PTPTN Loan Operations AI PlatformTechnical proposal
Prepared for Perbadanan Tabung Pendidikan Tinggi Nasional (PTPTN) Prepared by Kersani Global Sdn Bhd Document KG-PTPTN-TP-01, version 1.0 Issued October 2026

One timeline for every borrower.

An internal, AI-assisted platform that brings PTPTN’s application, disbursement, repayment and case records into a single auditable view, so officers decide faster, collections teams act earlier, and every action can be evidenced on request.

What an officer would see in one screen instead of six Synthetic borrower · loan PJ-2019-204118 · degree, Universiti Putra Malaysia
Aug 2019Application receivedLoan application system
Sep 2019Documents verified, 0.97 confidenceAI extraction + officer
Oct 2019First disbursementDisbursement records
Jul 2023Graduated, grace period startsInstitution feed
Aug 2024Repayment begins, salary deductionRepayment ledger
Mar 2026Employer remittance stopsSalary deduction file
Apr 2026Early warning raised, score 71Risk model + rules
Apr 2026SMS + call, restructure offeredCollections worklist
May 2026Back on scheduleLedger, bank confirmed
System eventRisk signalAI-assisted action, approved by an officer
Section 1

Executive summary

PTPTN already holds the data it needs to know where every loan stands. The difficulty is that this data lives in separate systems, and stitching it together for a single decision still takes an officer’s time, judgement and memory.

We propose an internal platform, used only by PTPTN staff, that sits over PTPTN’s existing systems of record. It does not replace the core loan or finance systems. It reads from them through APIs, change feeds or nightly batches, reconciles what it finds, and presents each borrower as one timeline: application, documents, approvals, disbursements, repayments, contacts and decisions, each with its source and its audit trail.

On top of that timeline, AI is applied where it removes repetitive work without removing accountability. Document intelligence reads the forms that arrive in the highest volumes and sends anything uncertain to a review queue. A case copilot drafts summaries, suggests next actions with reason codes and prepares letters in Bahasa Melayu or English. Early-warning scoring helps collections teams reach borrowers before arrears harden. A policy assistant answers staff questions only from PTPTN’s own SOPs, and always shows its sources.

Every AI output is a recommendation. Decisions that affect a borrower’s loan, status or record are approved by a named officer and logged.

What PTPTN gets in the first release (8 to 12 weeks from data access): the unified case timeline, document extraction for the highest-volume form types with an officer review queue, and a repayment dashboard that checks ledger, status and bank confirmations against each other. Workflow automation, reconciliation automation, risk scoring and automated audit evidence packs follow in two further phases.

Section 2

The operating context

PTPTN is funding a growing borrower base while recovering a large, ageing book, and the policy rules it operates under change every Budget cycle.

RM10.7 billionin repayment arrears as of 31 August 2026
1.16 millionborrowers carrying those arrears
4 new rulesannounced in Budget 2027 for repayment from 2027

Sources: Ministry of Higher Education written parliamentary reply, reported 6 October 2026; PTPTN statement on Budget 2027, 11 October 2026.

Budget 2027 changes the workload, not just the rules

The initiatives PTPTN welcomed on 11 October 2026 include a repayment deferment for borrowers earning up to RM2,500 a month, a RM50 minimum repayment for those earning RM2,501 to RM3,000, a 10% discount for active borrowers with no arrears who repay consistently, and tax deductions for employers who repay on behalf of staff.

PTPTN has stated that implementation depends on verifying borrower income through cross-checks with LHDNM. That is a data-matching, eligibility and evidence problem at the scale of hundreds of thousands of borrowers, which is exactly what this platform is built to handle.

