DEV Community: Dinesh Singh Panwar

Agentic AI in Mortgage Origination

Dinesh Singh Panwar — Mon, 22 Jun 2026 08:23:20 +0000

The Mortgage Process Already Knows Where AI Belongs

A technical guide to converting existing BPMN mortgage workflows to APMN for agentic AI

The Australian mortgage industry has significant AI investment and limited AI production deployment.

Most programmes are piloting document tools or building chatbots alongside existing processes. Few are asking the more productive question: what is already encoded in our existing process diagrams, and where does AI fit inside that?

Every lender running Appian, IBM BPM, Pega, or a comparable platform has BPMN diagrams for mortgage origination. Serviceability rules encoded. APRA CPS 234 requirements mapped. Exception paths tested against real borrowers over years.

Those diagrams are the starting point, not an obstacle to work around.

Where BPMN Falls Short for Mortgage AI

Income verification is probabilistic.

A human assessor reviews payslips, tax returns, and bank statements and returns a verified income figure. An AI model does the same work but returns a confidence-weighted output. BPMN cannot express "proceed if confidence above 90%, flag for assessor review if 70-90%, request additional documents if below 70%". Every implementation hacks this into service task variables and exclusive gateways -- obscuring the business logic and making the decision trail difficult to audit.

Document intelligence output is not a system event.

BPMN handles document receipt as a message event. AI document intelligence is different -- the model reads, extracts, validates, and returns a structured output that can be wrong. There is no BPMN boundary event for "the AI extracted an income figure that is implausible for the declared employment type". Catching this requires custom code that sits outside the process diagram and outside the audit trail.

APRA requires explainability. BPMN does not enforce it.

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" and says nothing about which model ran, what version, what confidence it returned, or what the assessor was shown before approving. This is a governance gap regulators are focused on.

What APMN Adds

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. Fully backwards compatible with existing BPMN tools.

Key constructs for mortgage:

ragTask -- retrieve financial context before assessment. Credit bureau data, ATO income data via CDR, previous application history, property records. The retrieval step is explicit in the diagram and in the audit trail.

agentTask -- AI performs the assessment. Specifies model, version, prompt context, and expected output schema. When a compliance officer asks what the AI assessed and why, the answer is in the process record.

confidenceGate -- route based on confidence score. Above threshold: straight-through processing. Mid-range: assessor review of AI output. Below threshold: full manual assessment. 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 referenced, and the key factors driving the output. Every decision is recorded against the process instance.

escapeGate -- automatic fallback. If confidence drops below minimum, if the model times out, or if output fails structural validation, the escapeGate routes to the manual assessor queue. No 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.

Full spec at apmn.kshetra.studio/spec/apmn-v0.1

The TwinTrack Architecture for Mortgage

Foundational separation between AI infrastructure and deterministic infrastructure, joined by lightweight orchestration ramps.

Your existing mortgage process -- the one that passes APRA audits, that assessors know, that compliance has signed off -- runs on the reliable track. It does not change.

AI runs on the innovation track in parallel. On-ramps divert selected applications based on routing criteria: standard residential, PAYG borrower, complete documentation. Low-risk, high-volume, high AI confidence. Start here.

Off-ramp triggers: confidence below threshold, document anomaly, borrower segment outside training distribution, escapeGate fires. Any of these returns the application to the assessor queue on the reliable track.

As evidence accumulates -- comparing AI outcomes to assessor decisions on the same applications -- the routing criteria widen at a pace your risk committee controls. Self-employed borrowers added when confidence on that segment is validated. Complex income structures added later.

The governance surface is reduced because AI decisions are isolated from deterministic decisions. Each has its own audit trail, its own escalation path. Frictionless innovation on the AI track. Unconditional reliability on the deterministic track.

A Worked Example: Income Verification

Standard BPMN: assessor reviews payslips and tax returns, calculates verified income, records in origination system. One human task. High volume. Time-consuming.

In APMN with TwinTrack:

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

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

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

humanInLoopTask for mid-confidence cases presents the assessor with AI-calculated income, documents referenced, figures extracted, and plain-language explanation of discrepancies. Assessor confirms, adjusts, or overrides. Decision and reason recorded.

escapeGate catches document read failures, model timeouts, output validation failures, figures outside plausible range. All return to manual track.

The compliance record shows: model, version, confidence, documents referenced, what the assessor saw, what decision they made. The audit trail APRA expects.

Getting Started

Upload your existing mortgage origination BPMN to TwinTrack. It identifies every human task where AI can act, scores conversion confidence, and generates APMN output plus deployment-ready Orkes Conductor JSON.

You control the confidence thresholds, routing criteria, and pace of adoption.

APMN spec v0.1 (Apache 2.0): apmn.kshetra.studio/spec/apmn-v0.1
APMN visual modeller (MIT): apmn-modeler.kshetra.studio
TwinTrack, free to try: bpmn2ai.kshetra.studio

Full article with APRA governance framework: apmn.kshetra.studio

Dinesh Singh Panwar, founder of Kshetra Studio. Creator of APMN and TwinTrack. Former Head of Technology, Westpac Group. Founder of askmybank.ai, AI-native mortgage document intelligence for Australian brokers and lenders.

Agentic AI in Clinical Processes

Dinesh Singh Panwar — Mon, 22 Jun 2026 07:54:34 +0000

Your Clinical Processes Already Know Where AI Belongs

A technical guide to converting existing BPMN healthcare workflows to APMN for agentic AI

Most healthcare AI programmes start by asking the wrong question.

