← Agent catalogue
AI Agent Profile · LendingIQ · Bengaluru
Bank Statement Analyst AI
Function: Credit Analyst — Statement ReviewRuntime: AWS Bedrock · ap-south-1Model: Claude Sonnet 4Context window: 200K tokens
DivisionOnboarding
Resume
What this agent does
The Bank Statement Analyst AI reads 12 to 24 months of bank statement data — from the Account Aggregator as the primary source and from OCR-extracted PDF statements as a fallback — and produces a structured financial analysis covering income calculation, EMI and obligation identification, bounce detection and pattern assessment, cash flow scoring, and a plain-language narrative that allows the credit officer to understand the borrower's financial health in under two minutes. It serves the Credit Underwriting Agent AI's decision process and, for referred cases, the human credit officer reviewing the file.
Primary functions
Income Calculation
Every application with statement dataInvoked when: bank statement data — AA feed or PDF OCR — is available for the application
- Reads all credit entries in the statement over the analysis window and classifies them by source type: regular salary credits (consistent amount, consistent timing, from a named employer), business income receipts (variable amounts, multiple counterparties, consistent with declared business type), rental income (regular fixed credits from consistent counterparties), and inward transfers that are not income (intra-account transfers from the borrower's own other accounts, family transfers, refunds, loan disbursements into the account).
- Calculates net monthly income by excluding non-income credits from the gross credit total — because a statement that shows ₹3 lakh in monthly credits but ₹1.2 lakh of those are the EMI bouncing back, ₹0.5 lakh is a family transfer, and ₹0.3 lakh is a mutual fund redemption has a genuine income of approximately ₹1 lakh, not ₹3 lakh. The income calculation shows the breakdown: what was included, what was excluded, and why — so the credit officer can see and verify the classification.
- Computes an average monthly net income across the analysis window and flags the income's consistency — a coefficient of variation above a configured threshold flags income as "variable" and presents the monthly series rather than a single average, so the credit officer can see whether income variability is seasonal (acceptable with context), trending downward (concerning), or erratic (higher risk). A single average that masks a declining trend is misleading; the agent presents the trend, not just the number.
- Compares the calculated income figure against the declared income on the application and against any ITR or GST data available in the file. Where the bank statement income is materially lower than the declared income, it flags the discrepancy with the specific figures — not to conclude that the application is fraudulent, but because the underwriting FOIR must be computed on the verified lower figure, not the declared higher figure.
Output: Income analysis — total credits by type with classification rationale, excluded items with exclusion reason, net monthly income series with average and coefficient of variation, income consistency classification (Stable / Variable / Declining), and discrepancy flag if bank-verified income differs materially from declared income.
EMI Bounce Detection
Every application — full statement windowInvoked simultaneously with income calculation — reads debit entries for obligation and bounce signals
- Identifies all ECS/NACH debit entries in the statement — the regular outgoing debits that represent EMI payments on existing loans — and checks each one for bounce or return entries in the 2–5 business day window following the debit date. An ECS debit on the 5th followed by an ECS return credit on the 7th is a bounce; both entries are required to confirm it. The agent does not flag a missing debit as a bounce — it requires both the attempted debit and the return credit to classify a bounce event.
- Analyses the bounce pattern: a single bounce in 24 months on a single obligation with no recurrence is categorically different from three bounces in 12 months across the same obligation, which is different again from bounces across multiple obligations in the same month (suggesting a liquidity crunch in that specific month). The pattern matters more than the count — and the narrative explains the pattern, not just the number.
- Identifies undisclosed obligations by cross-referencing ECS debit entries against the obligations declared on the loan application. A regular ECS debit to an NBFC or bank that is not listed in the declared obligations section is an undisclosed liability that affects the FOIR calculation — the agent flags it with the counterparty name and estimated monthly obligation amount, and the credit officer decides whether to request clarification before proceeding.
- Estimates the total existing EMI obligation from the statement — combining all identified ECS debits as the observed obligation baseline. This observed FOIR check is more reliable than the declared FOIR because it is based on actual outgoings rather than borrower self-declaration, and it includes obligations the borrower may have omitted from their application. Where the observed FOIR exceeds the declared FOIR by more than a configured tolerance, the discrepancy is flagged for the credit officer's attention.
Output: Bounce detection log — every bounce event with date, obligation, amount, and return confirmation; bounce pattern assessment (isolated / recurring / multi-obligation crunch); undisclosed obligation list with counterparty and estimated EMI; observed FOIR vs declared FOIR with discrepancy flag if material.
