AI May Know Everything About Underwriting. But Does It Know Everything About the Risk?

By Erik Leeming Know The Underwriting™ | Professional Insurance Foundation

Artificial intelligence may be capable of analyzing virtually every aspect of a commercial insurance policy. But what happens when the most important information about the risk isn't available for it to find?

The Underwriting Summary Looked Perfect. But Was It?

Imagine an underwriter receiving an AI-generated assessment of a commercial property.

The report is impressive.

Construction details. Occupancy information. Fire protection capabilities. Surrounding exposures. Historical property information. Even an evaluation of potential hazards.

It's organized, thorough, and supported by information gathered from numerous sources.

Everything suggests that AI has done an extraordinary job.

But there's something the report doesn't mention.

Three years ago, the building's occupant introduced a new manufacturing operation involving combustible materials. The change wasn't reflected in public records, the company's website, or any of the databases AI consulted.

The underwriter doesn't know about it either.

How can AI evaluate an exposure it doesn't know exists?

And there's an even more interesting question.

If AI cannot find the information, will it recognize that something might be missing?

These questions get to the heart of a discussion I recently had during an insurance industry conference.

When Will AI Take Over Underwriting?

During a presentation exploring artificial intelligence and its expanding role in underwriting, someone asked:

"At what point will AI take over all of these functions?"

The conversation naturally turned toward the extraordinary speed at which AI is evolving.

And there's no disputing that progress.

But I offered a different perspective.

What if the greatest limitation of AI in underwriting isn't its ability to analyze information, but its ability to obtain all the information necessary to properly evaluate a risk?

Those are two very different things.

Before exploring that limitation, however, let's recognize just how valuable artificial intelligence can be.

Where AI Can Be Extraordinary

Consider the evaluation of competing commercial liability coverage forms.

An underwriter receives two proposals offering $10 million in umbrella liability protection.

Both are described as umbrella policies.

But are they?

One contains an independent insuring agreement potentially providing coverage for certain exposures excluded by underlying insurance, subject to a self-insured retention.

The other includes restrictive language that effectively eliminates that broader protection, making it function much more like an excess liability policy.

An AI assistant, provided with the complete forms and endorsements, can compare the provisions in extraordinary detail.

For example, the ISO Commercial Liability Umbrella Coverage Form, CU 00 01 04 13, contains independent insuring agreements. Under Coverage A, the insurer agrees to pay covered ultimate net loss exceeding the applicable retained limit.

The ISO Commercial Excess Liability Coverage Form, CX 00 01 04 13, operates differently, with its coverage more directly dependent upon the underlying insurance.

Those distinctions, including the possible application of a self-insured retention, can materially affect coverage.

AI can examine both forms, their endorsements, exclusions, definitions, defense provisions, and conditions.

It can identify provisions restricting coverage, highlight inconsistencies, and explain why two policies bearing similar descriptions may offer substantially different protection.

The name on the policy doesn't determine the coverage. The actual language does.

This is an area where AI can be enormously beneficial.

It can perform a detailed comparison that might take an experienced insurance professional hours, and identify provisions deserving closer attention.

And while its conclusions should still be verified, the potential efficiency and analytical value are remarkable.

But that example reveals something important.

AI performs best when the information necessary for its analysis is actually available.

How Does AI Find All This Information?

Many people may not appreciate the lengths to which modern AI systems can go to gather information.

First, AI draws upon knowledge developed during its training, including terminology, concepts, and relationships learned from extensive collections of information.

Second, it can analyze materials supplied directly by a user: applications, financial statements, loss runs, policy forms, inspection reports, and other underwriting documents.

Third, when equipped with appropriate tools and permissions, AI can search beyond those materials.

It may examine publicly available property records, business websites, aerial imagery, historical weather information, government databases, and other accessible sources.

Some systems can connect directly to carrier databases, proprietary information services, and other internal resources. Others can analyze photographs, video, telematics, and information collected through connected devices.

Importantly, these capabilities vary among AI systems. Not every platform has unrestricted internet access or access to proprietary insurance information.

Nevertheless, the potential is extraordinary.

Imagine AI combining years of loss experience, an insured's financial records, property information, underwriting guidelines, and relevant policy language into a single analysis.

It might even identify inconsistencies that an underwriter overlooked.

But all that capability raises a fundamental question:

What happens when the information AI needs has never been recorded, isn't accessible, or simply hasn't been disclosed?

Let's look at a few familiar underwriting challenges.

Commercial Property: Does AI Really Know the COPE?

Consider that 35-year-old commercial building.

AI might identify its original construction, square footage, location, fire protection characteristics, and neighboring exposures.

But does it know that the electrical system was partially updated?

That the sprinkler protection was modified?

That a manufacturing process changed?

Or that valuable machinery was installed without any corresponding update to the insured's property schedule?

The ISO Building and Personal Property Coverage Form, CP 00 10, distinguishes Building from Business Personal Property, including how permanently installed machinery and equipment may be classified.

AI can interpret those provisions and help evaluate the adequacy of reported values.

But it cannot reliably establish the existence and value of property nobody has identified or documented.

Understanding what should be insured isn't the same as knowing everything the insured owns.

Business Income: When the Numbers Aren't Enough

Now consider Business Income coverage.

Provide AI with several years of financial statements, and it may produce an impressive analysis of revenue, operating expenses, and projected losses.

It can evaluate the ISO Business Income (and Extra Expense) Coverage Form, CP 00 30, which addresses Business Income, continuing normal operating expenses, Extra Expense, and the period of restoration.