"Which AI platform should we adopt?" "What should we pilot?" "Let's build something new and see what sticks."

Meanwhile, sitting in a process repository, is a BPMN diagram for patient preadmission. Another for prior authorisation. Another for clinical triage. Each one represents years of institutional knowledge -- every compliance requirement codified, every exception path tested against real patients.

That knowledge gets ignored. Teams reinvent from scratch. Pilots stay as pilots.

There is a better starting point.

What BPMN Cannot Express

BPMN 2.0 handles deterministic healthcare processes well. Sequential care steps, parallel workstreams, human tasks, service calls, exception handling -- it covers all of this.

The problem is that AI agents are not deterministic. Three specific gaps matter for health:

Probabilistic outputs. A radiologist returns a diagnosis. An AI model returns a probability distribution. BPMN has no way to say "proceed if confidence exceeds 90%, escalate to clinician if below". Teams hack this into service task variables and exclusive gateways. It works but makes the decision trail hard to audit.

AI-specific failure modes. BPMN handles system failures well. It handles AI failures badly. There is no boundary event for hallucination, no error handler for model drift, no standard catch for "the model was confident but the output was clinically implausible".

Transition states. A hospital cannot flip a switch and hand clinical decisions to AI. BPMN has no concept of "run AI alongside human for six months, compare outcomes, then decide". A task is either human or automated. Nothing in between.

What APMN Adds

APMN -- AI Process Model and Notation -- is an open extension of BPMN 2.0 that addresses these gaps while remaining fully backwards compatible with existing BPMN tools and diagrams.

It uses BPMN 2.0's official extension mechanism. Your existing diagrams remain valid.

The key constructs for healthcare:

agentTask -- an AI model performs clinical reasoning, document analysis, or administrative processing. The model, version, prompt context, and output schema are explicit in the diagram.

ragTask -- retrieve clinical context before reasoning. Patient history, previous imaging, medication records, clinical guidelines. The retrieval step is a first-class process node, not hidden inside a service call.

confidenceGate -- route based on AI confidence score. Thresholds are configurable per process and per risk appetite. High confidence proceeds. Mid-range flags for clinician review. Low confidence falls back to the human task.

humanInLoopTask -- structured clinician review of AI output before the process continues. Designed specifically for AI oversight -- the clinician sees the AI recommendation, the confidence score, the source documents, and the key factors driving the output.

escapeGate -- automatic safety net. If AI confidence drops below a minimum floor, if the model times out, or if output fails structural validation, the escapeGate catches it and routes to the deterministic fallback. The reliable process that was running before AI was introduced.

modelVersionGate -- run two model versions in parallel on a percentage of cases and compare outcomes before committing to an upgrade across the full patient population.

Full spec at apmn.kshetra.studio/spec/apmn-v0.1

The TwinTrack Architecture

The architectural principle behind APMN adoption is foundational separation between AI infrastructure and deterministic infrastructure, joined by lightweight orchestration ramps.

Your existing clinical process -- the one that passes audits, that clinicians trust -- runs on the reliable track. It does not change.

AI runs on the innovation track in parallel. On-ramps divert selected cases to the AI track based on routing criteria you define. Off-ramps return cases to the reliable track if confidence drops or an escapeGate fires.

The separation means a failure in the AI track cannot propagate to the reliable track. Governance of AI decisions is isolated from governance of deterministic decisions. Clinicians trust the reliable track unconditionally -- which is the prerequisite for AI adoption in clinical settings.

As evidence accumulates, the routing criteria widen at a pace determined by your risk appetite.

A Worked Example: Patient Preadmission

Standard BPMN: receive referral, verify insurance, review medical history, assess clinical priority, schedule, confirm. Steps 2, 3, 4 are human tasks.

In APMN with TwinTrack:

Verify insurance: ragTask retrieves policy documents. agentTask verifies eligibility. confidenceGate routes on confidence score. escapeGate catches failures and returns to human task.

Review medical history: ragTask retrieves EHR records and clinical guidelines. agentTask summarises and flags risk factors. escapeGate validates output before it reaches the clinician. humanInLoopTask presents the AI summary alongside source documents for clinician sign-off.

Assess clinical priority: agentTask applies urgency scoring. confidenceGate routes high-confidence scores to expedited scheduling, borderline scores to consultant review, low confidence to the human task on the reliable track.

Steps 1, 5, 6 are unchanged. The business logic and compliance requirements are unchanged. Three bottlenecks now have AI handling preparation work, with human oversight calibrated to confidence level and clinical risk.

Full worked example with before/after diagrams at apmn.kshetra.studio/examples/patient_preadmission

Getting Started

Upload your existing BPMN to TwinTrack. It identifies every human task where AI can act, scores conversion confidence, and generates APMN output plus deployment-ready Orkes Conductor JSON.

You control the confidence thresholds, the routing criteria, and the pace of adoption.

APMN spec v0.1 (Apache 2.0): apmn.kshetra.studio/spec/apmn-v0.1
APMN visual modeller (MIT): apmn-modeler.kshetra.studio
TwinTrack, free to try: bpmn2ai.kshetra.studio

Full article with regulatory context and governance framework: apmn.kshetra.studio

Dinesh Singh Panwar, founder of Kshetra Studio. Creator of APMN and TwinTrack.Former Head of Technology, Westpac Group.