The Underwriting Backlog Problem Nobody Talks About Honestly
A mid-sized commercial lines carrier we worked with last year had roughly 1,200 new submissions sitting unacknowledged at any given moment. Not declined. Not quoted. Just sitting there, aging in a queue. Their underwriters were burning around 60% of each day pulling loss runs, cross-referencing ISO data, and reformatting information that brokers had already submitted in a slightly different layout. That is not underwriting. That is data janitorial work, and it is surprisingly common across carriers of every size, not just the ones you would expect.
Serious attention is finally coming to ai in insurance underwriting because carriers are starting to see what the bottleneck actually is. It is not underwriter judgment. It is everything that happens before judgment even enters the picture.
This piece covers how modern automated underwriting pipelines actually work in production, where they still fall short, and what compliance looks like heading into 2026. If you are somewhere in the build-versus-buy decision, there is a comparison table below that might save you a few vendor calls.
What Automated Underwriting Actually Replaces (And What It Doesn’t)
Let’s be specific about the before state. Traditional underwriting on a small commercial account typically goes like this: broker submits an ACORD form, usually as a PDF, an underwriter manually keys that data into a rating system, pulls third-party data from Dun and Bradstreet or LexisNexis Risk Solutions separately, checks internal loss history, applies judgment to edge cases, then routes approvals through one or two managers. That whole chain takes anywhere from two days to two weeks depending on complexity and how deep the queue happens to be on a given Monday morning.
Underwriting automation does not replace underwriter judgment on complex accounts. What it replaces is that manual chain. A well-built pipeline ingests the ACORD form through document AI, something like AWS Textract or a purpose-built extractor, calls third-party data APIs in parallel, runs a risk-scoring model, applies business rules, and either auto-approves, auto-declines, or routes to a human with a pre-populated file and a recommended decision already attached. The underwriter reviews an exception instead of building a file from scratch.
That shift alone can cut straight-through processing time on eligible accounts from days to under ten minutes. One national specialty carrier publicly cited 80% straight-through processing on their small business BOP line after rolling out an automated workflow. That is not universal. Complexity kills STP rates fast. But for standard-risk accounts, those numbers are real and repeatable.
How the Data Ingestion Layer Works
Most of the failures we have seen in underwriting automation happen at the ingestion layer, not in the model itself. Carriers pour money into the ML component and then massively underestimate how messy incoming data actually is in production. This happens more than vendors will ever admit in a demo, and it almost never comes up until after go-live.
A live ingestion pipeline typically needs to handle broker submissions arriving as PDFs, Excel files, and increasingly via API. It also has to pull third-party enrichment from credit bureaus, geospatial hazard data vendors like CoreLogic or Verisk, public records, and prior carrier loss history. Add internal policy and claims data, and sometimes real-time IoT feeds for commercial property or fleet risks, and you start to appreciate why ingestion is genuinely hard to get right.
The normalization step is where things get fragile. Broker submissions use different field names, different date formats, different NAICS codes for the same industry. If your normalization logic is brittle, the model is scoring garbage. We have seen carriers hit 95% accurate document extraction in testing, then drop to 72% on live broker submissions because real-world PDFs are messier than the training set. That is not a model problem. The model just gets blamed for it, which is its own kind of frustrating.
Good pipeline design validates and flags low-confidence extractions before they ever reach the model, rather than letting the model silently score a file with missing or corrupted fields. That sounds obvious. It is not done consistently, even at carriers with otherwise sophisticated tech stacks.