Where time and accuracy are lost today

From comparable public-sector loan operations, we expect discovery to confirm the following patterns, which the platform is designed to address:

  • Officers re-key fields from scanned forms and check names, NRIC numbers and dates by eye.
  • Answering “where is my case” means opening several systems and reading notes written by others.
  • Payments confirmed by banks, salary deduction files and the ledger disagree, and variances are resolved in spreadsheets.
  • Arrears are seen only once instalments are already well overdue, and outreach varies by team.
  • Audit and investigation requests require evidence to be gathered manually, record by record.
Section 3

Who the platform serves

Four internal groups, each with its own workspace, permissions and view of the same underlying record.

Loan officers and case managers

Application handling, verification, status decisions
  • A prioritised review queue with extracted fields and confidence scores
  • A case summary written from the timeline, with sources
  • Suggested next action with reason codes, which they accept, edit or reject

Repayment and collections teams

Monitoring, reminders, restructuring, escalation
  • Worklists ranked by early-warning score and days past due
  • Recommended outreach channel and script per borrower
  • Referral tracking for debt consultancy agencies (APH)

Operations and compliance

Process control, approvals, audit responses
  • Approval workflows with maker-checker controls
  • Complete read and write audit trail for sensitive data
  • One-click evidence packs for audits and investigations

Support and call centre

Case inquiries, “where is my case”
  • Plain-language case status with the next step and expected date
  • Masked identity fields by default, revealed only with a logged reason
  • Handover to the owning officer without re-explaining the case
Section 4

Outcomes and how we will measure them

Targets are relative to baselines we measure with PTPTN in the first three weeks. We commit to the measurement method now and agree the final numbers after baselining.

OutcomeMeasureIndicative targetWhen
Faster processing with fewer manual checksMedian days from complete application to decision−40%End of Phase 2
Fields keyed by hand per in-scope application−60%End of MVP
Accurate document extractionField-level accuracy on auto-accepted fields, sampled weekly≥ 98%End of MVP
End-to-end case visibilityShare of active loans with a complete, reconciled timeline≥ 95%End of MVP
Average handling time for “where is my case” calls−35%End of MVP
Reconciliation accuracyLedger vs bank variances resolved automatically≥ 70%End of Phase 2
Audit-ready traceabilityTime to assemble an evidence pack for one caseUnder 1 hourEnd of Phase 3
Reduced delinquencyBorrowers contacted within 7 days of a first missed instalment≥ 90%End of Phase 2
Roll rate from 30 to 90 days past due, versus control group−15%6 months after Phase 3
Section 5

Platform modules

Six modules share one data layer, one permission model and one audit log. Each is available in the staff web interface and as documented internal APIs, so PTPTN’s other systems can consume the same results.

Application and document processing

MVP · extended in Phase 2

Receives documents from existing intake channels, classifies them, extracts fields and validates them against each other and against PTPTN’s records before an officer sees the case.

Classification of NRIC copies, offer letters, guarantor forms, income statements and deferment applications
Field extraction with per-field confidence and the image region it came from
Validation rules for NRIC format and embedded date of birth, name match across documents, programme and institution codes
Review queue ordered by confidence, age and service level

Loan account and repayment tracking

MVP

Shows each loan’s principal, disbursements, schedule, ujrah, payments and adjustments, and checks the account status against the ledger and bank confirmations.

Repayment dashboard with next due date, history and arrears ageing
Channel view across salary deduction, direct debit, JomPAY, counter and online
Consistency checks between ledger, account status and bank confirmations
Schedule simulation for restructuring and deferment scenarios

Case management and timeline

MVP · workflow engine in Phase 2

The centre of the platform: one chronological record per borrower and loan, assembled from every connected system, with case ownership, tasks, notes and service levels.

Unified timeline of statuses, events and document references, each tagged with its source
Configurable workflows for application stages, deferment, restructuring and appeals
Maker-checker approvals with reason codes and delegated limits
Service-level tracking with escalation when a stage overruns

Budget 2027 eligibility engine

Proposed for MVP, subject to PTPTN’s final rules

A rules service that applies the 2027 deferment, minimum repayment and discount criteria to verified income data, explains each result, and routes exceptions to officers.

