How Upstart's AI Underwriting Vision Misses the Bank Verification Reality for MCA Lenders

Upstart's AI Thesis Sounds Great Until You Underwrite an MCA

David Roitblat's recent critique of Upstart CEO Dave Girouard's AI underwriting vision, published on deBanked, landed a direct hit on a blind spot that MCA professionals have been talking about for years. Girouard's argument is seductive: AI models trained on enough data can outperform any human underwriter. Feed the machine credit bureau data, bank transactions, and alternative signals, and it will price risk better than a team of analysts ever could.

For consumer installment loans with predictable repayment structures, that thesis holds weight. For merchant cash advances, it falls apart. AI underwriting for merchant cash advance deals confronts a fundamentally different challenge. MCA repayment is tied to future revenue, not a borrower's credit score or debt-to-income ratio. The merchant's bank portal, viewed live, reveals patterns that no API feed or PDF statement can fully replicate: daily deposit velocity, seasonal swings, evidence of stacking, and whether the account even belongs to the applicant standing behind the application.

This article breaks down exactly where the pure-AI underwriting thesis fails for MCA funders, where AI genuinely adds value, and how async verification creates the connective tissue between speed and accuracy that the industry needs in 2026.

Where the Pure-AI Underwriting Model Breaks for MCA

MCA Repayment Hinges on Cash Flow, Not Credit Scores

Upstart's model was built on a simple premise: traditional credit scores are incomplete, and machine learning can find signals that FICO misses. That is true. But MCA repayment is not a function of creditworthiness in the traditional sense. A merchant with a 580 FICO score and $40,000 in monthly deposits can be a better MCA candidate than someone with a 720 score and $8,000 in deposits. The underwriting question is not "will this person repay a structured loan" but rather "does this business generate enough daily cash to sustain a percentage-based remittance without defaulting."

AI models trained on consumer lending data do not transfer cleanly to this problem. The features that matter are different. The time horizons are different. The fraud vectors are different.

Bank Statement Manipulation That Models Cannot See

One of the most persistent fraud challenges in MCA underwriting is manipulated bank statements. Fraudsters use PDF editing tools, synthetic transaction generators, and even AI itself to produce statements that look legitimate on paper. A machine learning model analyzing extracted data from these documents is, in effect, analyzing the fraudster's output. If the fabricated numbers are internally consistent, the model has no basis to flag them.

This is precisely why detecting manipulated bank portals during live verification remains essential. Video evidence of a merchant navigating their actual banking portal in real time provides a layer of authenticity that no document analysis pipeline can replicate. You can fake a PDF. Faking a live, interactive session with a bank's web application is orders of magnitude harder.

Stacking Fraud Is Invisible to Isolated AI Models

Stacking, where a merchant takes multiple advances from different funders simultaneously, is arguably the single largest risk factor in MCA. An AI model underwriting in isolation sees one application, one set of bank statements, and one applicant profile. It has no visibility into the three other advances the merchant funded last week through different brokers.

Network-level fraud detection helps, but even the best network models depend on data completeness. When a merchant deliberately routes applications through different brokers, uses slight name variations, or opens secondary accounts to obscure existing obligations, the signal degrades. The bank verification step, where an underwriter reviews the actual account and looks for unfamiliar ACH debits, remains the most reliable stacking check available. As we explored in our analysis of how SMB lending fraud is concentrating in MCA, the verification step is the last line of defense before funding.

Where AI Genuinely Adds Value in MCA Underwriting

Transaction Categorization and Deposit Velocity Analysis

None of this means AI is useless for MCA. Far from it. Machine learning excels at categorizing transactions across hundreds of bank statements per day, flagging anomalous deposit patterns, identifying round-number deposits that suggest fabricated revenue, and scoring cash flow stability across rolling windows. These tasks are tedious, error-prone, and slow when done manually. AI handles them in seconds.

The key distinction is that AI transaction analysis works downstream of verification, not instead of it. You first confirm that the data is real. Then you let the model analyze it. Reversing that order means the model might be analyzing fiction.

AI-Guided Recording and Step Validation

A more practical application of AI for MCA funders is in the verification workflow itself. Exact Balance uses AI-powered step detection to guide applicants through the recording process: confirming that each required screen is captured, that the correct date ranges are shown, and that the session is continuous. This is not AI replacing human judgment. It is AI ensuring that the human reviewer receives complete, usable evidence every time.

