Quality Control Agent AI

Invoked when: the daily file batch is available in the LOS — sampling runs and audits complete within business hours

• –Stratifies every day's processed files into three risk tiers before sampling — high-risk (first-time applicants, high loan amounts above the 80th percentile of the product's disbursed amount distribution, applicants from geographies with elevated historical error rates, or files processed by agents with a current error rate above 10%), medium-risk (standard applicants with no elevated risk factors), and low-risk (returning borrowers with clean prior file history processed by agents with error rates below 5%). The stratification is recalculated daily based on current data — a geography's risk tier is updated weekly as its error rate trend changes.
• –Applies differential sampling rates to each tier: 100% of high-risk files, 30% of medium-risk files, and 10% of low-risk files. This concentrates audit resources where errors are most likely and most consequential — a first-time high-value applicant file with an error that reaches disbursement creates a much larger credit and compliance exposure than the same error in a low-risk file. The 100% high-risk coverage ensures no high-risk file escapes audit; the statistical sampling of lower tiers provides error rate visibility without proportional resource cost.
• –Escalates any agent to 100% file audit coverage when their personal error rate exceeds 15% across the prior week's audited files — ensuring that the agent's full output is checked until their error rate returns to threshold. An agent whose 30% sample shows 15% errors has an expected full-file error rate that is unacceptable; the 100% coverage confirms the extent of the problem and provides the supervisor with a complete picture of the agent's output quality.

Invoked when: a file is selected for audit — error tagging is completed and tags are applied to the file record within 24 hours of selection

• –Checks every audited file against the 20-error taxonomy — the standardised set of operational errors that covers the full range of file quality issues observed across the lending operation: missing mandatory document (5 sub-categories by document type), expired document, illegible document, name mismatch (minor and significant variants treated separately), address inconsistency, income figure transcription error, income figure inconsistency across documents, bank statement period gap, application form field incomplete, application form field incorrect (data entry error rather than missing), signature absent, photograph non-compliant, co-applicant KYC incomplete, guarantor declaration absent, credit bureau report stale (older than the 30-day policy limit), LTV calculation error, interest rate applied incorrectly to product, disbursement amount mismatch vs sanction letter, and processing fee calculation error. Each error is tagged to exactly one category; multi-error files receive one tag per error.
• –Distinguishes between critical errors (those that would prevent the file from being legally enforceable or regulatorily compliant — missing mandatory KYC document, guarantor declaration absent, disbursement amount mismatch vs sanction letter) and standard errors (those that reduce operational quality but do not affect enforceability — transcription errors, bank statement gaps, photograph non-compliance). Critical error files are escalated to the supervisor immediately; standard error files receive the standard 24-hour feedback routing.
• –Records the agent responsible for each error — matched from the LOS processing record — enabling agent-level error rate tracking. Where a file has been processed by multiple agents at different stages, the error is attributed to the agent who was responsible for the stage in which the error occurred, not to the final agent in the processing chain.

Invoked when: an error is tagged — feedback card dispatched to the responsible agent within 24 hours; weekly pattern analysis runs every Monday

• –Dispatches a structured feedback card to each agent for every error in their audited files — identifying the specific file (by reference number, not borrower name), the specific error category, the exact field or document where the error occurred, the correct standard that was not met, and the corrective action required. The feedback card is not a performance warning — it is a coaching tool. The language is instructive, not punitive: "the bank statement attached covers March–May; the policy requires the 3 most recent months, which at the application date (June 14) means April–June. Please request the updated statement." Specific, actionable feedback is more effective at reducing recurrence than a generic error notification.
• –Tracks each agent's error history across a 12-week rolling window — building the dataset that distinguishes a one-off error (an agent who makes this error once in 12 weeks) from a recurring pattern (an agent who makes the same error consistently). Recurring errors — the same error category appearing in an agent's audited files for three or more consecutive weeks — are escalated to the supervisor with the full 12-week history, indicating that the agent has received feedback on this error but has not corrected it. The supervisor's intervention is the next level of coaching, not an HR action.
• –Produces a weekly ops QC dashboard for the ops head — showing error rates by category, by agent, and by branch for the current week and the prior 4-week trend. The category breakdown tells the ops head where process or training gaps exist across the team; the agent breakdown tells them where individual coaching is needed; the branch breakdown tells them whether quality problems are concentrated in specific locations that may have supervisory or infrastructure issues.