AI Statement Analysis: Why Every Figure Should Cite Its Page
A number without a source is a number you have to take on faith. In underwriting, faith is not a control. Citations turn AI output into something you can check.
Picture the underwriter's reply: "Your summary says $47,200 in monthly revenue. Where does that come from?" If the answer is "the software said so", you have a problem. If the answer is "here are the deposit lines, page two through four of the April statement, minus this transfer on page three", you have a conversation.
Automated statement analysis has made reading a stack of PDFs far faster. It can pull deposits, balances, NSFs and recurring debits in seconds where a person might take half an hour. But speed creates a new risk. When a machine produces a clean summary, it is tempting to trust it without looking. And when it is wrong, it is often wrong quietly.
The fix is a simple design principle: every figure the system produces should point back to the exact place on the statement it came from.
Pre-flight summary
- Automated extraction is fast but can be quietly wrong on messy statements.
- Linking each figure to its source page makes verification quick and routine.
- Treat AI output as a first draft that a person reviews and corrects.
- Citations plus an audit trail make a file explainable months later.
- Judge analysis tools on accountability, not only speed.
Leg 01Where automated extraction goes wrong
Bank statements are not standardised. Layouts differ between banks and change over time. Columns wrap, tables split across pages, and descriptors are truncated. Scanned or photographed statements add their own noise. Any extraction system, however good, will occasionally misread something.
The errors are rarely dramatic. More often a credit is misclassified, a page is skipped, a running balance is read as a deposit, or an internal transfer is counted as revenue. Each of these can move a revenue estimate by thousands of dollars, and none of them announces itself in a tidy summary table.
- A deposit line split across two pages counted twice or not at all
- An internal transfer classified as customer revenue
- A recurring debit missed because its descriptor changed
- A returned-item fee counted as an NSF event twice
- A scanned page read with a misplaced decimal point
Leg 02What a citation actually changes
When each figure links to the page and line it came from, checking it becomes a click instead of a search. You can verify a suspicious number in seconds, and you can spot-check a few figures on every file without redoing the whole analysis.
Citations also change behaviour. When reviewers know they can see the source, they actually look. A summary with no sources invites a skim; a summary where every number is one click from the evidence invites verification. That is the difference between an assistant and an oracle.
Every number should answer the question: where did you get that?
Leg 03Human review is the point, not a formality
Cited figures do not remove the need for a person to review the file. They make that review practical. The analysis does the tedious extraction; the person brings judgement: is this large credit really a customer payment, or the merchant moving money from savings? Is this daily debit an advance, or an equipment lease?
A healthy workflow treats the automated output as a first draft. The broker or underwriter reviews it, corrects what is wrong, adds context the statements cannot show, and only then relies on it. If a system nudges people to accept its output without looking, it is creating risk rather than reducing it.
- Run the analysis and read the summary.
- Click through on any figure that looks surprising.
- Spot-check a sample of routine figures, too.
- Correct misclassifications and record why.
- Add the merchant's explanations for anything unusual.
- Confirm the reviewed figures before anything is submitted.
Leg 04Citations and the audit trail
Funding files get revisited. A lender asks a question weeks after funding. A renewal comes up a year later. A dispute arises. In each case, someone needs to understand how the original numbers were reached.
A figure with a citation, plus a record of who reviewed it and what they changed, answers that question without guesswork. Combine that with a full audit trail of document uploads, views and edits, and you have a file that can be explained to a lender, a manager or anyone else with a legitimate reason to ask. That is good practice for a well-run floor, and it is far easier to build as you go than to reconstruct later.
Leg 05Questions to ask any analysis tool
If you are evaluating automated statement analysis, ask how it handles accountability, not only how fast it is. A fast answer you cannot check is not much help when a lender pushes back.
- Does every figure link to the specific page it came from?
- Can reviewers correct a figure, and is the correction recorded?
- Does it show existing positions, NSFs and balances with sources, not just totals?
- Who can see the statements and the analysis, and is access logged?
- Where is the data processed, and can it run privately if you need it to?
Leg 06How FundPilot approaches it
This principle is built into FundPilot's statement underwriting. Deposits, balances, NSFs and existing positions are extracted from the statements, and each figure is linked to the source page so a reviewer can verify it immediately. The results sit on the deal record, alongside the documents, call transcripts and notes, inside an isolated workspace with role-based access and a full audit trail.
The same philosophy runs through FundPilot's AI assistant, which answers from live CRM data and drafts applications, emails and texts but never sends or saves anything without a person confirming. For firms that want it, the AI can run privately on private hardware. The idea is consistent: the software does the reading, and people stay responsible for the decision.
Questions from the floor
Does AI statement analysis replace an underwriter?
No. It removes much of the manual extraction work, which frees people to spend time on judgement: classifying ambiguous credits, understanding existing positions and talking to the merchant. Funders still make their own decisions, and brokers remain responsible for what they submit. Think of it as a fast first draft that still needs a qualified person to review and confirm it.
How accurate is automated statement extraction?
Accuracy depends on statement quality, bank formats and the system used, so be wary of any blanket claim. Clean digital statements usually extract better than scans or photos. Rather than relying on a promised accuracy figure, build verification into your process: check surprising numbers against their source page and spot-check routine ones on every file.
Is it safe to put merchant statements into an AI system?
It depends on how the system handles data. Ask where statements are processed and stored, who can access them, whether access is logged, and whether data is used for any other purpose. Options such as isolated workspaces and private AI on private hardware reduce exposure. Your privacy obligations vary by jurisdiction, so confirm your requirements with qualified counsel.
General information only — not legal, tax or financial advice. Rules for telemarketing, texting, recording, disclosures and data retention vary and change; confirm current requirements with qualified counsel.
Connecting flights