Income verification from the LHDNM cross-check, matched to borrowers with confidence scoring
Versioned rules, so every outcome can be traced to the rule set in force on that date
Borrower-level explanations for call centre and appeal handling
Batch runs with before-and-after impact reports for management

AI copilots

Officer copilot in MVP · operations copilot in Phase 3

Assistants embedded in the screens staff already use. They summarise, suggest and draft; they do not decide. Details are in Section 6.

Reporting and compliance evidence

Dashboards in MVP · evidence packs in Phase 3

Operational dashboards for managers and an evidence service that assembles, seals and exports everything relevant to a case or an audit question.

Operational reporting on queues, service levels, arrears and collections outcomes
Evidence packs with timeline, documents, decisions, approvers and access log
Integrity sealing using cryptographic hashes recorded in the audit store
Scheduled extracts for management, the Ministry and auditors
Section 6

AI capabilities, in order of internal value

We prioritise AI that saves officers time on work they already do, uses data PTPTN already holds, and can be checked by a person before it has any effect.

Case copilot for officers

Reads the case timeline and the extracted document fields, then produces three things on request.

  • A case summary of the borrower’s position, written in plain language with links to each source record
  • A suggested next action: approve, reject, request documents, offer restructuring or escalate, with reason codes drawn from PTPTN’s code list
  • A draft letter or message in Bahasa Melayu or English, built from approved templates and filled only with facts from the case
Guardrail: suggestions cannot change a record. The officer accepts, edits or rejects each one, and the choice and reason are logged, which also gives us the data to measure and improve the copilot.
Copilot summary · synthetic caseGenerated from 14 records
Borrower is 54 days past due after employer salary deductions stopped in August 2026 SD-file 08/26. Payment history before that was fully regular for 19 months Ledger. Verified 2025 income is RM2,860 a month LHDNM check, which falls within the Budget 2027 RM50 minimum repayment band.
Suggested next action: contact borrower to confirm employment status and offer a restructured schedule at the RM50 minimum.
Reason codes: R-EMP-02 employer remittance stopped · E-27-B income band eligible · Confidence 0.86 · Officer approval required

Document intelligence

OCR and structured extraction for the required forms, combined with deterministic rules.

Scan→Classify→Extract→Validate→Auto-accept or review
  • Extracts NRIC, names, programme, institution, dates, amounts and guarantor details
  • Format checks and cross-field consistency, such as name and date of birth matching the NRIC
  • Fields below the confidence threshold go to the review queue with the source region highlighted
Guardrail: thresholds are set per field from measured accuracy, so a misread NRIC digit is never accepted silently.

Repayment and delinquency assistance

Detects risk early and recommends a consistent response.

  • Rules for missed instalments and days past due, with arrears ageing buckets
  • Early-warning score from repayment behaviour, channel changes and case history
  • Recommended outreach: SMS or email reminder, officer call, restructuring offer or referral to collections
  • Each score shows its main contributing factors so officers can explain it
Guardrail: scores are tested for unequal impact across states, institution types and income bands before release and monthly after.

Policy assistant

A searchable, question-and-answer interface over PTPTN’s SOPs, circulars and guidelines.

  • Answers only from the approved policy library, with the document, section and effective date cited
  • Can link an answer to the case the officer has open
  • Says plainly when the library has no answer and suggests the owning unit
Guardrail: the library is versioned; superseded circulars are excluded unless the question is about a past date.

Operations copilot

For supervisors and compliance, in Phase 3.

  • Explains queue build-ups and service-level breaches from workflow data
  • Gathers evidence for an audit question across many cases
  • Drafts management summaries that cite the underlying reports
Guardrail: read-only by design; it cannot trigger workflows or change records.
Section 7

Data architecture on PTPTN’s internal data

The platform reads from PTPTN’s systems of record and writes back only through approved interfaces. Data stays in PTPTN-controlled infrastructure in Malaysia.