Traditional vs. Automated Underwriting: A Side-by-Side Look
| Dimension | Traditional Underwriting | Automated Underwriting (AI-Assisted) |
|---|---|---|
| Decision time (standard risk) | 2 to 14 days | Under 10 minutes (straight-through) |
| Data sources consulted | Manual pull, typically 3 to 5 sources | Automated API calls, 10 to 20+ sources in parallel |
| Consistency across underwriters | Variable (different interpretations of same criteria) | High consistency within model; variance only at edge cases |
| Scalability under volume spikes | Limited by headcount | Scales horizontally with compute |
| Handling of novel or complex risks | Human judgment, often superior | Model degrades; should route to human |
| Auditability | Underwriter notes, often incomplete | Full decision log; model inputs and outputs recorded |
| Regulatory explainability | Narrative justification by underwriter | Requires XAI layer (SHAP, LIME, etc.) for adverse action |
| Override rate risk | Low (human owns decision) | High if model trust is low (40%+ override rates observed) |
That override rate row deserves more attention than it usually gets. We have worked with carriers where underwriters were overriding automated recommendations more than 40% of the time. At that point you are running an expensive recommendation engine that nobody actually trusts, not a real automated workflow. The fix is almost always about calibration and transparency rather than model accuracy. Underwriters override when they cannot see why the model said what it said. That is a communication problem, not a technical one, and it is fixable if you treat it that way rather than keep tuning hyperparameters and hoping the trust issue resolves itself.
Model Explainability: Why SHAP Isn’t Enough on Its Own
SHAP (SHapley Additive exPlanations) has become the default explainability method for insurance underwriting software, and for good reason. It gives feature-level attribution: this account scored higher risk because its loss frequency over the prior three years was elevated, its location carries above-average catastrophe exposure, and its revenue-to-employee ratio is an outlier for that NAICS code. That is genuinely useful information for an underwriter reviewing an exception file.
But SHAP values explain the model’s math, not the business logic behind it.
The NAIC’s model bulletin published December 2023, which remains the operative framework heading into 2026, requires that insurers be able to explain adverse underwriting decisions in terms a consumer can actually understand, not in terms of feature contributions. That means you need a translation layer converting something like “SHAP value for feature X was 0.34” into plain language: “your prior claim history in this coverage category was a significant factor in this decision.” Most insurance underwriting software platforms do not do this out of the box. You build it yourself, or you buy a compliance wrapper from a vendor who already did the work.
There is also a subtler problem that does not come up enough. SHAP explanations are local, meaning per-prediction, but model drift is a global phenomenon. A model trained on 2021 to 2023 loss data may have learned that Gulf Coast properties in a certain construction class are acceptable risks. After the 2024 storm seasons, that assumption is probably wrong. The local explanations still look coherent. The model is quietly mispricing risk at scale. Catching that requires ongoing population-level monitoring, not just per-decision explainability. We saw this exact failure mode cause meaningful reserve inadequacy at a regional carrier that went 18 months without a proper model performance review. Not a fun conversation to be part of.
What Compliance Actually Requires in 2026
Regulatory pressure on ai in insurance underwriting has intensified since the NAIC adopted its model bulletin in late 2023. Several states, with Colorado being the most aggressive through SB21-169 already in effect for life insurers, have enacted or are advancing algorithmic accountability rules that touch underwriting decisions directly.
Here is what you practically need to have in place:
- A documented model inventory covering every model touching underwriting or rating decisions, including version history and performance benchmarks tracked over time, not just at launch. Snapshots at deployment do not cut it.
- Adverse action notice capability, so that if a model contributes to a declination or a rate increase above a defined threshold, the system generates a consumer-readable explanation in plain language.
- Bias testing across protected class proxies (geography, credit where restricted, and similar factors) at least annually, and ideally more often when market conditions shift quickly
- Human review triggers, with defined criteria for when a submission bypasses automated scoring entirely and goes straight to a person. These criteria need to be documented, not just understood informally by the team that built the workflow.
- Immutable audit logs: model inputs, outputs, the version used, and any human override with a reason code attached
That last one matters more than most people realize until they actually need it. If a regulator asks why account 47382 was declined in March 2025, you need to reproduce that decision exactly. Which model version was live, what inputs it received, what score it generated, whether a human touched it afterward. If your insurance underwriting software does not store all of that, you are exposed in ways that are genuinely hard to explain to a board after the fact.
