The bottleneck in most lending operations is not the credit decision; it is the preparation for it. A trained underwriter can read a bank statement, map the income, identify the obligations, and form a view. In typical micro-lending operations, getting from a PDF uploaded by the applicant to a decision-ready summary takes an estimated 30 to 45 minutes per file. At ten applications a day, that is manageable. At fifty, the queue starts backing up. At two hundred, the operation either stalls or the quality of review starts to slip.
The decision pack exists to solve that problem. Not by removing the underwriter from the process (that would create a different set of problems, both regulatory and practical), but by eliminating the preparation work that should never have required a human in the first place. Manual bank statement review costs 45 minutes per application. The decision pack reduces the underwriter's active review time to five minutes or less on a clean approve, and to under ten on a referral, while giving them more information than they had before.
What the pack contains
The decision pack is the output of AffyScore's full processing pipeline, structured for immediate human consumption. It has four components, each designed to replace a specific step in the manual workflow.
1. Behavioural score with reason codes
The behavioural score runs from 300 to 850. It is derived from the applicant's transaction history: income stability, balance trajectory, obligation coverage ratio, dishonour frequency, and spending volatility. Where a bureau score tells you what happened in the credit accounts over the past year, the behavioural score tells you what has been happening in the bank account over the past three months.
Alongside the score sits a recommendation (Clear, Caution, or Review) and a set of reason codes that explain the primary drivers. The reason codes are not opaque reference numbers. They are plain-language explanations: "Income deposits show high month-to-month variability (CV 38%)", "Three dishonoured debit orders in the last 60 days", "Average end-of-month balance has been declining for 90 days". Each code points to the specific behaviour that influenced the score.
This is what replaces reading 200 pages of transactions. The analyst does not need to scroll through three months of data looking for patterns; they have been identified, quantified, and surfaced as discrete, auditable signals.
2. Reg 23A affordability calculation
The Regulation 23A affordability section, drawing on the requirements set out in GN R202 of 2015, is structured to align with the evidence types that Regulation 23A requires credit providers to document. It presents gross income (validated against bank deposits), net income after statutory deductions, total identified existing obligations, estimated necessary living expenses by category, and the resulting disposable income available for new credit.
Every line is sourced from the bank statement data. Salary deposits, rental income, grant payments are categorised and validated. Debit orders for existing loan repayments, vehicle finance, insurance are mapped and totalled. Grocery, utilities, transport, medical spend are identified from transaction descriptions and averaged across the statement period. Categorisation is pattern-matched from transaction descriptions. Ambiguous or unrecognised patterns are flagged for manual review rather than silently classified.
This is what replaces the payslip plus the declaration form plus the manual calculation on a spreadsheet. The traditional approach requires the applicant to declare their income and expenses honestly, the credit provider to verify what they can, and an analyst to reconcile the two. The bank statement affordability section does that reconciliation automatically and flags discrepancies when the declared figures do not match what the transactions show.
If the applicant declares a monthly income of R28,000 but the bank statement shows net deposits of R19,400, that gap appears explicitly in the affordability section. If the applicant declares no existing loan obligations but the statement shows three active debit orders matching the repayment pattern of personal loans, those obligations appear in the output. The underwriter does not need to discover these discrepancies; they are handed them, with the evidence attached.
3. Collection date forecast
The collection date forecast identifies the optimal debit order date based on the applicant's salary timing and post-salary balance patterns. It includes a collectability percentage (a measure of how often sufficient funds were available on the recommended date during the statement period, based on the balance trajectory observed after each income deposit) and a predictability percentage that reflects how consistent the pattern has been across the statement period.
A high collectability with high predictability means the applicant receives regular income on a consistent date and maintains a sufficient buffer after receiving it. A high collectability with low predictability means the funds are typically there, but the timing varies enough that the debit order date carries more risk than the balance alone suggests. A low collectability is a flag regardless of predictability; the income arrives, but the balance gets spent down before a mid-month debit order date would hit.
The collectability percentage reflects observed historical patterns and should be weighed alongside other risk factors when setting the collection date.
This is information that exists in the bank statement for any analyst who knows what to look for. But surfacing it manually requires more than reading transactions; it requires calculating averages, identifying patterns, and forming a probabilistic view. The collection date forecast does that calculation explicitly, so the underwriter can see the recommendation and the evidence behind it without doing the arithmetic themselves.
4. Tamper assessment
Every decision pack includes a tamper assessment for each statement submitted. The assessment is one of three outcomes: clean, flagged, or failed. Flagged and failed assessments include specific anomaly codes (metadata inconsistencies, balance reconciliation failures, font anomalies, sequence gaps) that identify what triggered the assessment.
A flagged statement is not automatically a declined application. Some flags are minor (a document converted between PDF versions, for instance, which can alter metadata without changing the underlying data). Others are material: a balance that does not reconcile, or a date sequence with an unexplained gap. The underwriter sees the flag and the specific reason, and can make an informed judgement rather than treating every anomaly the same way.
What the tamper assessment replaces is not a manual check most credit providers were performing. Most were not checking at all because the tools required to do it properly (metadata analysis, cryptographic verification, sequence validation) are simply not part of a manual review workflow. The decision pack includes it by default. Document integrity issues that would otherwise require manual forensic review are now flagged as a standard line item in every assessment. Not all flags indicate fraud; some may result from legitimate format conversions. The underwriter makes the determination.