Cash Flow Scoring
Every application — synthesised from income and debit analysisInvoked after income calculation and bounce detection are complete — requires both to produce a meaningful cash flow score
- Computes a cash flow health picture across five dimensions: income stability (how consistent is the income month-to-month), surplus adequacy (average monthly surplus after all observed obligations as a multiple of the proposed EMI — is there room in the cash flow for the new EMI?), balance buffer (average month-end balance — does the borrower maintain a meaningful balance or do they run the account to near-zero each month?), obligation discipline (bounce rate as a proportion of total ECS debit events over the window), and balance trend (is the average month-end balance growing, stable, or declining over the analysis window?).
- Produces a composite cash flow score across these five dimensions — not as a single number (which hides the dimension-level picture) but as a five-dimension assessment with a classification per dimension: Strong / Adequate / Marginal / Weak. A borrower with Strong income stability and Strong surplus but Weak obligation discipline presents a different risk profile from one with Moderate income and Adequate everything else — and the credit officer needs to see both profiles clearly, not a single aggregate score that obscures the difference.
- Produces a plain-language cash flow narrative — a 4 to 6 sentence paragraph that tells the story of the borrower's financial health as revealed by their statement, written for the credit officer to read before making the credit decision. The narrative states what the numbers mean, not just what the numbers are: "The account shows consistent salary credits of approximately ₹X per month from [employer]. Existing EMI obligations of ₹Y per month leave a pre-proposed-EMI surplus of ₹Z. A single ECS bounce in month 8 appears isolated rather than structural, coinciding with a period of reduced salary credits — possibly a salary delay rather than income disruption. Month-end balances have been gradually improving over the last 6 months."
- Computes the FOIR headroom for the proposed loan: the proposed EMI as a percentage of the verified average monthly net income, and the FOIR including all observed existing obligations. States clearly whether the proposed loan fits within the credit policy FOIR limit and, where it does not, by what margin — because an application that is 2% over the FOIR limit requires a different response from one that is 15% over.
Output: Five-dimension cash flow assessment — each dimension rated Strong/Adequate/Marginal/Weak with supporting data; FOIR calculation — current observed FOIR and post-proposed-loan FOIR against policy limit; plain-language 4–6 sentence narrative for credit officer; and overall cash flow classification (Strong / Adequate / Marginal / Weak) for underwriting routing.
Knowledge base
Account Aggregator Statement Data
Machine-readable, tamper-proof 12–24 month transaction data from the AA framework. The primary and highest-reliability source. Available only where borrower has consented to AA data sharing at origination.
PDF Statement Data (via Doc Verification AI)
OCR-extracted transaction data from self-submitted PDF bank statements. Lower reliability than AA data — used as fallback when AA consent is absent. All outputs from PDF source are clearly labelled as self-submitted.
Credit Policy Corpus (RAG)
FOIR thresholds by product and segment, EMI bounce tolerance norms, minimum average balance requirements, and income verification standards. Retrieved live — always the current policy version.
Transaction Classification Rules
Rules for classifying credits as income vs non-income and debits as obligations vs one-off payments. Maintained by the credit team. The quality of income and FOIR calculations depends directly on the completeness and accuracy of these rules.
Application Financial Declarations
The borrower's declared income and obligations from the loan application form — used as the baseline for discrepancy detection between declared and bank-verified figures.
Credit Analysis Knowledge
Pre-training knowledge of bank statement analysis methodology, Indian banking transaction patterns, EMI debit structures, NACH/ECS mechanics, and cash flow-based credit assessment practice up to knowledge cutoff.
Hard guardrails
Will notPresent a single income figure without the income trend and consistency classification. A declining income presented only as an average is misleading. The trend is always surfaced alongside the average — not an optional addendum.
Will notClassify cash deposits as income without flagging the lower reliability of cash as an income verification basis. Cash counter credits in a bank statement cannot be traced to a verifiable income source — they are included in the gross credit picture but noted as unverified income with an explicit flag for the credit officer.
Will notSuppress a discrepancy between bank-verified income and declared income even if the discrepancy is unfavourable to the application's prospects. Discrepancies are always reported — the credit officer decides whether to proceed, query, or decline based on the full picture.
