Hong Kong credit reference system (CRS) guide
香港信貸資料系統（CRS）指南
How credit reference agencies, Credit Data Smart, loan origination and loan management systems differ — and which borrower record belongs in each.
Free assessment checklist
How it works
Concrete workflows from first search to borrower follow-up.
Each Covenant Desk workflow starts from a real operational trigger, uses scoped borrower data, and ends with a visible action, record, or audit point.
See it running
Two walkthroughs of the platform.
Recorded in our demo environment. Every borrower, company and figure is fictional, and there is no sound.
Where each stage stands
One case, from application to a monitored facility
Each stage is available now, and each one is configured to your credit policy, documents and systems during implementation.
- 01
Intake
The company, its directors, shareholders and guarantors in one case. Staff open cases or the company applies on your own online form; a secure upload link and a checklist collect the documents. iAM Smart and CorpID are available, enabled for your institution on request. Identity numbers are stored encrypted and shown masked.
In the platform
- 02
Read
Financial statements, bank statements, certificates, HKID and business registration documents are read by OCR and AI document extraction. Figures go to the right fields with the page and line each came from, and a person confirms every value before it is used.
In the platform
- 03
Assess
Financial spreading by period; DSCR, leverage, current ratio and interest cover; bank-account analysis; a versioned 100-point scorecard in which missing evidence shows as unassessed; a standard downside case and structuring options compared side by side.
In the platform
- 04
Decide
Approval limits by amount, escalation, and a second-person check on exceptions. A file that cannot be graded cannot be approved. AI prepares the decision and your credit officers approve it, under your delegated authority matrix and credit committee, configured for your bank.
In the platformConfigured for your bank
- 05
Document
A facility letter with a section for your standard terms, a credit memo, and an origination report with Excel export. Your own templates and language versions are built into your implementation.
In the platformConfigured for your bank
- 06
Monitor
Collateral, haircuts and loan-to-value; covenants recorded and tested, with breaches and waivers logged; the company re-screened at the final letter and at booking. The AI banking agent monitors accounts and raises early-warning flags, most urgent first.
In the platform
Workflow paths
Seven workflows the demo should cover.
- 01
Borrower search and credit review
Trigger: a new or returning borrower applies. Staff search by HKID, name, mobile or loan number, open Customer 360 and review DSR, partakers, credit reference, loan history and AI support. Output: decision context for human approval or rejection, with access scope and lookup audit.
- 02
Cross-company exposure check
Trigger: a borrower has activity across a lender group. Staff inspect active and total loan counts by company and merged loan history. Output: concentration context before new lending, under exact-HKID lookup and company permission rules.
- 03
Repayment notice follow-up
Trigger: an overdue or upcoming notice sync. Operations filter notices, sort by priority, open matched borrower profiles and log follow-up. Output: a coordinated collections queue, with match confidence and sync-run history.
- 04
Borrower FPS repayment
Trigger: a borrower opens an instalment. The client app shows the FPS QR and reference code, the borrower pays in their bank app, and staff reconcile the received payment. Output: repayment status and audit trail.
- 05
Intake to live record
Trigger: an email, WhatsApp, Excel, phone or walk-in enquiry. The lead enters the queue, staff correct fields, then approve, reject or request more information. Output: a customer record or a rejected lead with its reason.
- 06
AI-assisted review
Trigger: a reviewer needs a summary or portfolio answer. The AI banking agent summarises the file, looks up records, ranks risk and lists overdue notices. Output: a prepared review for staff, approved by a named credit officer.
- 07
Admin and permissions
Trigger: a lender adds users or changes team scope. Admins configure roles, staff levels, company units, data sources, sync jobs, prompts, audit and reports. Output: a controlled operating environment.
Technology
Methods from the research literature, used where a person can check them.
We pick methods with published evidence behind them, and we put each one where its output can be verified: AI prepares, people confirm, rules grade.
- 01
Document reading with arithmetic checks
Statements are read into a standard spread, then checked against their own accounting identities: a stated total that disagrees with its parts is flagged before anyone relies on it, and bank inflows are reconciled with declared revenue. Scans are read by layout-aware OCR for mixed English and Traditional Chinese.
In the platform
Huang et al., LayoutLMv3, ACM MM 2022; PaddleOCR 3.0 technical report, 2025
- 02
Cash-flow evidence
Account behaviour moves months before accounts do. Monthly inflows, outflows and balances are analysed beside the statements, with transaction-level categorisation of mixed-language narratives.
In the platform
Norden and Weber, Review of Financial Studies, 2010; Gambacorta et al., BIS Working Paper 834, 2019
- 03
Scoring you can read
The scorecard is explicit rules with reason codes, and a missing input shows as unassessed rather than as a guess. Calibration on your own loan history, with a constrained challenger model explained factor by factor, is part of your implementation.
In the platformConfigured for your bank
Rudin, Nature Machine Intelligence, 2019; Lundberg and Lee, NeurIPS 2017
- 04
Early warning
Covenants are tested and breaches logged. A hazard model over account behaviour, ratio drift and registry events lets relationship managers see the few files that need attention first, with the reasons.
In the platform
Shumway, Journal of Business, 2001; Norden and Weber, 2010
- 05
Guarantor and director networks
Directors, shareholders and guarantors are linked on every case and across cases, so a shared guarantor or director under stress shows up on every exposure it touches.
In the platform
Cheng et al., networked-guarantee loan risk, IJCAI 2019
- 06
Language models as drafters, not scorers
Benchmarks show language models answering financial questions wrongly when left to generate numbers. So the model proposes figures from a document, with identifiers masked, and every figure is confirmed by a person before rules use it.
In the platform
Islam et al., FinanceBench, 2023
- 07
Property collateral
Valuations, haircuts and loan-to-value are recorded on each case, with index-based re-marking of loan-to-value and flags where a valuation departs from the market.
In the platform
Hong Kong Rating and Valuation Department price indices; US interagency AVM quality-control rule, 2024
- 08
Tamper-evident record
Each action is chained to the one before it by a cryptographic hash, so a changed or deleted entry breaks the chain and the break can be shown to an auditor.
In the platform
Crosby and Wallach, USENIX Security 2009
Decision boundary
AI prepares the decision. Your credit officers approve it.
Covenant Desk is software and does not provide loans. The AI banking agent prepares the file, the credit memo and the recommendation; your credit officers approve every facility under your own credit policy, with four-eyes, approval matrices and a full audit trail.
Hong Kong lending software guides
Deeper reference reading for Hong Kong lending teams.
Four in-depth guides covering credit reference data, loan origination, loan servicing, and Credit Data Smart readiness ahead of the rules proposed to take effect on 1 June 2027. All four are written in Traditional Chinese (zh-HK) for Hong Kong readers.
Loan origination system (LOS) guide for Hong Kong lenders
香港貸款發起系統（LOS）指南
The application-to-disbursement path: intake, documents, KYC, DSR, credit review, approval, and the handover into servicing.
Free process blueprint
Loan management system (LMS) guide for Hong Kong lenders
香港貸款管理系統（LMS）指南
Repayment schedules, collection reconciliation, arrears follow-up, notices, statements and portfolio monitoring after disbursement.
Free health check
Credit Data Smart (CDS) readiness guide for 1 June 2027
信資通（CDS）2027 年準備指南
What lenders should prepare before the proposed Credit Data Smart rules take effect on 1 June 2027, including the record fields to structure now.
Free data checklist
All Hong Kong lending software guides on the Covenant Desk homepage
Walk through these workflows on your own book.
We set up a personalised demo around your own products and approval steps, then your team tries it before you decide.
Book a personalised demo