The output formats
The decision pack is available in three formats: JSON, PDF, and XLSX. All three contain the same information. The format choice depends on who is consuming it and how.
The JSON response is for lending platforms and loan origination systems. The structured data feeds directly into the credit provider's workflow: scoring thresholds can be applied programmatically, decisions can be logged automatically, and the full pack can be stored against the application record without any manual transcription. This is the integration path for credit providers who process high volumes and want the decision pack to be invisible to the analyst. The system queries, the system receives, and the analyst sees only the output that requires a human eye.
The PDF report is for the underwriter's desk and the compliance file. It is formatted for reading; the score and recommendation are prominent, the reason codes are annotated, the affordability calculation is laid out clearly, and the tamper assessment appears as a summary with detail available on the following pages. The PDF can serve as a record of the analytical component of the affordability assessment. The credit provider remains responsible for ensuring the overall assessment documented in their credit file meets NCA requirements. The PDF is versioned, timestamped, and reproducible.
The XLSX workbook is for analysts who want to work with the underlying data directly: to query the transaction categorisation, to review the income and obligation mapping line by line, or to reconcile the affordability numbers against their own calculations. The workbook exposes the full transaction register with category tags, the income validation detail, and the obligation mapping alongside the summarised outputs.
The underwriter's workflow with the pack
Here is what a bank statement application review looks like with the decision pack in hand.
The PDF opens. The first thing visible is the score (a number on the 300-850 scale) and the recommendation: Clear, Caution, or Review. This takes ten seconds.
If the recommendation is Clear, the underwriter scans the reason codes to confirm there are no material flags that warrant a second look despite the positive recommendation. A score of 720 with reason codes indicating strong income stability and a healthy balance trajectory would typically support a positive decision under most lender policies. A score of 720 with a reason code flagging a single month of elevated dishonours (which the system has noted but not weighted heavily enough to change the recommendation) might warrant a brief review of the relevant transactions. That review takes another minute, not fifteen.
The underwriter checks the affordability ratio. The disposable income figure, set against the proposed instalment, confirms whether the application is within the lender's policy parameters. If the ratio is borderline, the underwriter looks at the income validation detail: is the income stable enough that the disposable income figure is reliable, or is there volatility that warrants a more conservative view? That check takes another minute.
The tamper assessment has a green indicator. No further action required on document integrity.
The collection date forecast recommends the 25th with an 84% collectability and 91% predictability. The proposed debit order date in the application is the 26th; this is close enough that no adjustment is needed.
Total time: under five minutes. The underwriter approves the application, logs the decision against the PDF that is already in the file, and moves to the next one.
That workflow describes a straightforward approve with clean data. Complex cases, referrals, or applications with ambiguous flags will require deeper investigation and typically take eight to fifteen minutes. The five-minute figure reflects optimal conditions, not a guaranteed outcome.
For a Caution or Review outcome, the process adds a step: the underwriter reviews the specific reason codes that triggered the referral and makes a judgement call. That judgement is the part that still requires a human. But it is informed by the same decision pack, and it takes eight minutes rather than forty-five.
The compliance benefit
The NCA requires credit providers to perform an affordability assessment before granting credit. It does not specify the exact method, but it does require that the assessment be defensible: that the credit provider can demonstrate, if challenged, that they gathered sufficient information and applied it correctly. Documentation quality is a significant factor in affordability defence, as legal commentary has noted. Cliffe Dekker Hofmeyr's analysis of the duty to conduct proper affordability checks discusses the importance of contemporaneous documentation. AffyScore aims to provide structured documentation of the analysis performed; the credit provider remains responsible for ensuring their overall process meets NCA requirements.
Manual affordability assessments are defensible in principle but difficult to reconstruct in practice. The analyst's spreadsheet may or may not have been saved. The payslip may be in a physical file or may have been scanned and lost in a folder structure. The notes from the review (if they exist at all) live in someone's email thread or handwritten on a form. When a compliance audit asks for the affordability evidence behind a specific loan granted eighteen months ago, finding it takes time. Reproducing the calculation to verify it was done correctly takes more.
Every AffyScore decision pack is versioned and tied to the statement batch it was generated from. The score, the reason codes, the affordability calculation, the tamper assessment: the output is designed to be reproducible from the same input data. If the underlying transactions are retained (which the NCA requires for the relevant period), the full affordability evidence chain can be reconstructed at any point. The PDF that went into the credit file is not the only copy; the structured data that generated it is also available.
This matters for three reasons. First, it means the credit provider's affordability evidence is standardised across every application, not dependent on which analyst processed it or how thoroughly they documented their work. Second, it means audit responses are mechanical rather than archaeological. Third, it means that as regulatory expectations around affordability documentation evolve, the credit provider's documentation already meets the emerging standard, not just the current minimum.
What does not change
The decision pack does not make the credit decision. That remains with the underwriter, and intentionally so. The NCA's provisions on reckless lending place the responsibility for affordability assessment squarely on the credit provider, not on any tool the credit provider uses. Full automation of the credit decision (removing the human entirely) creates a regulatory exposure that carries real NCA risk.
What the pack removes is the preparation burden that was consuming the underwriter's capacity without requiring their judgement. The 40 minutes that used to go into transaction reading, spreadsheet building, payslip cross-referencing, and manual balance reconciliation. That is the time the pack eliminates. The five minutes that remain are the part that actually requires a human: reading the structured output, applying lender policy, and making a defensible decision.
The underwriter's value is their judgement. The decision pack is what gives them the time to use it.