Core data domains

DomainKey contentsMain sourceSync method
Borrower masterIdentity, contact, institution, programme, study periodLoan applicationsAPI / CDC
Loan accountPrincipal, tenure, schedule, disbursements, status, defermentsLoan accountsCDC
Repayment ledgerPayments by channel, adjustments, ujrah, fees, reversalsFinance ledger, bank filesCDC Batch
Document repositoryUploads, document types, validity windows, OCR text, extracted fieldsDocument storeEvent feed
Workflow and auditWho did what, when, from which system, with what reasonAll systems and the platform itselfNative
External verificationsIncome cross-checks, employer remittances, bank confirmationsLHDNM, employers, banksBatch

Data quality and reconciliation

Matching with confidence. Records are joined on NRIC, normalised name, programme and loan reference. Each match carries a confidence score; exact NRIC and loan reference matches link automatically, partial matches go to a resolution queue rather than being merged by guesswork.

Reconciliation. Scheduled jobs compare account status, ledger postings and bank or employer confirmations. Where they disagree, the platform creates a discrepancy object instead of overwriting anything.

Discrepancy objects. Each has a type, the amount at stake, links to the conflicting records, a suggested resolution and an owner. Resolution is recorded with the evidence used, so the same issue can be explained months later.

Quality monitoring. Completeness, freshness and match rates are measured for every source and shown on an operations dashboard, with alerts when a feed is late or a match rate drops.

Section 8

Security, privacy and governance

The platform concentrates sensitive borrower data, so it is designed to make every access deliberate, minimal and recorded.

Control areaHow the platform implements it
Access controlRole-based access by team and function, with attribute rules for branch, case ownership and delegated limits. Single sign-on against PTPTN’s directory, with multi-factor authentication.
PII protectionField-level permissions. NRIC, contact details, income and bank data are masked by default for roles that do not need them; reveal requires a reason and is logged.
EncryptionTLS 1.2 or higher in transit, AES-256 at rest, keys held in PTPTN-controlled key management or HSM, with rotation.
Audit loggingEvery read and write of sensitive data is logged with user, role, purpose, time and source system. Logs are append-only and hash-chained so tampering is detectable.
AI data handlingModels run inside PTPTN’s hosting boundary. Borrower data is not sent to public AI services and is not used to train third-party models.
Segregation of dutiesMaker-checker on approvals, write-offs, restructures and rule changes. Administrators cannot approve their own changes.
ResilienceBackups with tested restores, defined recovery time and point objectives agreed in discovery, and a disaster recovery site where PTPTN has one.

Frameworks we design to. The Personal Data Protection Act 2010 and its 2024 amendments; the Cyber Security Act 2024 and NACSA guidance where applicable to PTPTN; Malaysian public-sector ICT security directives; ISO/IEC 27001 controls; and the National Guidelines on AI Governance and Ethics. We will map each requirement to a control and a test in the security design document, and support PTPTN’s own security assessment and penetration testing before go-live.

Section 9

Responsible AI and model operations

AI in a public lending context has to be measurably accurate, explainable to the borrower, and stoppable.

Human in the loop

Any action that changes a borrower’s status, balance, schedule or eligibility requires an officer’s approval. High-impact categories, such as rejections, legal escalation and blacklisting, require a second approver. The officer always sees the reasons and the evidence behind a suggestion.

Evaluation harness

Before any model or prompt reaches staff, it runs against a labelled test set drawn from PTPTN’s own historical cases. We report extraction precision and recall per field, copilot suggestion agreement with officers, false alert rates for early warnings, and fairness metrics across borrower groups. Release requires agreed thresholds to be met.

