What Bank Statement Analysis Is (and When Lenders Should Automate It)

About the author

AI and decision-systems operator with experience building large-scale systems at Uber and Cruise.

Bank statement analysis is the credit work of reading a borrower's deposits, withdrawals, and balances to decide whether the business can carry a new payment. For SMB lenders it is often the primary file, not a side check. Tax returns are a year late. The P&L was built for the application. The last 90 days of actual cash are sitting in a PDF.

If you still do this by hand, you already know the job. This page names it, says what it is not, and marks when it should stay on an analyst's desk versus when it should be automated. The step-by-step method lives in Bank statement analysis for lenders. This is the definition and the decision, not that walkthrough.

What bank statement analysis is

In lending, bank statement analysis is a cash-flow underwrite from transaction history. You are not reconciling the account. You are deciding whether the deposit stream is real operating revenue, whether existing debt already claims that cash, and whether the balances survive a slow week after a new payment is added.

A $75,000 equipment note at 9.5% for 48 months is about $1,880 a month. The question is not whether deposits exceeded $1,880. The question is whether verified operation revenue, after owner draws and existing funder debits, can absorb $1,880 without the account going negative the week a customer pays late. That is credit. Adding credits and subtracting debits is arithmetic.

Analysts who still work statements by hand are usually doing five things: stripping non-revenue credits, mapping the true debt burden, reading average daily balance and NSF behavior, checking the PDF against the rest of the package, and writing a finding a committee can use. If your process stops at a credit-minus-debit total, you are not doing bank statement analysis.

What it is not

Three things get sold as bank statement analysis. None of them is the credit job.

Not reconciliation

Reconciliation asks whether the book's cash matches the bank. Underwriting asks whether the bank's cash can repay you. A perfectly reconciled set of books can still be a decline if the only large credits are MCA proceeds and owner transfers. A clean tie-out is not repayment capacity.

Not personal finance

Consumer tools group spend into groceries and gas. An SMB credit file needs industry-aware labels: card settlements, related-party wires, funder ACH, owner sweeps. A Zelle from a regular at a restaurant is often revenue. A Zelle from the owner of a trucking company is often not. The label depends on the business, not the payment rail.

Not an OCR dump

Extracting every line into a spreadsheet is useful input. It is not analysis. If a tool hands you 1,400 rows and no opinion on which deposits are operating revenue, your analyst will still spend 40 minutes classifying them. That is the work you were trying to get off the desk.

What the analyst is actually looking for

The useful output is a short finding, not a highlighted PDF. On a typical six-month package you want verified monthly operating revenue, average daily balance, NSF and negative-balance days, existing MCA or daily ACH, owner draws, and any reason the PDF itself is not trustworthy.

How to run that review line by line is covered in the automation and method piece. A committee can use "verified revenue about $31,000, one daily ACH at $287, ADB compressed to $4,100" in two minutes. It cannot use a transcript of 1,100 lines.

A $75,000 equipment file, worked by hand

Take a six-year HVAC shop in Phoenix applying for $75,000 to buy a service van and a recovery machine. The broker package has an application, an invoice, a voided check, six months of Wells Fargo operating statements (May through October), and last year's tax return showing $412,000 gross receipts.

A competent analyst spends about 45 minutes. Months one and two look clean: average deposits around $38,400, ADB around $18,000, two NSFs in May and none after. July then shows a $28,000 credit from Kabbage. August shows a $22,000 credit labeled Lendr. Those are advances, not customers. Strip them and verified operating revenue is closer to $31,000 a month, not $38,000.

On the debit side, a $287 daily ACH to a funder starts in late July and does not stop. That is about $6,000 a month already leaving. Owner draws run $4,200 to $5,800. After those two, the new $1,880 equipment payment is tight. ADB in September drops to $4,100.

The tax return still says $412,000. The last 90 days of the bank say the shop took two advances and is paying one of them every business day. The committee needs: verified revenue about $31,000, existing daily ACH about $6,000, ADB compressed, two early NSFs, PDF looks clean, recommend structure or decline. A junior who treats the Kabbage and Lendr credits as revenue will approve a file the senior will kill. That single classification error is why this work has a name.

When manual still wins

Keep a human on the pages when the file is the exception, not the queue.

• The industry is new to your shop and you do not yet trust anyone else's labels. First cannabis-adjacent file. First owner-operator fleet paid in cash and Zelle. First seasonal contractor with four entities.
• The PDFs are ugly: photographed at an angle, cropped, password-protected, or three accounts stitched into one scan.
• Related-party mess. A spouse account, a payroll company, and the operating account weaving transfers that look like sales.
• Ticket size or exception policy requires a senior to sit with the statements. Anything over your auto-decision cap. A decline the salesperson wants overturned.
• You are writing or rewriting policy and need to see raw behavior, not a summary.

Manual review is slow and inconsistent. It is still the right tool on those files. Automating a one-off you do not understand will give you a confident number you cannot defend in committee.

When lenders should automate

Automate when the work is the same work, repeated. If the desk sees more than a handful of statement packages a day, at 30 to 60 minutes each, you are buying consistency and an exam trail as much as speed: the same rules on every deal, every number tied back to a page, NSF counts that do not depend on who had the file on Thursday.

The working model is hybrid. Software classifies and flags. The analyst decides the exceptions. That is true whether your rules live in a spreadsheet or in an underwriting OS.

Kaaj treats bank statement analysis as one agent inside a full package: statements plus SOS, invoices, and a source-linked memo, landing in the LOS or CRM you already use. No rip and replace. Live in about two weeks. SOC 2 Type II. The statement is never the whole file, and the decision is never the model's.

If you want the method and the output fields, read how bank statement analysis works and how to automate it. If you are choosing vendors, use the AI bank statement analysis tools comparison. Do not buy OCR and call it underwriting.

FAQs

How is bank statement analysis different from cash-flow spreading?

Spreading puts numbers into a template so you can compute DSCR. Statement analysis is how you decide which numbers are real before they go in the template. Spread a file that treated MCA proceeds as sales and the ratio is fiction.

How many months of statements do lenders actually need?

Three months is a snapshot. Six months is the default for most SMB equipment and working-capital files. Twelve months if the business is seasonal: landscaping, HVAC, tax prep, retail. Take every account the borrower actually uses, not just the one on the application.

Can you underwrite from statements alone if tax returns are stale?

Often yes, and many MCA, broker, and small-ticket equipment shops already do. Statements will not replace tax returns for every policy. They will tell you what happened after the return was filed. If the return is 14 months old and the bank shows two new funders, believe the bank.

When should a credit analyst stop reviewing statements by hand?

When the queue is larger than the exceptions. If every file is a one-off, stay manual. If you are running 20 similar $50,000 to $150,000 packages a week and the misses are classification and stacking, automate the grind and keep the analyst on the flags.

Ready to see Kaaj in action?

Book a demo and walk through a live deal with our team — from intake to credit memo.