This kind of narrow, task-specific AI is where the real ROI lives for MCA operations. It eliminates the most common failure mode of manual verification, which is not fraud but incompleteness: recordings that miss a required page, applicants who scroll too fast, or sessions that cut off before showing the full transaction history.

Fraud Pattern Detection in Pre-Screening

AI also adds genuine value in the pre-screening phase. Machine learning models can flag applications that share device fingerprints with previously flagged merchants, identify broker submission patterns associated with higher default rates, and detect geographic or temporal anomalies that suggest coordinated fraud rings. These signals are useful for prioritization. They tell an underwriter which applications deserve extra scrutiny, which recordings to review first, and where to focus manual investigation.

But flagging is not deciding. An AI flag that says "this application has a 23% probability of being associated with a stacking pattern" is useful input. Automatically declining that application without human review is a liability.

Async Verification Bridges the Speed-vs-Accuracy Gap

The real tension in MCA operations is not human versus AI. It is speed versus accuracy. Funders need to close deals fast because merchants have alternatives, and brokers will route volume to whoever funds first. But rushing the verification step is how fraud losses happen.

Traditional live verification calls attempted to solve both problems at once, and solved neither well. Scheduling a call takes hours or days. The call itself is synchronous, meaning an underwriter is occupied for the duration. Time zone mismatches between Canadian funders and U.S. merchants add further delays. The result is a bottleneck that slows deal velocity and frustrates merchants.

Async verification eliminates the scheduling problem entirely. The merchant records their banking session at their convenience. The underwriter reviews the recording on their own schedule. Neither party waits for the other. The Federal Reserve's latest small business surveys continue to show MCA adoption holding steady at 7% of firms, which means funders are competing fiercely for a defined pool of merchants. Every hour saved in the verification cycle is an hour of competitive advantage.

What makes this approach defensible is that async verification does not sacrifice evidence quality. The recording is timestamped, stored securely, and available for audit. If a deal goes bad, the funder has video proof of exactly what the merchant's banking portal showed at the time of underwriting. No call log or PDF screenshot provides that level of documentation.

Frequently Asked Questions

Can AI fully replace bank verification for MCA underwriting?

No. AI models can analyze transaction data, flag anomalies, and categorize cash flows with speed and consistency that humans cannot match. But they cannot confirm that the underlying data is authentic. Manipulated bank statements with internally consistent numbers will pass model checks. Live bank portal verification, whether synchronous or async, remains the only reliable way to confirm that statements reflect real account activity.

What is async bank verification and how does it work for MCA lenders?

Async bank verification replaces live, scheduled verification calls with browser-based screen recordings. The funder sends the applicant a secure link with custom instructions specifying which accounts and date ranges to show. The applicant records their banking session at their convenience using only a browser, with no software installation required. The funder's underwriting team reviews the recording on demand. Platforms like Exact Balance add AI-guided coaching during the recording to ensure completeness.

How do MCA lenders detect stacking fraud during bank verification?

During a bank verification recording, underwriters look for unfamiliar ACH debits that indicate existing advances from other funders. Daily or weekly debit patterns from known funder names, unexplained large outflows, and multiple recent small-dollar deposits designed to inflate average daily balances are all red flags. AI-powered transaction categorization can pre-flag these patterns, but a trained underwriter reviewing the live portal footage makes the final determination.

Why doesn't Upstart's AI underwriting model work for merchant cash advances?

Upstart's model was designed for consumer installment lending, where credit bureau data, income verification, and employment history are strong predictive features. MCA repayment depends on daily business revenue, which is volatile, seasonal, and difficult to model without real-time bank data. The fraud landscape is also different: MCA fraud relies on manipulated documents and stacking rather than synthetic identities, which means the detection methods need to be different too.

Conclusion

The debate over AI in MCA underwriting is not about whether to use AI. It is about where to use it. Machine learning transaction analysis, fraud pattern detection, and AI-guided verification workflows all deliver measurable value. Fully automated underwriting decisions based on unverified data do not.

The funders who will win in this market are the ones who pair AI speed with verification rigor. They will use models to prioritize, categorize, and flag. They will use async verification to confirm. And they will close deals faster than competitors who are still scheduling phone calls or, worse, skipping verification entirely because their AI told them the numbers looked fine.

Visit exactbalance.ca to see how async bank verification fits into your AI-augmented underwriting workflow. Send your first verification request in minutes, and let your underwriters focus on the decisions that actually require human judgment.