The Settlement Reconciliation Agent: How Finance AGI Catches What Manual Review Misses
Every D2C brand operating across multiple marketplaces and payment gateways is owed money it does not know it is owed. The settlement process where Myntra, Amazon, Flipkart, Razorpay, PayU, and other platforms reconcile payments against shipped orders and transfer the net amount contains systematic discrepancies that manual review catches inconsistently and automated rule-based tools catch incompletely. Short settlements, double TDS deductions, commission overcharges, return deductions without corresponding return receipts, and currency conversion errors accumulate quietly across thousands of transactions per month. At moderate volumes, this leakage runs to lakhs per quarter. At scale, it runs to crores annually. SuperManager AGI's Finance AGI was built to find every rupee of it.
Finance AGI achieved a 22.4 percentage point hallucination reduction in financial analytics in the ADA evaluation the largest improvement of any domain tested. This piece explains why structured financial data retrieved via direct database access is the single most accurate foundation for autonomous financial reconciliation.
Why Manual Reconciliation Fails at Scale
The settlement reconciliation problem is fundamentally a data joining problem at a scale and complexity that exceeds what human reviewers can execute reliably. A single marketplace settlement report may contain 4,000 to 40,000 line items, each representing a specific transaction: a payment, a return deduction, a commission charge, a TDS deduction, a shipping fee adjustment, or a promotional reimbursement. Reconciling this against the brand's own order management system, warehouse management system, and accounting ledger requires joining three to five datasets on multiple keys order ID, AWB number, SKU, transaction date with tolerance logic for timing differences, currency rounding, and partial fulfilments.A skilled finance analyst can reconcile a single marketplace's monthly settlement in two to four hours if the data is clean. In practice, settlement data is never fully clean. Marketplace reports use inconsistent column names across report versions, contain encoding errors, split transactions across multiple rows without clear indicators, and apply deduction logic that changes without advance notice. An analyst who reconciles Myntra settlements monthly has learned to work around Myntra-specific data quirks. When that analyst leaves, the institutional knowledge leaves with them. The next analyst makes errors for two to three months before learning the same workarounds.Beyond the volume and complexity problem, manual reconciliation has a sampling problem: analysts working under time pressure reconcile a representative sample of high-value transactions and assume the remainder is correct. This is a reasonable heuristic under resource constraints, but it systematically misses the category of discrepancy that Finance AGI is best at finding small, consistent, systematic errors that individually are too minor to flag but in aggregate represent significant leakage. A 0.12% commission overcharge on every transaction is invisible in a line-by-line review but visible immediately in a statistical analysis across the full settlement dataset.
The Accuracy Foundation: Direct Database Access vs. Document Parsing
The 22.4 percentage point hallucination reduction that Finance AGI achieved in the ADA evaluation relative to document-parsing approaches to financial reconciliation comes from a single architectural decision: Finance AGI accesses structured financial data directly from the source database rather than parsing settlement PDFs, CSV exports, or email attachments. This distinction is more significant than it initially appears.Document parsing approaches including many AI-powered reconciliation tools work by ingesting settlement reports as documents and extracting data from them using OCR, regex patterns, or language model parsing. This approach introduces error at every step of the extraction pipeline. OCR misreads numeric characters in poor-quality PDFs. Regex patterns break when report formats change. Language model parsing produces plausible-sounding but incorrect values when document structure is ambiguous. Each of these error sources is small individually but compounds across thousands of transactions, producing a reconciliation output that has high apparent coverage but material accuracy gaps in exactly the high-stakes transactions where accuracy matters most.Finance AGI connects directly to the organisation's ERP, OMS, and accounting database through secure, read-only API connections and to marketplace settlement APIs through direct integration where available. When direct API access is not available, Finance AGI ingests structured data exports with full schema awareness rather than treating them as unstructured documents. The reconciliation logic executes against data that has already been validated, typed, and structured not data that has been extracted from a presentation-layer report with all the ambiguity that implies. This is why the accuracy improvement in the ADA evaluation was largest in the financial domain: it is the domain where the gap between 'document that looks right' and 'data that is right' is most consequential.