Monitoring in production

SignalAction when breached
Data drift in source feedsAlert data owner, pause affected scoring
Model drift in score distributionRe-validate, retrain on approval
Officer override rate risingReview suggestions with team leads
Extraction accuracy below thresholdRaise confidence threshold, route to review
Latency above targetScale services, fall back to manual queue
Audit chain verification failsFreeze writes, notify security

Each AI feature has a switch that returns the process to manual handling without affecting the rest of the platform.

Section 10

Delivery roadmap

Three phases after a short discovery. Each phase ends with a working release in production use, not a document.

MVP

8 to 12 weeks, if data access is ready
  • Unified case timeline per borrower and loan: statuses, events and document references
  • Document extraction for the two to four highest-volume form types
  • Officer review queue with confidence scores
  • Repayment dashboard: next due date, history, ledger vs status checks
  • Officer copilot: case summaries and draft letters

Exit criteria: timeline complete for at least 95% of active loans in scope; extraction meets agreed accuracy; pilot team using it daily.

Phase 2: automation

2 to 4 months
  • Workflow engine for application stages with standard approvals and audit trail
  • Payment reconciliation automation to reduce manual variance work
  • Delinquency alerts and collections triage lists
  • Budget 2027 eligibility engine at full scale
  • More form types in document intelligence

Exit criteria: agreed share of variances auto-resolved; contact within 7 days of a missed instalment for most accounts.

Phase 3: AI optimisation

3 to 6 months
  • Risk scoring v1 to predict delinquency or need for assistance
  • Copilot upgrades: richer summaries, faster evidence gathering, next-action suggestions
  • Operations copilot for supervisors and compliance
  • Automated compliance evidence packs for audits and investigations

Exit criteria: risk model passes accuracy and fairness thresholds; evidence pack assembled in under an hour.

Operational readiness, throughout

Human-in-the-loop approvals for all high-impact AI outputs, configured per action type
Evaluation harness for accuracy, extraction precision and false alert rate on every release
Monitoring of data drift, model drift, latency and audit integrity, with on-call response
Training and change management for each user group, with floor support at go-live
Section 11

Technology and hosting

Open, widely supported components that PTPTN’s IT division can operate, with no lock-in to a single AI vendor.

Staff web interface

React with TypeScript, responsive, accessible to WCAG 2.1 AA, Bahasa Melayu and English.

Platform services

Python (FastAPI) services for timeline, workflow, rules, evidence and AI orchestration.

Data

PostgreSQL for curated data and audit, object storage for documents, OpenSearch for search.

Ingestion

Debezium change data capture to Kafka, plus scheduled batch loaders with contract tests.

Document AI

Self-hosted OCR and layout models, fine-tuned on PTPTN forms, with rules-based validation.

Language models

Open-weight models hosted on-premise, or a Malaysia-region managed service if PTPTN approves; retrieval with pgvector.

Identity

Keycloak or PTPTN’s existing identity provider, SAML or OIDC, with MFA.

Operations

Kubernetes, GitOps deployment, Prometheus and Grafana monitoring, MLflow for model registry.

Security tooling

Vault for secrets, image scanning, SAST and DAST in the pipeline, SIEM integration.

Hosting options

PTPTN data centre

Full on-premise deployment. Maximum control; requires GPU capacity for language and document models.

Recommended

Hybrid

Data and audit on-premise; AI inference in a government-approved Malaysian cloud region under PTPTN’s tenancy. Faster to start, scales with load.

Government cloud

Deployment within the public-sector cloud arrangements available to PTPTN, subject to classification of the data.

Section 12

Team and project governance

A compact delivery team working alongside PTPTN’s process owners and IT division, with decisions taken in a fixed governance rhythm.

