How Square's 10-Year Sub-4% Loss Rate Proves AI Underwriting for Merchant Cash Advance Works

A Decade of Sub-4% Losses Sets a New Benchmark for MCA Underwriting

Square Loans has done something no other online small business lender can claim. For more than ten consecutive years, the company has held loss rates below its 4% target. That data point, highlighted in a recent deBanked report, should stop every MCA funder and ISO broker in their tracks. AI underwriting for merchant cash advance is not a theoretical advantage anymore. Square proved it works at scale, across economic cycles, and with billions in originations.

The natural reaction from independent funders is to dismiss the comparison. Square has embedded commerce data. They see every swipe, every deposit, every refund. Of course their models are better. But that framing misses the real lesson. Square's edge is not just data access. It is the systematic, machine learning-driven approach to turning raw transaction signals into underwriting decisions that improve with every funded deal. And in 2026, the tools to build a similar feedback loop are available to funders of every size, if the data feeding those tools is accurate.

This article breaks down what Square's loss rate actually tells us about AI-driven underwriting, where independent funders fall short, and how closing the verification gap upstream makes downstream models dramatically more reliable.

Why Square's Model Wins on Loss Rates

The First-Party Data Advantage

Square does not ask merchants to upload bank statements. It already knows what they earn. Every transaction processed through a Square terminal feeds directly into the underwriting engine. That means no gaps in coverage, no risk of fabricated documents, and no lag between when revenue is earned and when the lender sees it. The model ingests daily sales volume, average ticket size, refund frequency, seasonal patterns, and even time-of-day transaction clustering. Each of these variables contributes to a risk score that updates continuously.

This is the gold standard for cash flow data depth in MCA underwriting. The model does not rely on a snapshot of four months of bank statements. It watches the business breathe in real time. When a merchant's daily volume drops by 20% over two weeks, the system recalibrates before a human would even notice.

Machine Learning Feedback Loops That Compound Over Time

What makes ten years of sub-4% losses so remarkable is not the starting accuracy. It is the compounding. Every funded deal that performs well or defaults becomes a training example. Square's models are not static scorecards. They are gradient-boosted decision trees and neural network ensembles that retrain on fresh outcome data quarterly, sometimes more frequently. Over a decade, that feedback loop produces a model that has seen virtually every pattern of merchant distress, recovery, and fraud.

Independent funders using manual underwriting or simple rule-based systems do not get this benefit. A senior underwriter might have strong instincts, but those instincts do not scale, and they do not update systematically when market conditions shift. The tariff-driven SMB slowdowns earlier this year caught many funders off guard. Square's models, trained on granular transaction velocity changes, likely flagged exposed merchants weeks before traditional underwriters noticed the softness in bank statement deposits.

The Gap Between Embedded and Independent Lenders

Shopify Capital originated $1.5 billion in Q3 alone, putting it at $4.3 billion for the year. Lightspeed is compressing remittance periods. These platform lenders share Square's structural advantage: first-party commerce data, automated decisioning, and closed-loop repayment via revenue splits. Independent MCA funders face a different reality. They depend on broker-submitted applications, manually uploaded bank statements, and phone conversations to assess risk.

That does not mean independents cannot compete. It means they need to be ruthlessly efficient about the data they do have. And the most critical data source for an independent funder is still the bank statement.

Closing the Verification Gap With AI-Powered Bank Statement Analysis

Upstream Data Quality Determines Downstream Model Accuracy

Here is the uncomfortable truth about AI underwriting for independent funders: the model is only as good as the data it ingests. If your bank statements arrive as blurry phone photos, if they are missing pages, if they have been altered, or if they are from accounts the merchant no longer uses, your underwriting model starts with corrupted inputs. No amount of algorithmic sophistication compensates for garbage in.

This is why the verification and document collection layer matters so much. Before an AI model can score cash flow patterns, someone or something needs to confirm that the statements are authentic, complete, and legible. Traditionally, that meant an underwriter opening each PDF, eyeballing the formatting, cross-referencing totals, and flagging anomalies manually. At volume, this process breaks. Industry data suggests that one in eight SMB loan applications contains suspicious documents. If your team is reviewing hundreds of applications per week, the math is brutal.

Automated bank statement analysis changes the equation. AI-powered extraction tools parse transaction-level data from PDFs and images, categorize deposits and withdrawals, flag inconsistencies in formatting or metadata, and surface red flags like round-number deposits or missing transaction days. When these tools sit upstream of your underwriting model, the data arriving at the decisioning layer is already clean, structured, and validated.

Async Document Collection Compresses the Funding Cycle

Speed matters in MCA. Merchants shopping for capital often go with whoever funds first. But speed without accuracy produces defaults. The challenge is compressing the time between first contact and funded deal without cutting corners on verification.

Async bank verification solves this by decoupling document collection from the sales conversation. Instead of waiting for a callback, emailing back and forth, or chasing a merchant for missing pages, funders can send a secure upload link the moment a lead shows interest. The merchant drops their last four bank statements, a government ID, and a void cheque from their phone. AI extraction pulls the numbers. By the time an underwriter opens the file, the application is already structured.

