How AI Underwriting Works: A Practical Walkthrough for Agents
Six years with a carrier. Solid relationship with the underwriter. A straightforward light manufacturing risk — same SIC code, similar revenue, nothing unusual in the loss history. Then a declination email arrives, and when you call the underwriter to ask what happened, the answer is: "the model scored it outside our appetite."
What model? What did it score? What specifically drove the score? Can the decision be appealed? The underwriter is sympathetic but cannot tell you much. The decision came from the platform, and the details are not easily accessible to them either.
This experience is becoming common in commercial lines, and agents who do not understand how AI underwriting actually works are at a disadvantage when it happens to them. You cannot advocate for a client with a carrier's model if you do not know what the model is doing or what inputs are driving its output. This article is a ground-level explanation — not a technical deep dive, but enough to help a working agent respond effectively.
The Problem AI Underwriting Is Solving
To understand why carriers are adopting AI underwriting tools, you need to understand the problem they are trying to solve. It is not primarily a cost problem, though efficiency gains are part of the pitch. It is a consistency and loss ratio problem.
Human underwriting at scale is inconsistent. Two underwriters presented with the same submission will produce different risk assessments with material frequency. That inconsistency is not random — it correlates with underwriter experience level, workload, the sequence in which submissions arrive, and time-of-day effects that are well-documented in behavioral research. An underwriter who reviews a submission at 4:30 PM on a Friday following a difficult week will apply different standards than the same underwriter reviewing the same submission on a Tuesday morning.
For carriers managing large books of business, this inconsistency is a loss ratio problem. Underwriters who are too lenient on a particular risk profile produce adverse selection in that segment; underwriters who are too conservative lose good business to competitors. Neither outcome is good.
The second driver is volume. Many carriers, particularly in small commercial, receive more submissions than their underwriting staff can evaluate carefully. The submission backlog creates pressure to make quick decisions on incomplete information, which produces more inconsistency. An AI model that can pre-screen submissions and prioritize the ones that need careful human attention addresses both the volume problem and, in theory, the consistency problem simultaneously.
The third driver is loss run data exploitation. Carriers are sitting on years or decades of claims and loss data. A model that can identify risk characteristics correlated with adverse loss outcomes — in ways that individual underwriters cannot detect by pattern-matching manually — represents genuine incremental information. Whether models actually deliver this in practice is the subject of ongoing debate, but the theoretical case is sound.
The Data Layer: What AI Underwriting Models Actually Use
The quality of an AI underwriting model is almost entirely determined by the quality and breadth of its training data. This is not a marketing claim — it is the fundamental constraint on what models can and cannot do, and it has direct implications for which carriers can deploy effective AI underwriting and which cannot.
Models in commercial lines draw from several data categories:
Internal loss history is the foundation. The carrier's own claims experience, mapped to risk characteristics from the original applications, is the training data the model knows best. Carriers with thin historical data in a segment — a regional carrier that recently entered a new line — will produce worse models in that segment than carriers with decades of claims data. This is the training data problem discussed in more detail in our coverage of AI underwriting tools.
Third-party data enrichment significantly extends what models can see beyond what the applicant disclosed on the application. Verisk is the dominant provider of enrichment data in U.S. insurance, supplying ISO commercial lines data, CLUE reports, and industry-specific risk data. Dun & Bradstreet business credit and financial stability data is commonly used for commercial accounts. Models use this enrichment data to fill gaps in the application and to cross-check what the applicant disclosed. If a business reports $800,000 in revenue but D&B data shows $3.2 million, the model will flag that discrepancy.
Geospatial data is used heavily in property and catastrophe-exposed lines. Flood zone designations, wildfire risk scores, wind exposure modeling, crime indices, and proximity to fire stations are all inputs that affect property underwriting scores. Catastrophe modeling data from providers like RMS and AIR (now Verisk) is often integrated directly.
