Many AI programs fail after a promising pilot.
Why is that?
You need to travel further up the data stream.
Many lending workflows have been automated. But it is still common for analysts to perform manual spreading, reconciliation, borrower follow-up, and validation. It is the Emperor’s New Spreadsheets — automation in appearance rather than practice.
Ask yourself the following question: Is your borrower financial data simple to access, standardized to use, and robust to scrutiny? The answer will determine the strength of your data foundations.
In reality, incomplete coverage, inconsistent inputs, exception-heavy workflows, manual interventions, and retrospective remediation are often hallmarks of operational failure. Even with increased digitization, processes remain hamstrung if staff must intervene whenever data is late, in the wrong format, or lacks sufficient provenance and control.
It is a waste all ways round.
This is increasingly important because AI is part of a broader modernization in commercial lending, where direct accounting and ERP connectivity is becoming a standard expectation. Larger banks are already adopting connected borrower-data flows. Those that lag are operating with slower, less scalable models in a market that now expects connected, reusable financial data.
Look at financial spreading. In reported coverage, Citi described a particular bottleneck: commercial borrower financials come from disparate sources in diverse formats and need to be converted from unstructured to structured data to support greater uniformity, analytics, and repeatability. This suggests that the value of AI in underwriting depends heavily on whether the data first has to be made usable. And even when automation improves turnaround time, the workflow still requires structuring messy inputs and validating the results.
Outside of lending, the OCC consent order on Bank of America’s transaction-monitoring program shows that the problem was not the absence of automation. It involved weak governance around inputs, thresholds, filters, and coverage. The order required independent validation of data inputs, recognized that some activity could fall outside automated monitoring, and required manual processes to catch what the system did not reliably capture. It also drove look-backs to determine whether suspicious activity had been missed and whether earlier filings needed correction. We do not see a dramatic system collapse, but incomplete coverage and expensive remediation after the fact.
Supervisory and research sources point in the same direction: if data cannot be obtained efficiently, normalized consistently, and defended during reviews or audits, automation is more likely to remain partial and exception-prone.
Better models cannot compensate for weak foundations in borrower data.
That is why Validis supports lending platforms, underwriting tools, and analytics systems by enabling banks to access borrower financial data directly from accounting and ERP systems, standardize it, and move it through a controlled, repeatable process in a fully secure environment that meets SOC2 Type II and ISO 27001 standards.