Let's Submit was built for exactly this workflow. Sabbie, our AI sales rep, texts and calls every lead in your pipeline. When a merchant expresses interest, Sabbie sends a branded upload link on the spot. Bank statements, IDs, and signed applications land in one place. AI parses revenue, deposits, and key fields automatically. Your underwriting team only touches deals that are ready to review, with clean, pre-extracted data.

Fraud Detection Must Happen Before Decisioning, Not After

Square's loss rate success is partly a fraud story. Embedded transaction data makes it extremely difficult for a merchant to fabricate revenue. Independent funders do not have that luxury. They rely on documents the merchant provides, which means fraud detection must be an explicit, automated step in the pipeline.

AI fraud detection for bank statements has matured significantly. Modern tools check for pixel-level inconsistencies, font mismatches, metadata tampering, and statistical anomalies in transaction patterns. As we explored in our analysis of how MCA lenders detect fabricated bank statements with AI document verification, generative AI has made it trivially easy to produce convincing fakes. The countermeasure is equally sophisticated detection, applied automatically to every document before it reaches underwriting.

Funders who bolt fraud detection onto the end of their process, as a post-approval audit, are already too late. The cost of a funded fraudulent deal dwarfs the cost of catching it at intake.

Building Your Own Underwriting Feedback Loop

Square has ten years of labeled outcome data. Most independent funders have some version of this, scattered across spreadsheets, CRMs, and the memory of their senior underwriters. The gap is not the existence of data. It is the structure.

To build a meaningful feedback loop, funders need three things. First, standardized data capture at intake. Every application should produce the same structured fields: average monthly revenue, average daily balance, NSF count over 90 days, time in business, number of existing positions. If your intake process produces inconsistent data depending on who reviewed the file, your model cannot learn. Second, outcome tracking tied to those fields. When a deal defaults or performs, the system needs to connect that outcome back to the intake variables. Third, periodic model retraining. Even simple logistic regression models improve meaningfully when retrained on six months of new outcome data.

The starting point for all three is clean, structured intake data. That is not a modeling problem. It is a document collection and extraction problem. Funders who automate this step, using tools like Let's Submit to collect and parse bank statements before underwriting, give their models the cleanest possible training data from day one.

Consider the practical difference. A funder processing 200 applications per month with manual data entry will have inconsistencies, typos, and missing fields scattered across their dataset. The same funder using AI extraction will have uniform, machine-readable records for every deal. Over 12 months, the second funder's dataset is dramatically more useful for model training, even if both funders use the same underwriting algorithm.

Frequently Asked Questions

How does AI underwriting reduce loss rates for MCA funders?

AI underwriting reduces loss rates by processing more variables, more consistently, than human reviewers can. Machine learning models evaluate transaction patterns, cash flow volatility, seasonal trends, and deposit regularity across thousands of applications, learning from every outcome to improve future predictions. The key is feeding these models clean, structured data. Automated bank statement analysis ensures that the inputs are accurate, complete, and free from manipulation before the model scores the deal.

Can independent MCA funders match the loss rates of platform lenders like Square?

Independent funders will likely not match Square's exact loss rates because they lack first-party transaction data. However, they can close much of the gap by combining AI-powered bank statement extraction, automated fraud detection, and systematic outcome tracking. The largest source of excess losses for independents is not weak algorithms. It is dirty data at intake: missing statements, undetected fabrication, and inconsistent manual review. Fixing the upstream verification layer produces the biggest marginal improvement in portfolio performance.

What is async bank verification for MCA lending?

Async bank verification decouples document collection from live conversations between sales reps and merchants. Instead of gathering bank statements over email or during a phone call, the funder sends a secure upload link that the merchant can complete on their own time, from their phone or computer. AI extraction then parses the uploaded documents automatically. This approach reduces cycle time, eliminates back-and-forth chasing, and ensures documents are collected in a standardized format every time. Let's Submit provides this exact workflow, with branded upload links, AI-powered extraction, and a unified pipeline view.

How do MCA lenders detect fake bank statements in 2026?

Detection relies on multiple layers. AI tools analyze document metadata for signs of editing software, check font consistency across pages, flag statistical anomalies in transaction patterns such as perfectly round deposits or missing weekends, and compare formatting against known templates from major banks. Some systems also cross-reference reported balances against open banking data where available. The most effective approach applies these checks automatically at the point of document upload, before the application reaches an underwriter. As explored in our piece on detecting fabricated cash flow patterns with AI, the sophistication of both fraud and detection tools has escalated rapidly.

Conclusion

Square's decade of sub-4% loss rates is not just a headline. It is a blueprint. The principles behind that performance, clean data, machine learning feedback loops, and fraud detection at intake, are available to every MCA funder willing to invest in their verification infrastructure. Independent funders will never have Square's embedded transaction feed, but they do not need it to dramatically improve portfolio outcomes. They need structured, validated bank statement data flowing into models that learn from every deal.

Let's Submit gives funders the upstream layer that makes everything downstream work better. From AI-powered lead outreach to async document collection to automated extraction, the platform ensures your underwriting team starts with clean data on every file. Visit letssubmit.ca to see how async verification and AI extraction fit into your funding workflow.