Engagement directorAccountable to PTPTN’s steering committee
Project managerPlan, risks, reporting, PTPTN coordination
Solution architectIntegration design, security architecture
Data engineers (2)Ingestion, matching, reconciliation
ML and AI engineers (2)Document AI, copilot, risk scoring, evaluation
Full-stack engineers (3)Staff interface, services, APIs
Business analystProcess mapping, rules, acceptance criteria
QA and security engineerTest automation, security testing
Change and training leadUser training, adoption, floor support
ForumMembersFrequencyPurpose
Steering committeePTPTN senior management, engagement directorMonthlyScope, priorities, phase sign-off
Project boardPTPTN project owner, IT, process owners, project managerFortnightlyProgress, risks, decisions
Design authorityPTPTN IT and security, solution architectFortnightlyArchitecture and security approvals
AI review panelCompliance, legal, operations, ML leadBefore each AI releaseEvaluation results, fairness, go or no-go
Sprint reviewPilot users and delivery teamEvery 2 weeksWorking software demonstration and feedback
Section 13

Risks, assumptions and dependencies

RiskLikelihoodMitigation
Source system data access is delayedMediumStart with batch extracts in week 1; prioritise the three tables the MVP needs; agree data contracts in discovery.
Data quality lower than expectedMediumMatching with confidence and discrepancy queues; quality dashboard from week 3; phase scope adjusts to measured quality.
Extraction accuracy varies on older scansMediumPer-field thresholds; low-quality scans routed to review; fine-tuning on PTPTN samples.
Staff trust and adoption of AI suggestionsMediumPilot team co-designs screens; suggestions always show reasons; override feedback improves the model.
Budget 2027 implementation rules changeHighVersioned, configurable rules engine; PTPTN policy owners edit rules without code changes.
Bias in risk scoringLowFairness testing before release and monthly; excluded attributes; AI review panel sign-off.

Assumptions

  • PTPTN provides read access to required tables, or extracts, within two weeks of kickoff.
  • A pilot group of officers, collections staff and call centre agents is available for testing.
  • Sample documents, SOPs and code lists are provided for model training and the policy library.
  • PTPTN provides hosting environments per the selected option.

Dependencies

  • Interface specifications for source systems and write-back endpoints.
  • Access arrangements for LHDNM income verification data, as determined by PTPTN.
  • PTPTN security review and penetration test windows before each go-live.
  • Decisions from the steering committee at each phase gate.
Section 14

Commercial framework

Pricing is submitted separately in the financial proposal, in line with PTPTN’s procurement procedures. The structure below shows how it is organised.

ComponentBasisPayment milestones
Discovery and MVPFixed price20% at kickoff · 30% at user acceptance · 50% at go-live
Phase 2Fixed price per agreed scope30% at design sign-off · 70% at go-live
Phase 3Fixed price per agreed scope30% at design sign-off · 70% at go-live
WarrantyIncluded12 months defect correction from each go-live
Support and model operationsAnnual service agreementQuarterly in advance, with service levels for response and resolution
Change requestsAgreed day ratesOn approval of each change request

All source code, configurations, models trained on PTPTN data, and documentation produced for PTPTN become PTPTN’s property, with full knowledge transfer to PTPTN’s IT division.

Section 15

Next steps

  1. Walk through the demo with PTPTN stakeholdersA 60-minute session covering each user group’s workspace using synthetic data.
    Week 0
  2. Sign a non-disclosure agreement and data access arrangementDefines what data the discovery team can see and where it is held.
    Week 1
  3. Run a three-week discoveryProcess maps, baselines for each measure, data contracts, and confirmation of the MVP form types.
    Weeks 1 to 3
  4. Prove extraction on PTPTN’s own documentsA measured accuracy report on two form types, using a sample PTPTN selects.
    Weeks 2 to 4
  5. Approve MVP scope and start the buildSteering committee sign-off on scope, plan and acceptance criteria.
    Week 4

See it working

The demo runs entirely in the browser with synthetic borrowers. Switch between officer, collections, compliance and call centre roles to see the same records with different permissions.

  • Review an AI-extracted application and correct a field
  • Ask the copilot for a case summary and draft a letter in BM
  • Triage arrears and run the Budget 2027 eligibility check
  • Generate a sealed evidence pack from the audit log

Open live demo