Perfect Data, Wrong Risk

Oct 7, 2026

IIAI Article Hero Image

When AI Turns the Internet Into an Underwriting File, Accurate Evidence Can Still Produce an Inaccurate Decision

By James W. Moore

Key Takeaways

  • Underwriters have been checking websites and outside sources for decades. AI changes the economics by making that research practical across far more submissions.
  • The operational risk is not that AI makes inferences and humans do not. Both do. The risk appears when an inference becomes detached from the evidence and context that produced it.
  • Public web content is not a neutral description of normal operations. Stock imagery, unusual events, aspirational marketing, and thin online footprints can all distort the risk, and reused imagery can repeat the same misreading across many accounts.
  • Good systems can preserve sources, expose the inputs behind classifications, detect reused imagery, and make correction easier. The failure mode is a workflow choice, not an inevitable consequence of AI.
  • Underwriting managers should measure how much useful signal externally generated flags produce, how often review changes the conclusion, and how much friction the correction process creates.

A Crane on the Website

In the late 1990s, I submitted a construction account to an underwriter.

The prospect told us it did not use cranes. The underwriter went to the company’s website and found a stock photograph of a construction site with cranes and other heavy equipment. He came back to me with a reasonable concern: the application said one thing, while the prospect’s own website appeared to say another.

The website was real. The photograph was really there. The underwriter had found exactly what the prospect had chosen to publish.

The photograph just did not describe the prospect’s operations.

I went back to the insured and obtained a written statement confirming that it did not use cranes and would notify me and the carrier if that changed.

The account was written.

Almost thirty years later, the underwriting problem in that story is still recognizable. What has changed is how cheaply and consistently an insurer can look for the next crane.

Outside Research Is Old. Routine Outside Research Is New.

Insurance did not need artificial intelligence to discover external research.

Underwriters have always looked beyond the application when the economics of the account justified it. They checked property records, licenses, maps, photographs, business directories, loss information, and eventually company websites. Sometimes the outside evidence confirmed the submission. Sometimes it exposed a material omission. Sometimes, as with my construction account, it created a new question.

The constraint was time. One underwriter deliberately looked at one website on one account, and that kind of work did not scale easily across thousands of small commercial submissions.

It can now.

Coterie Insurance says its small-business underwriting platform uses publicly available business data to identify a business, detect its six-digit NAICS code, prefill operational details, and run automated underwriting checks in real time. Its SimplyBind materials also describe displaying some of the variables that affected underwriting so agents can review and update them.

In 2022, Coterie said its implementation of Carpe Data brought into the underwriting workflow the kind of external research that once belonged on an underwriter’s browser. According to the company, Carpe Data supplied business descriptions, services, operating hours, customer reviews, and other risk indicators that an underwriter might otherwise have searched for manually.

Carpe Data, now Carpe, makes the progression explicit with its current Minerva Reasoning Engine. The company says the system uses live business intelligence and carrier appetite to produce evidence-backed quote, decline, or refer recommendations.

That is not simply faster research. It is research becoming part of the underwriting production line.

When Evidence Becomes an Underwriting Fact

There is nothing inherently wrong with that progression. A modern system may preserve more evidence than the human process it replaces. A handwritten file note that says “crane exposure per website” may be far less useful later than a system that retains the URL, timestamp, source content, extracted observation, and reasoning path.

The important distinction is not human versus AI. It is attached versus detached workflow.

In my crane account, the underwriter made the wrong inference, but the inference remained attached to the evidence. He had seen the photograph. He told me what concerned him. I knew what needed to be answered.

An automated workflow can preserve that chain just as well, or better. It can also collapse the chain.

The underlying observation might be that an applicant’s website contains a photograph showing a crane. That can become “crane exposure detected,” then “applicant performs crane operations,” then a classification, referral, appetite failure, or decline.

By the time someone sees the underwriting output, the photograph itself may no longer be part of the conversation.

The categories are not philosophically clean. Even identifying a crane in an image requires interpretation. The practical point is simpler: can someone still get from the underwriting conclusion back to enough of the original evidence and reasoning to understand why the conclusion was reached?

Current vendors emphasize that connection. Carpe describes its recommendations as source-backed. Kalepa says its underwriting platform links insights and recommendations back to original sources and provides a digital trail from data to decision.

That is the right direction, but provenance answers only one question. It can tell you where the evidence came from. It cannot tell you what the evidence means.

The Internet Does Not Describe an Average Day

A company website is not an inspection report. Neither is its Instagram feed, a collection of customer photographs, or years of online reviews.

Imagine a restaurant that truthfully reports occasional live entertainment. An automated search finds hundreds of photographs and videos showing packed rooms, bands, dancing, alcohol, and special events.

Every image may be authentic.

But a hundred photographs of live music do not tell an underwriter whether it happens every Friday or twice a year.

The internet preserves what is interesting, not what is typical.

