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.
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.
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.
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
- 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
- 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
- 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
- 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
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.
| Outcome | Measure | Indicative target | When |
|---|---|---|---|
| Faster processing with fewer manual checks | Median 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 extraction | Field-level accuracy on auto-accepted fields, sampled weekly | ≥ 98% | End of MVP |
| End-to-end case visibility | Share 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 accuracy | Ledger vs bank variances resolved automatically | ≥ 70% | End of Phase 2 |
| Audit-ready traceability | Time to assemble an evidence pack for one case | Under 1 hour | End of Phase 3 |
| Reduced delinquency | Borrowers 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 |
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
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.
Loan account and repayment tracking
Shows each loan’s principal, disbursements, schedule, ujrah, payments and adjustments, and checks the account status against the ledger and bank confirmations.
Case management and timeline
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.
Budget 2027 eligibility engine
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.
AI copilots
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
Operational dashboards for managers and an evidence service that assembles, seals and exports everything relevant to a case or an audit question.
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
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.
- 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
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
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
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
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.
PTPTN source systems
Ingestion and sync
Data platform
Platform services
Access
Core data domains
| Domain | Key contents | Main source | Sync method |
|---|---|---|---|
| Borrower master | Identity, contact, institution, programme, study period | Loan applications | API / CDC |
| Loan account | Principal, tenure, schedule, disbursements, status, deferments | Loan accounts | CDC |
| Repayment ledger | Payments by channel, adjustments, ujrah, fees, reversals | Finance ledger, bank files | CDC Batch |
| Document repository | Uploads, document types, validity windows, OCR text, extracted fields | Document store | Event feed |
| Workflow and audit | Who did what, when, from which system, with what reason | All systems and the platform itself | Native |
| External verifications | Income cross-checks, employer remittances, bank confirmations | LHDNM, employers, banks | Batch |
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.
Security, privacy and governance
The platform concentrates sensitive borrower data, so it is designed to make every access deliberate, minimal and recorded.
| Control area | How the platform implements it |
|---|---|
| Access control | Role-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 protection | Field-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. |
| Encryption | TLS 1.2 or higher in transit, AES-256 at rest, keys held in PTPTN-controlled key management or HSM, with rotation. |
| Audit logging | Every 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 handling | Models 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 duties | Maker-checker on approvals, write-offs, restructures and rule changes. Administrators cannot approve their own changes. |
| Resilience | Backups 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.
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
| Signal | Action when breached |
|---|---|
| Data drift in source feeds | Alert data owner, pause affected scoring |
| Model drift in score distribution | Re-validate, retrain on approval |
| Officer override rate rising | Review suggestions with team leads |
| Extraction accuracy below threshold | Raise confidence threshold, route to review |
| Latency above target | Scale services, fall back to manual queue |
| Audit chain verification fails | Freeze writes, notify security |
Each AI feature has a switch that returns the process to manual handling without affecting the rest of the platform.
Delivery roadmap
Three phases after a short discovery. Each phase ends with a working release in production use, not a document.
MVP
- 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
- 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
- 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
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.
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.
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.
| Forum | Members | Frequency | Purpose |
|---|---|---|---|
| Steering committee | PTPTN senior management, engagement director | Monthly | Scope, priorities, phase sign-off |
| Project board | PTPTN project owner, IT, process owners, project manager | Fortnightly | Progress, risks, decisions |
| Design authority | PTPTN IT and security, solution architect | Fortnightly | Architecture and security approvals |
| AI review panel | Compliance, legal, operations, ML lead | Before each AI release | Evaluation results, fairness, go or no-go |
| Sprint review | Pilot users and delivery team | Every 2 weeks | Working software demonstration and feedback |
Risks, assumptions and dependencies
| Risk | Likelihood | Mitigation |
|---|---|---|
| Source system data access is delayed | Medium | Start with batch extracts in week 1; prioritise the three tables the MVP needs; agree data contracts in discovery. |
| Data quality lower than expected | Medium | Matching with confidence and discrepancy queues; quality dashboard from week 3; phase scope adjusts to measured quality. |
| Extraction accuracy varies on older scans | Medium | Per-field thresholds; low-quality scans routed to review; fine-tuning on PTPTN samples. |
| Staff trust and adoption of AI suggestions | Medium | Pilot team co-designs screens; suggestions always show reasons; override feedback improves the model. |
| Budget 2027 implementation rules change | High | Versioned, configurable rules engine; PTPTN policy owners edit rules without code changes. |
| Bias in risk scoring | Low | Fairness 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.
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.
| Component | Basis | Payment milestones |
|---|---|---|
| Discovery and MVP | Fixed price | 20% at kickoff · 30% at user acceptance · 50% at go-live |
| Phase 2 | Fixed price per agreed scope | 30% at design sign-off · 70% at go-live |
| Phase 3 | Fixed price per agreed scope | 30% at design sign-off · 70% at go-live |
| Warranty | Included | 12 months defect correction from each go-live |
| Support and model operations | Annual service agreement | Quarterly in advance, with service levels for response and resolution |
| Change requests | Agreed day rates | On 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.
Next steps
- Walk through the demo with PTPTN stakeholdersA 60-minute session covering each user group’s workspace using synthetic data.Week 0
- Sign a non-disclosure agreement and data access arrangementDefines what data the discovery team can see and where it is held.Week 1
- Run a three-week discoveryProcess maps, baselines for each measure, data contracts, and confirmation of the MVP form types.Weeks 1 to 3
- Prove extraction on PTPTN’s own documentsA measured accuracy report on two form types, using a sample PTPTN selects.Weeks 2 to 4
- 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