Colorado’s approach has attracted attention across the industry because it requires carriers to test for unfair discrimination in any external data or algorithm used in underwriting, not just protected classes under federal law. Other states are watching closely. Build audit infrastructure now, before your state makes it mandatory. Retrofitting logging into a production pipeline is genuinely painful and always more expensive than doing it right the first time.
Accuracy Trade-offs You Should Actually Expect
Automated underwriting models tend to outperform human underwriters on high-volume, standard-risk accounts. Data from several large personal lines carriers suggests loss ratios on auto-approved accounts are often better than on manually underwritten accounts in the same segment, partly because models are consistent and partly because underwriters have limited bandwidth and cut corners when volume pressure is high. That is just reality, and it is worth saying plainly rather than dancing around it.
Where models underperform is a different story: new risk types, unusual industries, accounts with thin data histories, anything involving meaningful judgment about forward-looking risk. Like a restaurant expanding into catering, or a tech startup moving into hardware manufacturing. These are also exactly the accounts where skilled underwriting creates the most value. So the honest framing is this: automate the standard, preserve human bandwidth for the complex. Not a new insight. But a lot of carriers still have not operationalized it in any real way, and I am genuinely not sure why.
One thing worth flagging before you talk to vendors: model accuracy benchmarks are almost always measured on the vendor’s training distribution. Your book of business may not look anything like that distribution. Always request out-of-sample performance data on accounts similar to your actual submission mix before committing to any platform. If a vendor will not share that data, that tells you something worth paying attention to.
FAQ
How much does underwriting automation actually speed up decisions?
For standard-risk accounts that qualify for straight-through processing, decisions that used to take two to five business days routinely complete in under ten minutes with automated insurance underwriting. The speed gain is real, but it only applies to the subset of accounts the model handles with high confidence. Complex or unusual risks still route to humans, though often with a pre-populated file that cuts manual prep time by 60 to 70%, which is still a meaningful win even if it is less dramatic than the headline STP numbers suggest.
What’s the biggest risk of deploying AI in insurance underwriting without proper governance?
Model drift is the most underrated risk, full stop. A model can look fine on per-decision explanations while quietly mispricing an entire segment because the real-world risk environment shifted after training. Without population-level monitoring and a scheduled recalibration cadence, you probably will not catch it until loss ratios surface the problem, which could easily be 12 to 24 months later. Governance is not a compliance checkbox. It is genuinely how you avoid an actuarial surprise that nobody wants to explain to a reinsurer.
Does insurance underwriting software need to explain every automated decision to regulators?
Not every decision, but adverse ones, yes. Under the NAIC model bulletin framework and state-level rules like Colorado SB21-169, any automated decision resulting in a declination or a materially higher rate requires a consumer-facing explanation in plain language. Most platforms require custom development to produce these notices in a compliant format. It is not something you want to discover you are missing after a regulatory inquiry lands on your desk.
Where This Actually Leaves You
The case for underwriting automation is solid on standard-risk volume. Decision speed improvement is real. Consistency gains are real. The capacity unlock for experienced underwriters to focus on complex accounts is genuinely valuable and probably undersold in most vendor pitches.
We are not skeptical of the technology itself.
What we are skeptical of is implementations that treat the model as the hard part and treat data quality, explainability, and audit infrastructure as afterthoughts. That is where most projects actually fail. Not in the algorithm. In the plumbing around it, which nobody wants to budget for until something breaks and the post-mortem is uncomfortable for everyone in the room.
If you are early in evaluating options, the most useful thing you can do right now is audit your current submission intake process and document exactly where manual data assembly is consuming underwriter time. That is your automation opportunity map. The model comes second, not first.
Sources and References
- NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers (December 2023)
- Colorado Senate Bill 21-169: External Consumer Data and Information Sources in Insurance
- McKinsey and Company: Insurance 2030, The Impact of AI on the Future of Insurance
- Verisk ISO Commercial Lines Data Products
