring for each proposed AI replacement
  • A structured report you can take to your risk committee

This is not a vendor demo. It is your own process, analysed against your own data, producing a structured roadmap specific to your origination workflow. The output is Orkes Conductor JSON -- ready to run as soon as your team is ready to pilot.

You control the pace. You control the routing thresholds. You control when evidence is sufficient to widen the on-ramp.

The reliable track never stops. The AI track proves itself. The ramps between them are yours to configure.

Resources

APMN spec v0.1 (Apache 2.0, open source): apmn.kshetra.studio/spec/apmn-v0.1

APMN visual modeller (MIT, open source): apmn-modeler.kshetra.studio

TwinTrack -- BPMN to APMN compiler, free to try: bpmn2ai.kshetra.studio

Questions or contributions to the APMN standard: github.com/kshetra-studio/apmn

Dinesh Singh Panwar is the founder of Kshetra Studio and the creator of APMN and TwinTrack. Former Head of Technology and Chief Engineer, Westpac Group. Founder of askmybank.ai, an AI-native mortgage document intelligence platform for Australian brokers and lenders.