How Upstart's AI Underwriting Vision Exposes What MCA Funders Actually Need

Upstart's AI Vision Sounds Great, Until You Apply It to MCA

A recent industry response published on deBanked laid bare a tension that has been simmering in alternative lending for years. Upstart's CEO has championed a vision of AI underwriting that replaces traditional credit scoring with machine learning models trained on thousands of consumer data points: education history, employment patterns, behavioral signals. The premise is compelling. The problem is that it was never designed for merchant cash advance.

AI underwriting for merchant cash advance operates in a fundamentally different universe. MCA funders do not care about a business owner's alma mater or how long they held their last W-2 job. They care about daily deposit consistency, average bank balances, existing position stacking, and whether the three months of bank statements in front of them are even real. Upstart's model optimizes for predicting consumer default probability. MCA underwriting optimizes for predicting whether a merchant's future receivables can support a specific purchase amount and holdback percentage.

This distinction matters more than ever in 2026, as MCA funders face simultaneous pressure from institutional capital partners demanding audit-ready files, regulators circling the industry with new disclosure requirements, and a fraud landscape that grows more sophisticated by the quarter. The question is not whether AI belongs in MCA underwriting. It absolutely does. The question is what kind of AI, trained on what data, solving which specific problems.

Why Consumer AI Models Fail in MCA Underwriting

Credit Scores and Alternative Data Miss the Point

Upstart's core innovation was proving that variables beyond FICO scores could predict consumer loan repayment. That was a genuine breakthrough for personal lending. But MCA is not a loan. It is a purchase of future receivables, and repayment capacity is tied directly to business cash flow, not the owner's personal creditworthiness.

A merchant with a 580 FICO score running a restaurant that deposits $4,000 daily is a fundamentally different risk profile than a merchant with a 720 score whose deposits are erratic and declining. Consumer AI models trained on credit bureau data, employment stability, and education cannot distinguish between these two scenarios because they were never designed to parse bank statement transaction data at the line-item level.

Cash Flow Pattern Recognition Requires MCA-Specific Training

Genuine AI underwriting for MCA demands models trained on merchant banking data: deposit frequency, average daily balances, NSF counts, existing ACH debits from other funders, seasonal revenue fluctuations, and the ratio of cash deposits to card processing volume. These signals are invisible to consumer lending AI.

Purpose-built models can identify patterns that human underwriters might miss across hundreds of pages of statements. A gradual shift from card deposits to cash deposits, for instance, might signal that a merchant is diverting receivables. A sudden spike in large round-number deposits before an application could indicate manufactured cash flow. These are MCA-specific fraud indicators that require MCA-specific training data.

As we explored in our analysis of purpose-built AI models outperforming general LLMs in MCA document verification, the gap between generic AI and domain-trained AI is not marginal. It is the difference between catching a fabricated statement and funding a fraudulent deal.

The Document Verification Gap Upstream

Even the most sophisticated AI decisioning engine is only as good as the data feeding it. This is perhaps the most overlooked flaw in the Upstart-style vision: it assumes clean, structured, verified data arrives at the model's doorstep. In MCA, that assumption is dangerously wrong.

Bank statements arrive as PDFs emailed from brokers, photographed on phones, downloaded from online banking portals in wildly different formats, or sometimes stitched together from multiple accounts. Before any AI model can analyze cash flow, someone or something needs to extract the data accurately, verify the document's authenticity, and flag inconsistencies.

This is where the real bottleneck lives. Deals do not stall because an AI model takes too long to score a merchant. They stall because the bank statements never arrived, arrived incomplete, or arrived in a format that requires 45 minutes of manual data entry before anyone can even begin underwriting. Let's Submit was built to solve exactly this problem: a secure upload link captures documents asynchronously, AI-powered extraction pulls key financial data automatically, and the underwriting team receives clean, structured information ready for review.

What AI Underwriting for MCA Actually Needs to Do

Detect MCA-Specific Fraud, Not Just Generic Anomalies

Fraud in MCA lending looks nothing like fraud in consumer lending. Stacking, where a merchant takes advances from multiple funders simultaneously without disclosure, is the industry's most persistent problem. Detecting it requires analyzing bank statements for ACH debits from known funder names, identifying multiple daily debits that suggest undisclosed positions, and cross-referencing application data against known stacking patterns.

Synthetic identity fraud has also evolved. Fraudsters now create business entities with manufactured banking histories, depositing funds between controlled accounts to simulate legitimate revenue. General-purpose anomaly detection will flag obvious outliers, but it takes AI fraud detection specifically trained on business lending patterns to catch the subtler schemes that slip through.

Settlement fraud adds another layer. Merchants enter debt settlement arrangements and stop honoring their MCA obligations, but the signs often appear in bank statements weeks before a default: declining balances, new legal fee payments, and deposits redirected to settlement accounts. AI models trained on these patterns can provide early warning signals that save funders significant losses.

Pair Automation With Human Judgment

The most dangerous version of the Upstart vision is the idea that AI can fully automate the underwriting decision. In consumer lending, where products are standardized and regulated, a fully automated decision pipeline is feasible. In MCA, where deal structures vary widely, where merchants operate in dozens of verticals with different cash flow profiles, and where the legal landscape is shifting rapidly, removing human judgment entirely is reckless.