Web signals and digital presence data are a newer category. Some models, particularly those built for small commercial, scrape and analyze the applicant's website, review sites like Google and Yelp, social media presence, and regulatory filings to identify signals about business operations, risk practices, and potential undisclosed exposures. A business whose website advertises services that are not disclosed on the application is a flag that a human underwriter would not routinely catch but a model can detect consistently.
Industry benchmarks and peer loss data allow models to compare a submission to the loss experience of similar businesses in the same industry code and geography. Gradient AI, Planck, and Cytora all use some form of industry benchmarking as part of their scoring approach. See our Gradient AI review and Cytora review for specifics on each platform's data layer.
This data enrichment layer is what gives AI underwriting models their practical advantage over rules-based systems. A rules-based system applies fixed eligibility criteria. A model can identify patterns in the data that no human sat down and specified as rules — which is both the source of its potential value and the source of its potential for error and bias.
Model Types in Practice: What the Technical Distinctions Mean
You will encounter references to gradient boosting models, neural networks, random forests, and occasionally large language models in descriptions of AI underwriting platforms. The technical distinctions matter less to a working agent than understanding the underlying principle that all of them share: they identify statistical patterns in historical data and use those patterns to score new submissions.
Gradient boosting models (XGBoost, LightGBM) are currently the most common architecture in structured insurance data applications. They handle tabular data well, are relatively interpretable compared to neural networks, and have a strong track record in loss ratio prediction tasks. Gradient AI and several actuarial platform vendors use gradient boosting variants.
Neural networks are more powerful but harder to interpret, which creates regulatory friction in underwriting applications where carriers need to explain decisions. Neural network approaches are more common in document processing and image analysis (auto damage assessment, property inspections) than in direct underwriting scoring.
Large language models are beginning to appear in submission intake and underwriting assistance — reading and summarizing broker submissions, extracting key risk data from loss runs, drafting coverage rationale. They are not yet widely used for the scoring decision itself.
The model type matters less than two other factors: the training data (discussed above) and the feature engineering — that is, how the model's inputs are constructed from raw data. Two gradient boosting models trained on different datasets will produce meaningfully different outputs on the same submission. A model trained primarily on workers comp claims will not score a professional liability submission well. Ask vendors about both the model architecture and the training data composition before drawing conclusions about accuracy.
The Submission Scoring Workflow: From Application to Decision
In practice, AI underwriting does not replace the underwriting decision — at least not in most current deployments. It stages the decision. Here is how the workflow typically looks in a carrier that has deployed a tool like Planck, Gradient AI, or Cytora:
- The submission arrives — either through the broker's AMS, a comparative rater, a wholesale platform, or direct carrier submission portal.
- The submission data is fed to the AI scoring model, either in real time or in a batch process. If intelligent intake tooling is in place, unstructured documents (loss runs, inspection reports, prior policy declarations) are parsed and extracted automatically.
- The model produces a risk score and, in most cases, a set of contributing factors — the top variables driving the score up or down.
- The score is compared against the carrier's current appetite parameters. If it falls within appetite, it may be auto-accepted (for simple, clean risks) or routed to an underwriter with the model's analysis as a starting point. If it falls outside appetite, it may be auto-declined or flagged for senior underwriter review.
- For risks that reach a human underwriter, the model output is presented as context — a dashboard showing the score, the contributing factors, the data inputs, and a comparison to similar risks in the carrier's book.
The critical point for agents is Step 4. When a submission gets declined at the model scoring stage without reaching a human underwriter, the decline may arrive very quickly (sometimes within hours of submission) and may contain limited explanation. That is not necessarily the underwriter being unhelpful — they may genuinely have limited visibility into the model's reasoning. For a head-to-head look at two of the leading scoring tools, see our Gradient AI vs. Planck comparison.
Explainability: What It Means and What Agents Can Request
Explainable AI is a term that gets used in several different ways in insurance contexts. At the technical level, it refers to methods for decomposing a model's output into the contributions of individual input variables — "this submission scored 67 out of 100, and the three factors with the largest negative effect on the score were: prior loss frequency, revenue-to-payroll ratio, and the hazard class of the primary SIC code."
