How David Roitblat's AI Underwriting Critique Reveals What MCA Funders Actually Need

The Industry Debate That Exposes AI's Blind Spot in MCA

When David Roitblat, founder and CEO of AI My Advance, published his pointed response to Upstart CEO Dave Girouard's vision for AI underwriting for merchant cash advance, it struck a nerve across the alternative lending community. Writing on deBanked, Roitblat argued that Upstart's consumer-centric AI framework fundamentally misunderstands the mechanics of merchant cash advance underwriting. His critique isn't an anti-AI position. It's a precision argument: the wrong AI, trained on the wrong data, applied to the wrong product structure, creates more risk than it eliminates.

This matters in 2026 because the MCA industry sits at an inflection point. Capital is flowing in from investment-grade sources. Regulatory scrutiny is tightening. And every funder, from boutique shops to platform-scale operations, is being told that AI will solve their underwriting challenges. Roitblat's rebuttal forces a harder question: solve them how, exactly?

For funders evaluating their technology stack, this debate isn't academic. It directly shapes how you should think about where AI adds value, where it falls short, and what upstream infrastructure you need before any AI model can deliver reliable results.

Why Consumer Lending AI Breaks Down in MCA Underwriting

The Structural Mismatch Between MCA and Consumer Credit

Upstart built its reputation on using machine learning to predict consumer loan defaults more accurately than traditional FICO-based models. The approach works well for personal loans with fixed terms, predictable payment schedules, and borrowers whose income arrives in regular biweekly deposits. None of those conditions hold in merchant cash advance.

MCA deals are structured as purchases of future receivables. Repayment is variable, tied to daily or weekly revenue splits. A restaurant's Tuesday looks nothing like its Saturday. A contractor's cash flow swings wildly between project milestones. The data patterns that make consumer lending AI effective, stable income cadences, consistent debt-to-income ratios, and predictable utilization curves, simply don't exist in the MCA borrower population.

Roitblat's core argument centers on this structural mismatch. An AI model trained on consumer credit data will misclassify risk when applied to a business whose bank statements show erratic deposits, large lump-sum payments, and seasonal revenue gaps. These aren't signs of distress in MCA. They're normal operating patterns that require industry-specific interpretation.

Bank Statement Data Quality Is the Real Constraint

Even the most sophisticated AI underwriting engine is only as good as the data feeding it. This is where Roitblat's critique connects to a problem every MCA funder lives with daily: the bank verification bottleneck.

Before any model can score a deal, someone has to collect three to six months of bank statements from the merchant. Those statements arrive as PDFs via email, through broker submissions, or through a patchwork of upload links. They come in different formats from hundreds of different banks. Pages are missing. Files are corrupted. Statements from multiple accounts get mixed together. Sometimes the merchant sends screenshots instead of actual PDFs.

No AI decisioning model accounts for this chaos. As we explored in our analysis of what Upstart's AI vision actually means for MCA funders, the gap between what AI promises and what it delivers starts at the document intake layer. If your team spends 30 minutes chasing missing pages and re-keying data before a deal even reaches the underwriting queue, the speed advantage of automated decisioning evaporates.

Stacking Detection Requires MCA-Specific Context

One of the highest-value applications of AI in MCA underwriting is detecting stacking, the practice of merchants taking multiple cash advances simultaneously without disclosure. Generic fraud models flag unusual transaction volumes or new account openings. MCA-specific stacking detection requires something different entirely: recognizing the ACH debit signatures of competing funders buried in bank statement transaction data.

This means the AI needs to identify recurring debits that match known funder names, parse memo fields that often contain abbreviated or coded references, and distinguish between legitimate vendor payments and split-funding deductions. Consumer lending AI has no training data for this. Purpose-built MCA models do, but only when fed clean, complete, accurately extracted bank statement data.

The pattern recognition is useless if the underlying bank statements are incomplete or poorly parsed. As we've discussed in the context of preventing MCA stacking fraud with smarter bank verification, the detection chain starts with document collection, not with the model.

Why Smart Funders Separate Document Intake From AI Decisioning

Roitblat's critique implicitly points to an architectural principle that the most operationally efficient funders already follow: decouple your document intake and verification layer from your underwriting decisioning layer.

When these two functions are bundled into a single vendor's black box, funders lose visibility and control. They can't audit how data was extracted. They can't verify whether the AI's confidence score on a particular field reflects genuine accuracy or interpolation from incomplete data. They can't swap out a decisioning model without also ripping out their entire intake workflow.

The better approach is modular. A purpose-built intake platform handles the messy upstream work of collecting documents from merchants and brokers, extracting structured data from bank statements and applications, validating completeness, and flagging anomalies. Clean, structured data then flows into whatever downstream decisioning engine the funder chooses, whether that's a proprietary scoring model, a third-party AI underwriting tool, or a human underwriter reviewing a pre-populated deal summary.

