How Covenant Desk reads, checks and grades a credit file.
Each method below says what it does in an SME or property credit workflow, how it runs in your deployment, the safeguards around it, and the published research behind it.
Every number in a Covenant Desk credit file traces to a source a person confirmed. The grade comes from rules your credit team can read, and the decision from a named credit officer with the authority to make it, with the AI preparing the file. These rules follow the expectations of explainability, human oversight and auditability in the HKMA's High-level Principles on Artificial Intelligence, Taiwan's FSC guidelines for AI in the financial industry and the MAS FEAT principles; meeting your own obligations under them remains your institution's assessment.
Each method carries one or more labels: In the platformConfigured for your bankAvailable
01
Document reading with arithmetic checks
In the platformConfigured for your bank
What it does
Audited accounts, management accounts and bank statements are read into a standard spread. Each proposed figure carries the line it came from. The spread is then checked against its own accounting identities: where a stated subtotal or total disagrees with the sum of its parts, the check is shown before anyone relies on the figure. Bank inflows over twelve months are reconciled with declared revenue.
In your deployment
Text-based PDFs are read on the server. Scanned pages and phone photos are read by layout-aware OCR that handles English and Traditional Chinese on the same page, tested on your own document types during the implementation.
Safeguards
Extraction prepares the data and a person confirms every value, so every check runs on figures a named person has accepted.
Monthly inflows, outflows and balances from the company's bank statements are analysed beside the financial statements, so the analyst sees whether the business the accounts describe is the business the bank account shows.
In your deployment
Account behaviour changes months before annual accounts are published. Transaction-level categorisation of mixed Chinese and English narratives, and monitoring of facility usage after drawdown, are part of the same analysis.
Safeguards
Categorisation accuracy is measured for each deployment before it is used in a grade, so every derived ratio rests on tested categories.
The SME scorecard is a versioned set of explicit rules over ratios, cash flow, security and conduct, with reason codes. A missing input is shown as unassessed rather than filled with a guess, and a file that cannot be graded cannot be approved.
In your deployment
Your thresholds, weights and grade bands are loaded in your deployment. Calibration against your own loan history is part of the pilot, with any statistical challenger model constrained so each factor moves risk in the economically sensible direction and explained factor by factor.
Safeguards
Validation of any calibrated model belongs to your model-risk function, and we supply the evidence it needs.
Covenants are recorded on the facility and tested when a user runs them, with breaches and waivers logged against the case.
In your deployment
A discrete-time hazard model runs over account behaviour, ratio drift between periods and registry events, ranking performing loans by risk of deterioration so relationship managers review the few files that need it, each with its reasons.
Safeguards
An early-warning signal triages files for your credit officers; a grade or a limit changes only when a named person changes it.
Directors, shareholders and guarantors are linked to every case, and the company is re-screened at the final facility letter and at booking.
In your deployment
Across cases, the same person often guarantees several borrowers. A network view shows a shared guarantor or director under stress on every exposure it touches, with explicit contagion rules and graph models alongside.
Safeguards
Published graph models for guarantee networks were trained on mainland Chinese bank data, so explicit rules lead and the graph models are tested on your own book before use.
A language model proposes figures from a document and drafts text. Identifiers are masked before any call, the model prepares figures and text, rules grade the case, and a person confirms every figure before rules use it.
In your deployment
Your bank chooses the model: one it already approves, or one hosted in your region.
Safeguards
Benchmarks show language models answering questions about financial filings wrongly when left to generate numbers. Numbers in our documents come from the confirmed data, not from the model.
Valuations, haircuts, loan-to-value and guarantees are recorded on each case, with an expected-loss estimate and exposure across the borrower group.
In your deployment
The platform adds loan-to-value re-marking from published price indices and a flag where a surveyor's valuation departs from the market by more than a set margin.
Safeguards
An automated estimate is a check on a valuation, alongside the surveyor valuation your policy requires.
Every financial and credit action is written to an append-only log in which each entry carries a cryptographic hash of the one before it. A changed or deleted entry breaks the chain, and the break can be shown.
In your deployment
Every assessment, schedule and credit memo is stamped with the engine version that produced it, so a number can be traced to the code that computed it.
Safeguards
Hash-chaining shows that a record changed, and access control and retention follow your policy.
Vendors do not get models approved; banks do. Our job is to give your validators the evidence they need.
01
Calculation reference
Every formula with its day count, rounding and treatment of missing inputs.
02
Recomputation workbook
The spread and ratios in Excel, so your analysts recompute any figure and get the same answer.
03
Engine and policy versions
The version that produced each assessment, schedule and memo.
04
AI processing record
Which model was called, when, for which document, with what masking.
05
Known limitations
A written list of what each method does not do, kept current.
06
Control mapping
Our controls mapped to the rules your supervisor applies, for your own assessment.
Data
Why connected data matters for SME credit.
In an HKMA study of about 80,000 Hong Kong SME loans from 2022 to 2025, banks that used the Commercial Data Interchange more actively charged on average 36 basis points less and asked for fewer collateral items. Connecting your deployment to the CDI, your credit bureau and your core banking system is part of the implementation.