Not all carriers have invested in explainability tooling. Some are running models that produce scores but cannot easily attribute the score to specific inputs in a way that is legible to either the underwriter or the agent. Regulators in several states are beginning to require that carriers be able to explain adverse underwriting decisions in terms of specific, articulable factors — which is driving investment in explainability infrastructure.
As an agent, here is what you can reasonably request:
If a submission is declined without explanation: Ask the underwriter whether there is a model-based pre-screening score associated with the declination and whether specific risk factors contributed to that score.
If the underwriter can access the model output: Ask which factors had the largest negative effect on the score. This is meaningful information because it may identify something you can address — an error in the application data, a third-party data source that contains outdated information, or a risk characteristic that can be mitigated.
If the decline is driven by third-party data: You have a right to know if a specific data report affected an adverse underwriting decision in states that have enacted AI insurance regulations. Ask the carrier to identify which data sources contributed to the decline.
The predictive underwriting landscape is evolving quickly on this point, and the regulatory requirements vary by state. But the general principle is that explainability is a legitimate ask, not an unusual one.
How AI Affects Carrier Appetite Signals
One of the least-discussed effects of AI underwriting is its impact on how carrier appetite shifts — and how quickly that shift happens relative to the signals agents receive.
In traditional underwriting, carrier appetite changes slowly. An underwriter's guidelines update annually or after a major reinsurance change. The signals agents receive — appetite guides, underwriter conversations, accepted and declined submissions — create a learning process that, over months, gives good agents a reasonable sense of where a carrier's appetite sits.
AI underwriting models do not work on that cycle. A model that is retrained on updated loss data, or whose appetite parameters are adjusted by a portfolio management team, can shift the effective appetite for a segment within days. Submissions that were within appetite in January may score outside it in March without any formal appetite guide change — and without the underwriter necessarily being aware of the shift at the parameter level.
This creates a specific frustration pattern: an agent who placed a risk in a class with a carrier successfully six months ago submits a nearly identical risk today and gets a quick decline. The underwriter is apologetic but cannot explain the change. The model parameters shifted, and neither the agent nor the underwriter has good visibility into what changed.
The practical implication is that agents should not assume carrier appetite is stable just because a recent similar submission was accepted. With AI-scored carriers, appetite is a continuously updating parameter, not a static policy document. The carrier appetite concept has changed in practical terms — treat it as dynamic.
Agentic Underwriting: The Emerging Next Step
The current generation of AI underwriting tools augments human underwriters. The emerging category — what the industry is beginning to call agentic underwriting — goes further: AI systems that can draft the submission rationale, ask clarifying questions of the broker, order and analyze supporting documents, and in some cases route the submission to binding without human intervention for qualifying risks.
Platforms like Sixfold and Hyperexponential are building in this direction. The concept is that for standard, low-complexity submissions that fall clearly within appetite parameters, the AI agent handles the full workflow from intake to bind. For submissions that require judgment, the agent prepares a fully documented analysis for the underwriter.
For agents, the practical implication of agentic underwriting is that the speed of certain transactions will increase, but the human touchpoint with the underwriter will decrease. For standard small commercial BOP, a 10-minute bind is becoming realistic for carriers that have deployed this technology. For complex or specialty risks, the human underwriter becomes more important, not less — the model handles the commodity work, freeing underwriters for the cases that require judgment.
What Agents Can Do When a Deal Gets AI-Declined
A quick decline from an AI-scored carrier is not necessarily the end of the placement. Here are the practical steps:
Request human review explicitly. Ask the underwriter to review the submission manually rather than accepting the model-based decline as final. Many carriers have a formal exception process; others handle it informally but do handle it.
Identify and address specific data issues. If the underwriter can access the model's contributing factors, look for data that appears incorrect or outdated. Third-party data sources are not always current. A business that closed a high-hazard operation two years ago may still have that operation reflected in data that was updated at the time.