Will notVerify the origin of funds. Bank statement analysis can identify that credits exist and what their pattern is — it cannot determine whether the source of funds is legitimate, whether credits represent declared income or undisclosed sources, or whether the account has been used to cycle funds in preparation for a credit application. These questions require additional investigation beyond statement analysis.
Will notProduce an analysis on a statement that covers fewer than 3 months of data. Analyses on shorter windows are not meaningful for income assessment or bounce pattern identification. Applications with fewer than 3 months of statement data are flagged as requiring supplementary income verification — the agent does not extrapolate a 1-month statement to a 12-month picture.
Known limitations
Seasonal businesses produce cash flow patterns that look like income instability to a uniform analysis. A contractor, a mango exporter, a school uniform retailer, or a wedding photographer may show very low bank credits for 4 months of the year and concentrated high credits in the season — which the analysis will correctly show as high income variability without the context that this is normal for the business type. The seasonal pattern limitation is most significant for MSME borrowers and least significant for salaried borrowers.Build a sector-seasonality reference into the transaction classification rules — common business types with predictable seasonal cash flow profiles, so the analysis can contextualise variability as seasonal rather than erratic where the sector evidence supports it. This is a corpus addition, not a model change.
PDF bank statement analysis is subject to the quality limitations of OCR on financial documents. Multi-column table formats, merged cells, non-standard date formats, and scanned images of printed statements all degrade extraction accuracy. The confidence score per extracted transaction is the primary quality signal — but where overall confidence is low, the computed income and bounce figures may be unreliable even where they appear plausible.Where PDF statement OCR confidence falls below the configured threshold for more than 20% of transactions, route the application to Account Aggregator re-consent before proceeding. An analysis based on a heavily degraded PDF is worse than having no analysis — it provides false precision on unreliable numbers.
The agent cannot identify undisclosed obligations that are paid by cash or through a different bank account not submitted in the application. A borrower who services an informal borrowing or a separate credit facility through a second bank account will show a clean ECS debit record in the primary account — the FOIR observed from the primary statement will be understated. Multiple-account borrowers are a structural limitation of single-account analysis.For applications above a defined ticket size, make a multi-bank statement review mandatory — requiring the borrower to provide AA consent or statements from all accounts in their name. This reduces but does not eliminate the multiple-account blind spot, since the borrower must self-declare all accounts for the request to cover them.
The FOIR calculation derived from the statement is an approximation, not a precise measurement. ECS debit amounts include both principal and interest components that cannot be separated from the statement alone; the full sanctioned loan amounts and outstanding balances require bureau data for precision. The statement-derived FOIR is a useful cross-check against the bureau-verified FOIR — the two should be reconciled, and where they diverge materially the reconciliation explains which obligations are captured in each source and which are not.Build a reconciliation note into the standard output where the statement-observed FOIR and the bureau-verified FOIR differ by more than a configured tolerance. The reconciliation makes the source of divergence explicit rather than leaving the credit officer to manually reconcile two figures that should but do not match.
Important Reads
Learn more about how to deploy Bank Statement Analyst AI to your lending workflow.
- Use case #0001Income calculation: how Bank Statement AI handles irregular and gig incomeA salary slip states an income. A bank statement reveals one. The difference matters most for the 40% of loan applicants in India whose income does not arrive as a single monthly credit from a single employer — the gig worker, the consultant, the small business owner, the freelancer whose income arrives in multiple credits of variable amounts from multiple sources. Standard income calculation methods, designed for salaried borrowers, systematically underestimate or misclassify this income. The Bank Statement Analyst AI is built to read it correctly.Read article →
- Use case #0002EMI bounce detection: the bounce patterns that predict future defaultAn EMI bounce in a bank statement is not a number — it is a narrative. A single bounce followed by immediate clearance tells one story. Three bounces in the same calendar month tell another. A pattern of bounces on the 1st of the month that are cleared by the 5th, every month for six months, tells a third. The Bank Statement Analyst AI reads the narrative, not just the count, because the narrative predicts default far more accurately than any individual bounce event.Read article →
- Use case #0003Cash flow scoring: how Bank Statement AI builds a 12-month financial pictureA CIBIL score tells the underwriter what a borrower has done with credit in the past. A cash flow score tells them how the borrower has managed their actual financial life over the last 12 months — the inflows, the outflows, the savings rate, the spending patterns, the financial shocks absorbed, and the trajectory of every metric from month 1 to month 12. It is not a replacement for the credit bureau score. It is the context that makes the bureau score interpretable.Read article →