FICO was invented in 1989. It was designed to help consumer lenders — primarily credit card issuers and mortgage originators — quickly assess the repayment likelihood of individual borrowers. The model works well for its intended purpose: predicting whether a person will make their next credit card payment, given their history of making or missing prior consumer credit payments.
The problem is that FICO was never designed to evaluate a business. A sole proprietor's personal FICO score says nothing useful about their business's cash generation, their client concentration, their seasonal revenue cycle, or the quality of their receivables. Yet for decades, personal credit scores have been the primary gating mechanism for SMB loan decisions. This is not because it's a good signal — it's because it was the only readily available signal that could be processed quickly.
That constraint no longer exists.
Why credit scores produce systematic errors for SMB lending
The mismatch between FICO as a signal and SMB creditworthiness as a concept runs deep. Consider what FICO actually measures: payment history on consumer credit accounts (roughly 35% of the score), credit utilization on revolving accounts (30%), length of credit history (15%), recent inquiries (10%), and credit mix (10%). Every one of these factors reflects consumer financial behavior, not business operating performance.
An SMB owner who used personal credit cards to float operations during a difficult period — a common and often rational decision — carries a utilization signal that looks negative even after those cards are paid off. An owner who relocated and closed older credit accounts to consolidate their financial profile may have a shortened credit history that scores poorly even though their business is performing well. An owner who took on a personal mortgage to purchase a property that now generates rental income may have a high credit utilization signal from the origination inquiry.
None of these patterns tell a lender anything meaningful about whether the SMB can service a $200,000 working capital line. They are noise in the signal, systematically biased against owners whose personal financial lives don't conform to the consumer credit profile FICO was built to evaluate.
The data pipeline that makes accounting-based underwriting possible
Building an underwriting model on live operational data requires solving three distinct engineering problems: data ingestion, normalization, and signal extraction. Each layer is worth understanding because the quality of the underlying data pipeline determines the quality of the credit decision.
Data ingestion
Accounting platform APIs — QuickBooks Online, Xero, FreshBooks — expose structured access to general ledger data, accounts receivable aging reports, expense categorizations, and invoice records. A well-implemented API integration pulls this data in real time and handles the authentication, rate limits, and schema variations across platforms. Plaid and similar bank data aggregators provide normalized transaction-level access to business bank accounts — typically one to two business days of latency for the most recent transactions, with full history back several years.
The ingestion layer needs to handle incremental updates cleanly. A business's QuickBooks data changes every day as invoices are issued, payments are received, and expenses are recorded. The underwriting model should be reading the current state of the data, not a snapshot from the moment of initial application.
Normalization
Raw accounting data is not analysis-ready. A QuickBooks file from a construction company categorizes expenses differently than one from a food distributor. Revenue recognition practices vary — some businesses use accrual accounting, some use cash basis. A business that uses a mix of personal and business accounts (common among sole proprietors and early-stage businesses) requires deduplication logic to isolate business cash flows from personal intermingling.
Normalization means transforming the raw data into a consistent schema: monthly revenue by source, expense categories mapped to standard classifications, accounts receivable aging buckets (current, 30-60 days, 60-90 days, 90+), bank account balance time series, and payroll cost by period. Without this normalization layer, the model is comparing apples to completely different fruit across different businesses.
Signal extraction
This is where the underwriting model operates. The features that matter for working capital credit decisioning fall into several categories:
- Revenue stability: Coefficient of variation on monthly revenue over the trailing 12 months. A low CoV indicates predictable cash generation. High CoV requires deeper examination of whether the variation is seasonal (explainable) or erratic (concerning).
- Cash flow coverage: A version of DSCR calculated from actual operating cash flows rather than reported net income. This matters because depreciation, owner compensation adjustments, and one-time items can distort net income as a coverage proxy.
- Receivables quality: The proportion of AR in 60+ day aging buckets is a leading indicator of collection risk. A business where 40% of outstanding invoices are over 60 days may have lower effective revenue than their top-line figure suggests.
- Bank velocity and minimum balance behavior: Average daily balance over the trailing 90 days, and the frequency and depth of below-minimum-operating-balance days, tell a story about operating buffer discipline that a monthly P&L cannot reveal.
- Expense concentration: If a single vendor represents 35% of total operating expenses and that vendor relationship shows payment irregularity, that's a risk signal that doesn't appear in any credit bureau report.
What the model gains — and what it still can't see
The predictive improvement from replacing credit scores with operational data is material for a specific population: established SMBs with 12+ months of clean accounting data and connected bank accounts. For this population, cash flow-based underwriting produces a fundamentally more accurate picture of repayment capacity than FICO does, because it's measuring the thing that actually determines repayment — the business's ability to generate operating cash.
This does not mean the model is infallible, and it would be misleading to suggest otherwise. There are several categories of risk that live operational data handles poorly.
Forward-looking business risk — a key contract that's about to expire, a regulatory change that will affect the industry, a major client who has informally signaled they won't renew — doesn't appear in historical accounting data. The model sees excellent trailing cash flows right up until the week the wheels come off. This is why human review remains appropriate for large lines and for businesses showing unusual patterns that warrant qualitative investigation.
Fraud risk is also more complex in data-based underwriting. A sophisticated operator can manipulate accounting records in ways that pass automated checks. Reconciliation against bank transaction data provides a partial check — cash receipt patterns should correlate with invoiced revenue — but determined manipulation of connected data sources is a real attack surface. The model needs anomaly detection layers specifically designed to flag inconsistencies between accounting records and bank feed data.
The business owner's perspective: what this means practically
For an SMB owner applying for a working capital line through a data-based underwriting process, the practical change is significant in two ways.
First, the application itself is different. Instead of gathering two years of tax returns, preparing a business plan narrative, locating personal financial statements, and scheduling a banker meeting, the process involves connecting accounting software and authorizing a bank feed. The information burden shifts from the borrower to the data pipeline. A business with clean books on QuickBooks and a connected business bank account is already 80% of the way through their application before they answer a single question.
Second, the decision reflects current reality. A business that had a rough 2023 but recovered strongly in 2024 is evaluated primarily on 2024 and 2025 performance, not on the aggregate picture that a tax return summary produces. The model sees the recovery, not just the dip.
For the significant portion of SMB owners whose businesses perform better than their personal credit files suggest, this shift is the difference between access and exclusion. The data doesn't lie — it just requires a system capable of reading it.
Where this goes from here
The infrastructure for real-time operational data underwriting is still maturing. API access to accounting platforms has improved substantially over the past five years. Bank feed coverage via aggregators has expanded. But there remain gaps: businesses using legacy accounting software without API access, businesses in industries with complex revenue recognition patterns, and businesses where the separation between personal and business finances is minimal.
The direction of travel is clear. Every improvement in data connectivity — broader accounting platform coverage, faster bank feed latency, better transaction categorization — directly improves the signal quality available to underwriting models. The credit score will continue to be used as one input among many, particularly for identity verification and fraud detection. But as the primary decision signal for SMB credit? Its reign is ending, not because of ideology, but because better data exists and the infrastructure to use it is operational.