AI aids but can't replace
- A widely shared thread says AI expands underwriting diligence breadth but cannot substitute on‑site depth. - The post stresses human verification, like site visits, remains critical for nuanced risk judgements. - The thread argues underwriters will prefer decision‑support AI with traceable inputs rather than opaque black‑box automation (x.com).
The underwriting pitch for artificial intelligence is speed and scale, but the current argument inside insurance and lending is narrower: let the software widen the search, and let humans make the call after checking the real world. (verisk.com) Underwriting is the job of deciding whether to insure or finance a risk, and on what terms. In practice, that means pulling financials, loss histories, inspection reports, images, and outside data into one file and turning that file into a yes, no, or price. (federalreserve.gov) The case for AI is straightforward: carriers and lenders already use models to process large data sets, and newer generative systems can summarize submissions, flag missing documents, and spot anomalies faster than a person reading every page. Verisk said some carriers using GenAI in early underwriting tasks have seen 30% to 50% time savings. (verisk.com) The limit is that underwriting is not only a document exercise. Deloitte said multimodal models can analyze thousands of images and reduce some manual inspections, but firms still face questions about compliance, transparency, and fairness when those tools shape risk decisions. (deloitte.com) That is why the industry’s preferred model is increasingly “human in the loop,” a plain term for software that assists while a person keeps authority over the decision. Verisk’s October 23, 2025 write-up of an Insurtech New York panel said GenAI can act as a trusted assistant, but “decisions remain in human hands.” (verisk.com) Regulators are pushing in the same direction. The National Association of Insurance Commissioners adopted its AI model bulletin on December 4, 2023, and by April 30, 2024, 11 jurisdictions had adopted or issued related bulletins, including Alaska, Connecticut, Illinois, Maryland, Pennsylvania, and Washington. (content.naic.org) Bank regulators have long required the same basic discipline for models used in risk decisions. Federal Reserve and Office of the Comptroller of the Currency guidance says banks should use active model risk management, with validation, governance, and controls for quantitative systems that inform business decisions. (federalreserve.gov) The technical fight is less about whether to use AI than what kind to use. The National Institute of Standards and Technology says its AI Risk Management Framework is meant to build “trustworthiness” into design, development, use, and evaluation, which pushes firms toward systems with documented inputs, testing, and review trails. (nist.gov) Europe has drawn an even sharper line for some insurance uses. The European Insurance and Occupational Pensions Authority said the European Union’s AI Act treats AI used for risk assessment and pricing for natural persons in life and health insurance as “high-risk,” triggering requirements on data governance, transparency, and impact assessment. (eiopa.europa.eu) That leaves site visits, broker conversations, and other on-the-ground checks in a stronger position than the automation pitch sometimes suggests. The emerging workflow is not a black box replacing the underwriter; it is software doing more of the reading so the underwriter can spend more time verifying what the file cannot show. (verisk.com)