This is exactly the architecture Let's Submit is built around. The platform handles async document collection through secure upload links and email forwarding, uses AI-powered extraction to pull business information, financials, and owner details from uploaded documents, and presents everything in a reviewable dashboard. The funder's underwriting team, or their AI model, receives clean data ready for analysis. No chasing merchants for missing pages. No re-keying numbers from scanned PDFs.

The Audit Trail Advantage

Separating intake from decisioning also creates a stronger compliance posture. With regulatory pressure increasing across multiple states, as seen in recent deBanked coverage of New York and Connecticut legislative activity, funders need a complete audit trail showing what documents were collected, when, from whom, and what data was extracted before a funding decision was made.

Bundled AI platforms that ingest documents and produce a funding recommendation in a single step make it difficult to demonstrate that each stage was handled properly. A modular approach with a dedicated intake platform provides timestamped records of every document upload, every extraction result, and every manual review, independent of whatever decisioning logic comes afterward.

Speed Without Sacrificing Accuracy

The funders winning deals in competitive scenarios aren't necessarily the ones with the fastest AI models. They're the ones who eliminate dead time in the application pipeline. The biggest time sink isn't running a scoring model. It's waiting for documents, re-requesting incomplete submissions, and manually entering data that should have been captured automatically.

Async bank verification solves this by removing the synchronous dependency between the merchant's availability and the funder's workflow. A merchant receives a secure link, uploads documents on their own time, and the platform processes everything in the background. By the time an underwriter opens the file, extracted data is already populated and ready for review. That speed advantage compounds across every deal in the pipeline.

What This Debate Means for Your Technology Stack

Roitblat's public disagreement with Upstart's vision isn't just industry commentary. It's a practical framework for how MCA funders should evaluate AI tools. Three principles emerge.

First, demand domain specificity. Any AI tool you adopt for underwriting should demonstrate training on MCA-specific data: variable repayment structures, daily balance volatility, split-funding patterns, and industry-specific fraud typologies. Ask vendors what their training data looks like. If the answer is consumer credit data or generic small business lending data, proceed with caution.

Second, control your data pipeline. The quality of your AI outputs depends entirely on the quality of your inputs. Investing in a sophisticated decisioning model while tolerating a manual, error-prone document intake process is like buying a racing engine and mounting it on a bicycle frame. Fix the intake layer first.

Third, maintain human oversight at the decision boundary. Even Roitblat, who builds AI tools for MCA, argues that human judgment remains essential for edge cases. The most effective workflow uses AI to accelerate data extraction, flag anomalies, and pre-score deals, then puts a trained underwriter in front of the final decision with all the relevant data already organized.

Frequently Asked Questions

Why does generic AI underwriting fail for merchant cash advance?

Generic AI underwriting models are trained on consumer lending data, which assumes stable income patterns, fixed repayment terms, and predictable borrower behavior. MCA deals involve variable daily or weekly repayment tied to business revenue, irregular deposit patterns, and risk signals like funder stacking that consumer models don't recognize. Applying a consumer AI model to MCA data produces unreliable risk scores because the underlying assumptions about borrower cash flow don't hold.

How does bank verification improve AI underwriting accuracy for MCA?

Bank verification provides the foundational data that any AI underwriting model needs to function. Clean, complete, accurately extracted bank statement data allows AI models to assess daily balance trends, identify existing funder obligations, calculate true available cash flow, and detect fraud indicators. Without reliable bank verification upstream, even well-designed AI models produce garbage-in, garbage-out results. Automated extraction from platforms like Let's Submit ensures the data reaches the model in a structured, consistent format.

Should MCA funders use AI or human underwriters?

The most effective approach combines both. AI excels at data extraction, pattern detection, anomaly flagging, and pre-scoring deals at scale. Human underwriters excel at interpreting edge cases, applying contextual judgment, and making final funding decisions that account for factors AI models can't quantify. Funders who try to fully automate decisioning without human oversight take on unnecessary risk. Those who rely entirely on manual processes can't compete on speed.

What is async bank verification and why does it matter for MCA lenders?

Async bank verification allows merchants to upload bank statements and other documents on their own schedule through a secure link, rather than requiring real-time coordination with the funder's team. The funder's platform processes and extracts data from uploaded documents in the background. This eliminates the back-and-forth of email-based document collection, reduces time-to-decision, and ensures complete submissions before a deal reaches the underwriting queue.

Conclusion

David Roitblat's critique of Upstart's AI underwriting vision is a wake-up call for MCA funders shopping for technology solutions. The lesson isn't that AI doesn't work. It's that AI only works when it's built for the right problem and fed the right data. For MCA, that means purpose-built models trained on industry-specific patterns, fed by clean bank verification data extracted from complete merchant document packages.

The upstream problem of collecting, parsing, and validating bank statements before they ever reach an underwriting model is where most funders still lose time and accuracy. Let's Submit solves that layer with async document collection, AI-powered extraction, and a review dashboard built specifically for MCA workflows. Visit letssubmit.ca to see how it fits into your stack.