But consider what it must know to develop a meaningful Business Income exposure analysis.

Will payroll continue following a major loss?

Which expenses will cease?

Could operations temporarily relocate?

How long will replacement equipment take to arrive?

Could Extra Expense expenditures reduce the ultimate interruption loss?

And how long might rebuilding, permitting, or equipment installation actually take?

Some answers may be found in financial statements. Others require an understanding of the insured's operations, contingency plans, and management decisions.

AI can make assumptions.

It can model multiple scenarios.

But a mathematically correct calculation based on incorrect assumptions is still an incorrect analysis.

Commercial Auto: The VIN Isn't the Whole Story

Commercial Auto presents another challenge.

AI might identify a vehicle's make, model, gross vehicle weight, and other characteristics. It may understand the applicable rating rules and recommend a five-digit commercial auto classification code.

But suppose two businesses own identical Ford F-350 trucks.

One primarily transports tools between nearby job sites. The other regularly transports materials over much greater distances.

Their classification and rating characteristics may differ substantially.

AI can decode a VIN and analyze the classification rules.

But where will it find reliable information about actual vehicle use, operating radius, and the business activities involving that vehicle?

That information may exist nowhere outside the insured's own operations.

Knowing how to classify a vehicle isn't the same as knowing how that vehicle is actually used.

Workers Compensation: Knowing the Scopes Manual Isn't Knowing the Employees

Workers Compensation may provide one of the strongest examples.

Imagine an AI system with access to the applicable NCCI Scopes Manual, Basic Manual, classification rules, and state exceptions.

It could potentially become exceptionally proficient at evaluating Workers Compensation classifications.

But does it know what an insured's employees actually do?

Suppose an employee performs duties involving two different classifications.

AI may understand the rules governing interchange of labor and payroll division. However, those rules generally require accurate payroll records documenting actual time spent in qualifying classifications. Estimated percentage allocations are not sufficient.

The applicable NCCI rules and state exceptions determine when such division is permitted.

AI can interpret the rules.

But without knowing the employee's actual duties and examining the necessary records, how can it establish the proper classification or payroll allocation?

And what about multistate operations?

The standard NCCI Workers Compensation and Employers Liability Insurance Policy, WC 00 00 00 C, distinguishes states identified in Item 3.A. of the Information Page from those addressed under Item 3.C., Other States Insurance.

Part Three includes specific provisions concerning work begun after policy inception and notification requirements.

AI can explain those provisions perfectly.

But does it know that an employee is working remotely in another state? That a construction crew recently began work across state lines? Or that operations already existed in another jurisdiction when the policy began?

And what about uninsured subcontractors?

Does AI know which subcontractors the insured uses, whether they have valid Workers Compensation coverage, or whether new subcontractors have been hired since the original application?

Without that information, it cannot reliably evaluate the potential premium or statutory exposure.

Knowing the Scopes Manual is not the same as knowing the insured's operations.

And therein lies the distinction.

Isn't This Also True of Human Underwriters?

Of course it is.

An underwriter who overlooks missing information has an underwriting problem regardless of the tools being used.

In fact, AI may substantially improve the situation.

It can identify incomplete applications, recognize inconsistencies, formulate questions, review inspection results, and help underwriters recognize exposures requiring additional investigation.

Experienced underwriters have always relied on applications, inspections, producer discussions, financial information, and other resources to establish the facts necessary to evaluate a risk.

AI can enhance that entire process.

It may eventually conduct sophisticated underwriting interviews, automatically request documentation, arrange additional inspections, and identify discrepancies across numerous information sources.

But even then, one fundamental principle remains.

Asking the right questions doesn't guarantee that the answers are accurate or complete.

A completed application isn't necessarily a complete description of the risk.

A database record isn't necessarily current.

And a polished underwriting report isn't necessarily based on everything an underwriter needs to know.

Will More Powerful AI Solve the Problem?

Artificial intelligence will undoubtedly continue advancing.

It will become faster, more capable, and better equipped to identify missing information.

New technologies may provide access to information previously unavailable through traditional underwriting methods.

But increased analytical capability alone cannot eliminate the underlying issue.

An AI system may know every classification rule, interpret every policy provision, and analyze every financial statement with extraordinary precision.

Yet it may still lack one critical fact about the insured.

A new operation.

An undocumented building improvement.

An employee working in another state.

An uninsured subcontractor.

A vehicle being used differently than reported.

The distinction is between what is known, what can be reasonably inferred, and what has never been established.

That distinction applies to human underwriters, too.

And regardless of how sophisticated AI becomes, underwriting will still depend on the quality, completeness, and reliability of the information available.

The Question We Should Really Be Asking

I'm enthusiastic about the future of artificial intelligence in insurance.

It has the potential to dramatically improve underwriting research, coverage analysis, classification accuracy, risk evaluation, and operational efficiency.

Used effectively, it can make good underwriters even better.

But perhaps our industry's discussion about AI replacing underwriting professionals is focused on the wrong question.

Rather than asking when artificial intelligence will become capable of performing every underwriting function, perhaps we should ask:

When will it have the complete, accurate, and current information necessary to perform those functions reliably?

Because those are two very different challenges.

And no matter how remarkable the technology becomes, one distinction deserves our attention:

AI may eventually know everything there is to know about underwriting. That doesn't mean it knows everything there is to know about the risk being underwritten.

Know The Underwriting™ | Professional Insurance Foundation

Know The Policy™

Coverage and classification references are illustrative. Actual coverage and rating requirements depend on applicable policy editions, endorsements, jurisdictional rules, and individual risk characteristics.