AI in Insurance Underwriting: Benefits, Risks, and the Critical Role of Human Oversight
By Nick Johnson, Director of Technology
Artificial intelligence is reshaping insurance underwriting in ways that are hard to ignore. Carriers are moving faster, pricing more accurately, and processing more submissions than ever before. As independent agents, we need to understand what AI actually does well, and where it falls short, because both matter to our clients and to our own practices.
I’ve spent time building AI systems. Here is my honest read on where things stand.
What AI Does Well
The benefits of AI in underwriting are genuine, and the industry should embrace them thoughtfully.
Speed and capacity. A skilled underwriter might evaluate 30 to 40 files in a day. AI systems process thousands simultaneously, with no drop in consistency as the day goes on. For high-volume personal lines like auto and homeowners, that throughput changes the economics of the business.
Pattern recognition. When you have claims experience across millions of policies and decades of data, machine learning can surface risk signals that traditional actuarial methods miss. Property conditions, geographic micro-patterns, behavioral indicators; AI finds correlations in that data at a scale no human team could replicate. Done right, better risk discrimination means that responsible policyholders stop subsidizing higher-risk ones.
Consistency. Human underwriters are inconsistent. The same file might get different treatment depending on the day, the workload, or frankly the mood in the office. AI applies the same logic every time. That consistency, when the underlying logic is sound, is a genuine improvement.
Fraud detection. AI is very good at spotting anomalies across large volumes of submissions: inconsistencies in reported information, patterns that correlate with inflated claims, mismatches between stated facts and verifiable records. Fraud costs the industry billions a year and those costs flow through to honest policyholders in the form of higher premiums.
Operational efficiency. Automated data prefill, straight-through processing for straightforward risks, real-time integration of third-party data; these reduce cost and error and free underwriters to spend their time on accounts that actually require judgment.
The Bias Problem Is More Complicated Than You’ve Heard
Most discussions of AI bias in underwriting frame it as a data quality problem. Feed a model biased training data and you get biased outputs. Clean up the data and you’re fine.
That framing is incomplete, and it matters that we understand why.
Historical data carries historical decisions. When you train an AI model on decades of underwriting decisions and claims outcomes, you are teaching it to replicate the patterns in that history. That history includes a long period when redlining was federal policy, when race was an explicit underwriting criterion in many states, and when access to credit and insurance in certain communities was systematically restricted. The model may accurately describe historical loss patterns in certain ZIP codes or housing types, but those patterns were themselves shaped by discriminatory practices. The model finds a real signal. It just happens to be the fingerprint of past discrimination.
Neutral variables can act as proxies. Even when protected characteristics like race or national origin are explicitly excluded from a model, AI will find other variables that correlate with them. This is not a flaw, it is the model doing exactly what it was designed to do: finding any signal that predicts the outcome it is optimizing for.
ZIP code is the clearest example. It is not a protected characteristic. It also correlates strongly with race and income in most American cities, because of historical residential segregation. A model that uses geographic rating variables may be doing nothing technically prohibited while producing outcomes that fall disproportionately on certain communities. Credit score raises similar questions. It genuinely predicts loss experience and is a legally permitted rating factor in most states, but credit access has not been distributed equally in this country, and the correlations run in predictable directions.
AI can amplify bias over time, not just replicate it. This is the point I find most underappreciated in coverage of this topic. A machine learning model optimizes relentlessly for its target outcome. It will find and weight every available signal, including proxy variables, with more precision than any human actuary. When that model is deployed at scale and its decisions feed into the next generation of experience data, the bias can tighten rather than correct. Communities that were underserved become harder to insure, which affects behavior and maintenance, which affects loss experience, which the next model learns from. The bias becomes self-reinforcing in a way that is difficult to detect and harder to unwind.
Disparate impact testing is necessary but not sufficient. Checking whether a model’s outputs affect protected classes differently is an important step, but it is typically done at a single point in time on a defined test set. It does not necessarily capture how a model’s effects evolve over time in the real world. And because modern large language models (LLMs) often operate as black boxes, you can observe that disparate outcomes are occurring without being able to explain why, which creates real problems when a regulator or a policyholder asks for an explanation.
What Responsible Use Looks Like
None of this is an argument against using AI in underwriting. It is an argument for using it carefully, with sustained human oversight built into the process.
Require explainability. Any AI system driving underwriting decisions should be able to produce a plain-language explanation of why a given submission was rated, restricted, or declined the way it was. If a system cannot do that, it creates regulatory exposure and real problems for the agents who have to explain the decision to a client.
Human review must be ongoing, not just at launch. A model that passed a fairness evaluation when it was deployed and has not been reviewed since is not a model with a genuine fairness commitment. Outcome patterns by geography, demographic proxy, and line of business should be monitored continuously by people who have the authority to intervene when something looks wrong. This cannot be a one-time checkbox.
Consequential decisions need a human in the loop. Automated processing for routine renewals on straightforward risks is reasonable. Declinations, significant coverage restrictions, and material premium increases should have a human review step. Not because humans are always more accurate, but because accountability has to live somewhere, and “the model decided” is not an acceptable explanation to a policyholder or a regulator.
Know the regulatory environment. New York has been one of the more active states in scrutinizing algorithmic underwriting and pricing. The DFS has issued guidance on the use of external data and algorithms in insurance, and that guidance continues to evolve. What is compliant today may not be in two years, and both carriers and the technology vendors they work with need to be tracking that.
Represent your clients. As independent agents, we are the part of this industry closest to the people being insured. When a client receives a decision driven by an automated system and the explanation does not hold up, that is worth documenting and pushing back on. That ground-level visibility into how these models affect real people is something neither the carrier nor the vendor has, and it is genuinely valuable.
The Bottom Line
AI in underwriting is powerful technology being applied in a high-stakes domain with a complicated history, and where the consequences of getting it wrong fall disproportionately on people who are already underserved. The efficiency gains are real. The risk assessment improvements are real. The bias risks are also real, and they are more structural and more persistent than a single validation test will reveal.
Independent agents have always been the part of this industry with the closest relationship to the insured. As AI reshapes how underwriting decisions get made, that relationship matters more, not less.
Use the tools. Ask for explanations. And make sure someone is always watching the outputs.
Topics
