Most Insurtech AI is Theater

The Problem with AI in Insurance

Every insurtech pitch deck has an AI slide. Document ingestion. Automated underwriting. End-to-end claims processing. The promises are bold and the demos are impressive—watch the system parse a complex submission in seconds, extracting dozens of data points that would take an experienced underwriter the better part of an hour to compile manually.

Then you deploy it in production. Reality hits fast.

Complex specialty supplementals—the kind with handwritten notes in the margins, inconsistent formatting across pages, and domain-specific terminology that varies by region—break these systems with remarkable consistency. The AI confidently extracts incorrect values. It hallucinates fields that don't exist in the source document. It misinterprets context in ways that would be obvious to any human reader with industry experience.

The result? The underwriter now spends more time correcting the AI's mistakes than they would have spent simply reading the document themselves. The efficiency gain evaporates. The frustration compounds.

This is what I call AI theater: technology that performs brilliantly on stage during carefully orchestrated demonstrations, then crumbles under the weight of real-world complexity.

Why Theater Persists in the Market

The incentive structures in insurtech actively reward theatrical AI implementations. Investors want to hear "AI-powered" in the pitch. Buyers want to check the innovation box in their procurement process. Sales teams know that a flashy demo closes deals faster than a honest conversation about capabilities and limitations.

And the demos themselves? They're carefully choreographed performances. Clean documents with standard formatting. Common submission types that represent the easy 70% of the distribution. Predictable inputs that the system has been specifically tuned to handle.

But insurance—particularly specialty insurance—is inherently messy. Specialty lines exist precisely because they involve edge cases that don't fit neatly into standard underwriting boxes. The submissions that matter most are often the ones that deviate from the norm. An 80% accuracy rate that sounds impressive in a pitch deck means that 20% of submissions still require manual review—and those tend to be the most complex, highest-value, and most consequential submissions in the queue.

"Automation applied to an inefficient operation will magnify the inefficiency."

— Bill Gates

The same principle applies to AI in insurance workflows. If your underlying process is messy and full of exceptions, throwing AI at it won't magically create order. It will amplify the chaos and introduce new failure modes that didn't exist before.

Four AI Applications That Actually Deliver Value

After spending years building our own tools and systematically evaluating competitors across the insurtech landscape, I've identified four categories of AI applications that consistently deliver measurable value in production environments:

1. Summarization

AI excels at condensing lengthy, complex documents into digestible briefs that humans can quickly scan and evaluate. The key distinction here is important: this is not about extracting specific fields or structured data from unstructured documents. That's where AI theater lives. This is about synthesizing context and presenting the essential narrative of a submission.

An underwriter can read a well-crafted two-paragraph summary and make an informed decision about whether to invest time diving deeper into the full submission. For the significant percentage of submissions that will ultimately be declined, this represents hours of saved time. The AI handles the cognitive labor of initial filtering; the human retains full authority over the judgment call.

2. Underwriting Guidance Based on Existing Guidelines

When you feed AI your organization's actual underwriting guidelines—rather than training it to develop new ones from historical data—it becomes a genuinely powerful co-pilot for your underwriting team. The system can flag submissions that violate your stated appetite before an underwriter invests significant time in analysis. It can suggest relevant questions based on your established criteria. It can ensure consistency across a distributed team where individual underwriters might otherwise apply subtly different interpretations of the same guidelines.

The critical insight here is that the AI isn't making underwriting decisions. It's ensuring that human underwriters have the right information and context to make better decisions more efficiently. The guidelines come from human expertise; the AI simply operationalizes their consistent application.

3. Sourcing Data at Scale

AI is exceptionally good at gathering and organizing external data from multiple sources: property characteristics from public records, business information from various databases, market comparables from historical transactions, regulatory filings, news mentions, and dozens of other signals that inform underwriting decisions.

These are tasks that require breadth rather than depth. The AI finds and aggregates; the human reviews and judges. No complex interpretation is required—just efficient collection and organization of information that would take a human analyst hours to compile manually. The underwriter's time is then freed to focus on analysis and judgment rather than data gathering.

4. Validation of Submitted Assets

Checking whether a submitted photo actually matches a property description. Verifying that uploaded documents are what they claim to be. Confirming that certificates and attestations appear authentic. These are binary validation tasks where AI can flag anomalies for human review without making final determinations.

The pattern here is risk reduction through systematic verification. The AI doesn't decide whether a submission is fraudulent—it identifies submissions that warrant closer human scrutiny. False positives are acceptable because the cost of human review is low. False negatives are minimized because the AI applies consistent checking across every submission.

Deterministic vs. Non-Deterministic: You Need Both

There's a related debate in insurtech that often gets flattened into false binaries: should systems be deterministic (rule-based, predictable, same input always yields same output) or non-deterministic (probabilistic, AI-driven, adapting to context)?

The answer, unsurprisingly, is both. But knowing when to apply each is where most platforms get it wrong.

Deterministic tools excel where the rules are clear and the stakes of inconsistency are high. Compliance checks. Regulatory requirements. Appetite boundaries that should never flex. If your guidelines say you don't write restaurants with fryers over 50 gallons, that's not a judgment call—it's a hard rule. A deterministic system enforces it the same way every time, creates an auditable trail, and never hallucinates an exception that doesn't exist.

Non-deterministic tools—the AI and ML systems that dominate insurtech marketing—excel in a different domain entirely. They handle ambiguity. They synthesize unstructured information. They surface patterns that would be invisible to rule-based systems. The summarization, data sourcing, and validation tasks I described earlier all benefit from non-deterministic approaches precisely because the inputs are messy and the "right answer" isn't always binary.

The mistake I see repeatedly is platforms trying to use non-deterministic AI for tasks that demand deterministic consistency. An underwriter asks why the system flagged one submission but not another with similar characteristics. The answer—"the model weighted these features differently based on contextual factors"—is technically accurate and operationally useless. For compliance-critical decisions, you need to be able to explain exactly why something happened. Every time. To regulators, to auditors, to the underwriter who needs to trust the system.

The best architectures I've seen treat this as a layering problem. Deterministic rules form the foundation—the hard boundaries that never move. Non-deterministic AI operates within those boundaries, handling the ambiguity and complexity that rules can't capture. The human sits on top, applying judgment to the cases that neither system can resolve with confidence.

It's less exciting than "end-to-end AI underwriting." But it actually works.

The Underlying Pattern

AI works when it augments human judgment, not when it attempts to replace it.

Each of the four successful applications above shares a common structure:

• Summarization gives underwriters better inputs for their decisions
• Guidance ensures they're applying consistent criteria across the organization
• Data sourcing expands what they can realistically consider in their analysis
• Validation catches potential errors before they compound into larger problems

In every case, the human remains firmly in control of the judgment. The AI handles the cognitive labor that doesn't require expertise—the filtering, gathering, checking, and organizing. The human applies the expertise that AI cannot reliably replicate—the contextual judgment, the pattern recognition born of experience, the intuition about what feels wrong even when the data looks right.

None of these applications are as exciting as "fully automated underwriting." They don't make for compelling pitch deck slides or breathless press releases. But they ship. They work in production. They deliver measurable ROI. And they don't create new categories of expensive failures.

The Question to Ask

The companies that ultimately win in insurtech won't be the ones with the most impressive demos or the boldest claims about AI capabilities. They'll be the ones solving specific, well-defined problems where AI's genuine strengths align with the task requirements—and where human expertise remains central to the value chain.

The next time you evaluate an AI feature in an insurance platform, ask yourself a simple question: Is this theater, or is this work?

The answer will tell you everything you need to know about whether the capability will survive contact with your actual production environment.