Businesses publish their biggest projects, busiest nights, most impressive equipment, special events, and aspirational marketing. The contractor may use the photograph that looks like the kind of project it hopes to win next year rather than the work it performs today.

The distortion can run the other direction too. A complex business with little online presence can look deceptively simple because there is almost nothing to find.

This is not an argument against using public information. The same research can expose a genuine mismatch between the application and reality. The point is that web evidence is a sample selected for purposes other than underwriting.

The underwriting system still has to decide what that sample proves.

Better Technology Can Catch the Same Problem

The automated version of the crane story can make AI sound worse than it deserves. A current system could have checked whether that photograph appeared elsewhere before anyone raised a concern.

The same stock image may appear across hundreds of unrelated contractor sites. If a system misreads that image once, the same interpretive error can repeat wherever the same pattern appears. That is not new to insurance, but automation can make a familiar concentration problem easier to scale, which is the same concern explored in Noise Diversifies. Bias Accumulates.

Google Cloud Vision’s web-detection tools, for example, can identify full and partial matches of an image elsewhere on the web, find pages containing the same image, and return visually similar images.

The same technology that detects an apparent exposure can therefore test the evidence before elevating it. The failure is not inevitable.

The operational question is whether the system merely finds the crane or also asks whether the crane belongs to this risk.

There is also a tradeoff underwriting managers should not pretend away. No enrichment process needs to be perfect to be useful. A system can generate occasional false positives and still create substantial value if it catches enough real exposures, reduces manual research, and makes the misses inexpensive to correct.

The useful measure is not whether a bad flag ever occurs. It is what the entire workflow does with good flags and bad ones.

A system that catches ten material inconsistencies and creates one easy-to-resolve false alarm may be performing well. A system that produces a flood of weak referrals can erase its own efficiency gains through agent questions, underwriter rework, and abandoned submissions.

That is an operating problem, not a philosophical one.

Ask Before You Infer

My construction account did not get resolved because the underwriter searched harder.

He asked.

More precisely, he showed me the conflict. I went back to the insured. The insured made the representation explicit and put it on the record.

The carrier did not have to choose between blindly trusting the website and blindly trusting the applicant. That is still a useful model.

If outside evidence conflicts with the applicant’s description on something material and reasonably answerable, the workflow can surface the evidence, ask a targeted question, preserve the answer, and move on.

AI can make that process cheaper too.

It can identify the conflict, retain the source, generate the follow-up question, capture the response, and route the account only when the answer still requires underwriting judgment.

The threshold matters. If every weak signal becomes another question, the correction process becomes the new bottleneck. Agents and insureds will not celebrate automation that simply converts underwriter research into a longer application.

Materiality is itself a judgment call, whether that judgment sits in a rule, a model, or an underwriter. Underwriting managers already manage thresholds everywhere else. This is another one.

For independent distribution, there is an additional operational advantage worth preserving. The agent often knows enough about the account to recognize when external evidence does not fit the risk.

That contextual layer only works if the agent knows what triggered the concern.

Hide the evidence behind a score, classification, or generic decline reason and the carrier may preserve automation while discarding one of the most useful parts of the distribution relationship. It also gives the agent one more reason to take the account somewhere else rather than spend time untangling a decision no one can explain.

Measure the Correction Path

Underwriting managers do not need a new governance framework to find out whether this problem exists in their own operation.

Take a sample of submissions where external data produced a material flag, classification, referral, or other underwriting concern, then compare the source evidence with the inference that entered the workflow. Measure how often underwriter review changes the conclusion, how often an agent or insured provides context that materially changes it, how many flags prove useful, how many create avoidable rework, and how long correction takes.

The objective is not a universal benchmark. Different books, classes, and appetites will tolerate different tradeoffs.

The objective is to find out whether the enrichment system is actually creating underwriting signal or merely moving research cost into a different part of the process.

A sound workflow preserves five things: the source, what the system observed, what it inferred, what underwriting action followed, and how a material mistake can be corrected.

That is less glamorous than another argument about model intelligence. It is also much closer to how insurance actually works.

The regulatory questions around external data deserve their own treatment. This article is about the underwriting operation itself.

Grounded Is Not the Same as Right

The underwriter did not hallucinate the crane.

The website was real. The photograph was real. His concern was reasonable, and the risk description was still wrong.

AI makes it possible to perform the equivalent of that website visit across far more submissions, with far more sources, at far lower cost. That can improve underwriting. It can expose facts an application missed, reduce manual work, and make small accounts worth researching in ways they never were before.

It can also make an incorrect interpretation cheaper to repeat.

The answer is not to stop looking. It is to build the research into a workflow that keeps the evidence close enough to the conclusion that someone can tell when the two do not belong together.

The AI does not have to hallucinate to be wrong.

 

 

Sources

Subscribe to AI Insights Newsletter

AI Disclaimer: This content was created with assistance from artificial intelligence technology. While content is based on factual information from the source material, readers should verify all details directly with the respective sources before making business decisions.

Related Posts

468