What the Settlement Reconciliation Agent Catches
Short settlements and underpayments
Finance AGI joins every order in the brand's OMS against the corresponding settlement line in the marketplace report, flagging orders where the settlement amount is below the expected net payout after allowable deductions. It tolerates the timing difference between order completion and settlement cycle typically 7 to 15 days depending on the marketplace and distinguishes between orders that are genuinely pending settlement and orders that have been settled at a lower-than-expected amount. Short settlements accumulate most commonly from commission rate misapplication (the marketplace charges the wrong rate category) and from return deductions applied to orders where no return was received by the warehouse.
Duplicate and double deductions
TDS deductions, commission charges, and shipping fee adjustments are sometimes applied twice across a single settlement cycle once in the transaction-level report and once as a consolidated deduction in the summary. Manual reviewers rarely cross-reference transaction-level and summary-level data in the same settlement document. Finance AGI joins both levels and flags every instance where the same deduction appears to have been applied at both levels, with the full transaction chain for finance team review and dispute filing.
Return deductions without receipt confirmation
When a customer returns an order, the marketplace deducts the payout from the brand's settlement account. Finance AGI cross-references every return deduction against the brand's warehouse management system to confirm that the corresponding return was actually received, inspected, and accepted. Return deductions for items that were never received by the warehouse a common source of leakage, particularly for high-NDR categories are flagged immediately with the original AWB number, the marketplace deduction date, and the WMS receipt status.
Commission rate anomalies
Commission rates on marketplaces vary by category, subcategory, product price band, and promotional participation status. Finance AGI maintains a current commission rate table for every marketplace and applies it to every transaction, flagging instances where the applied commission rate deviates from the expected rate by more than a configurable tolerance. Rate anomalies most commonly occur when a product is miscategorised by the marketplace's taxonomy, when a promotional period ends and the promotional commission rate continues to be applied, or when a rate revision is applied retrospectively without advance communication.
The Dispute Filing Workflow
Identifying a discrepancy is only half the reconciliation problem. The other half is converting the identified discrepancy into a filed dispute that the marketplace finance team can process and resolve. Manual dispute filing is time-consuming, inconsistently documented, and frequently deprioritised by finance teams under deadline pressure which means that a significant fraction of identified discrepancies are never filed, and the leakage continues even in organisations that have invested in reconciliation processes.Finance AGI generates a structured dispute package for every flagged discrepancy: the transaction ID, the discrepancy type, the expected amount, the received amount, the variance, the supporting documentation (order record, WMS receipt, commission rate table), and the specific marketplace dispute submission format required for that platform. For marketplaces with API-based dispute submission Amazon Vendor Central, Flipkart Seller Hub Finance AGI can submit disputes autonomously, with the finance manager receiving a confirmation and approval request. For marketplaces requiring email or portal submission, Finance AGI generates the complete dispute submission document, pre-filled and ready for human review and send.Finance AGI also tracks dispute status through resolution. Outstanding disputes are followed up automatically at configurable intervals the agent sends a follow-up query through the appropriate channel if a dispute has been open for more than the marketplace's stated resolution SLA without a response. Resolved disputes are reconciled against the brand's accounting records when the credit appears in the next settlement cycle. Unresolved disputes are escalated to the finance manager's attention with a recommended escalation path. The entire dispute lifecycle is managed by the agent, with humans involved only at the decision points that require business judgment.
What Finance AGI Recovers: Representative Numbers
| Discrepancy Category | Typical Rate | Detection by Manual Review | Detection by Finance AGI |
|---|---|---|---|
| Short settlements (commission overcharge) | 0.8–1.4% of GMV | ~30% of instances | >95% of instances |
| Return deductions without WMS receipt | 0.3–0.6% of return volume value | ~20% of instances (sampled review) | >98% of instances |
| Duplicate TDS/commission deductions | 0.1–0.3% of settlement value | ~15% of instances | 100% (mathematical duplicate detection) |
| Commission rate anomalies | 0.2–0.5% of GMV | ~10% of instances | >90% of instances |
| Outstanding disputes never filed | 35–60% of identified discrepancies | N/A (human deprioritisation) | 0% all identified discrepancies queued for filing |