The right model is augmented underwriting. AI handles the heavy lifting: extracting data from documents, categorizing transactions, flagging anomalies, calculating key metrics like average daily balance and deposit consistency. The human underwriter then reviews the AI's output, applies contextual judgment, and makes the final call. This approach is faster than manual underwriting by an order of magnitude, but it preserves the expertise that catches edge cases AI models have never seen.

Platforms like Let's Submit facilitate this workflow. Documents upload asynchronously through a secure link. AI extracts business info, financials, and owner details. The underwriting team reviews and edits the extracted data on a dashboard with full visibility into every application's status. The AI accelerates; the human decides.

Build Compliance-Ready Audit Trails From Day One

Institutional investors and regulators increasingly demand documentation of how underwriting decisions were made. As we covered in our piece on how investment-grade capital raises the stakes for MCA bank statement verification, the funders attracting serious capital are the ones who can demonstrate systematic, auditable processes.

AI that operates as a black box creates compliance risk. If a funder cannot explain why a deal was approved or declined, they face exposure on multiple fronts: regulatory scrutiny, investor due diligence failures, and potential litigation from merchants or brokers. Every AI system used in MCA underwriting should produce an audit trail that documents what data was analyzed, what the model flagged, and what the human reviewer decided.

Where This Plays Out in Practice

Consider a mid-sized MCA funder processing 200 applications per week through a mix of ISO broker submissions and direct merchant inquiries. Without AI-assisted verification, each application requires manual review of three to six months of bank statements, typically 50 to 150 pages per deal. An underwriter spends 30 to 45 minutes per file on data entry alone, before any analysis begins. At 200 deals per week, that is roughly 150 hours of pure data entry, or nearly four full-time employees doing nothing but copying numbers from PDFs into spreadsheets.

Now layer in the Upstart-style promise: plug in an AI model and let it score every deal automatically. The problem is immediately apparent. The AI model needs structured input data. Who is preparing that data? If the answer is still those four full-time employees manually entering figures, the AI has not solved the bottleneck. It has simply added a faster step after the slowest one.

The actual solution requires AI at every stage. Document collection needs to be asynchronous and automated so merchants upload directly through a secure portal rather than emailing PDFs that sit in an inbox. Extraction needs to happen at the point of upload, with AI parsing bank statements, applications, and ID documents into structured data fields. Only then can an AI decisioning model receive clean inputs and produce reliable outputs.

This is why the debate about AI underwriting for MCA often misses the forest for the trees. The industry does not need a better scoring model nearly as much as it needs better data pipelines. The funder who can collect, verify, extract, and analyze documents in a single workflow will outperform the funder running a sophisticated model on messy inputs every time.

Frequently Asked Questions

How does AI underwriting for MCA differ from consumer lending AI?

AI underwriting for merchant cash advance focuses on business bank statement data, daily deposit patterns, existing funder positions, and cash flow consistency rather than personal credit scores or employment history. MCA AI models must be trained on merchant banking data and MCA-specific fraud patterns like stacking and synthetic deposits. Consumer lending AI models like Upstart's use alternative credit variables that have little predictive value for future receivables performance.

Can AI fully replace human underwriters in MCA?

Not reliably. AI excels at data extraction, transaction categorization, anomaly detection, and calculating financial metrics from bank statements. However, MCA deals involve contextual factors like industry-specific cash flow norms, evolving regulatory requirements, and non-standard deal structures that require experienced human judgment. The most effective approach is augmented underwriting, where AI handles data processing and humans make final decisions.

What types of fraud can AI detect in MCA bank statements?

Purpose-built AI models can detect fabricated or altered bank statements, stacking from undisclosed positions via ACH debit analysis, synthetic cash flow generated through round-tripping deposits between controlled accounts, and early indicators of merchant distress or debt settlement activity. Detection accuracy depends heavily on whether the AI was trained on MCA-specific data rather than general financial documents.

Why is document collection the real bottleneck in MCA underwriting?

Most MCA deals stall not at the decisioning stage but during document intake. Bank statements arrive in inconsistent formats, often incomplete, through email threads that require manual sorting. Even the best AI scoring model cannot function without clean, verified input data. Asynchronous document collection through secure upload portals, combined with AI-powered extraction at the point of upload, eliminates this bottleneck and ensures underwriting teams receive structured data immediately.

Conclusion

The debate triggered by Upstart's AI underwriting vision is healthy for the industry. It forces MCA funders to ask exactly what kind of AI they need, and where in the workflow it matters most. The answer is not a consumer credit model repackaged for commercial use. It is purpose-built AI that handles document verification, bank statement extraction, and fraud detection at the specific points where MCA deals break down.

For funders serious about building scalable, audit-ready operations, the starting point is fixing the intake pipeline. Visit letssubmit.ca to see how asynchronous document collection and AI-powered extraction turn chaotic application intake into a streamlined workflow, giving your underwriting team clean data and the confidence to make faster, better decisions.