Submit supporting documentation. Loss control reports, safety certifications, management experience documentation, and letters of explanation can sometimes be fed into the underwriter's manual review in a way that offsets model-based concerns. These documents do not change the model score, but they give the human reviewer a basis for an exception.
Consider wholesalers and surplus lines. If the carrier's AI model scores a risk outside admitted market appetite, that does not mean the risk is unplaceable. Wholesale brokers and E&S surplus lines markets exist specifically for risks that admitted markets decline. An AI declination from an admitted carrier is sometimes a routing signal, not a coverage impossibility.
Document the decline and the reason. If a carrier declines a risk and later you learn that the decline was driven by erroneous data, having documentation of the original decline and your subsequent inquiry strengthens any conversation about re-rating or correction.
The Regulatory Landscape
AI underwriting operates in a regulatory environment that is changing rapidly and unevenly across states. The key developments agents should know:
Colorado's SB 21-169, effective in 2023 and subsequently updated, requires carriers to audit insurance models for unfair discrimination and establish governance processes for AI-based decisions. Colorado is currently the most advanced state on this front, but several others — California, New York, Illinois — are at various stages of drafting AI insurance regulations.
The NAIC has issued model bulletins on the use of AI and machine learning in insurance, but NAIC bulletins do not have force of law — individual states must adopt them, and adoption is inconsistent. As of 2026, the U.S. AI underwriting regulatory landscape is a patchwork: aggressive in a handful of states, minimal in most others.
For agents, the regulatory context matters for two reasons. First, in states with strong AI underwriting regulations, you may have more leverage when requesting explanations for AI-based decisions. Second, carriers operating in multiple states face compliance costs that affect which AI underwriting tools they can deploy and how — a factor in understanding why carrier behavior may vary by state for the same risk class.
See our ai-underwriting-2026-adoption-roi coverage for the broader adoption and ROI context across the industry.
InsurAItools is editorially independent. We do not accept payment for placement or rankings. Our evaluation methodology is described at /methodology.
Editorial verdict: AI underwriting is not a black box by design — it is a black box by default in many deployments because carriers have not invested in explainability tooling or agent-facing communication processes. Agents who understand the underlying workflow are better positioned to respond to AI-based declines productively: by identifying data issues, requesting human review, and knowing when to route to the surplus market instead of continuing to work a declined submission. The carriers who have invested in transparent, explainable models will, over time, have better agent relationships than those who have not.
Marcus Reed is a senior insurance technology analyst with 12 years evaluating agency management and insurtech platforms. He previously served as operations director at a mid-size independent agency.
Frequently Asked Questions
Can an AI underwriting model be wrong?
Yes, and carriers know this. Models produce probabilistic scores, not facts. A model can score a risk outside appetite based on training data that does not reflect current conditions, third-party data that contains errors, or a legitimate edge case the training data underrepresents. The right response as an agent is not to accept the decline passively — it is to request human review and, if possible, identify the specific data inputs that produced the score. Models are decision-support tools, not infallible arbiters, and most carriers build exception processes for this reason.
As an agent, how do I know if AI scored my submission?
You often will not, at least not explicitly. Most carriers do not disclose whether a specific decision was AI-assisted. Signs include very fast declinations (minutes to hours rather than days), decline language that references "model output" or "scoring criteria," or carrier appetites that seem to shift without explanation. If you have a relationship with the underwriter, ask directly whether there is a pre-screening model in place for that risk class. Some carriers will answer this openly; others treat model deployment as proprietary.
Does AI underwriting discriminate against certain risk types?
This is a live regulatory concern. Models trained on historical loss data can encode historical biases — geographic patterns that correlate with demographic factors, industry codes that are underrepresented in training data, or risk characteristics that vary by region in ways that proxy for protected classes. Several states have begun requiring carriers to audit AI models for unfair discrimination. If you believe a client has been repeatedly declined by AI-scored carriers in a way that feels structural rather than specific to their individual risk profile, the regulatory context discussed in this article is relevant background.