The End of the One-Size-Fits-All Insurance Questionnaire?

For decades, insurance underwriting has relied on a familiar ritual: long forms, repetitive questions, and a process that often feels as frustrating for applicants as it is operationally expensive for insurers.

But what if the questionnaire itself could think?

A new research paper, AI in Insurance: Adaptive Questionnaires for Improved Risk Profiling, points to a future in which underwriting journeys become shorter, smarter, and more personalised. Rather than asking every applicant the same fixed set of questions, the proposed framework uses AI to adapt the conversation in real time based on what is already known about the applicant.

For insurers, this is more than a better user experience story. It could signal a meaningful shift in how risk is assessed, how fraud is detected, and how underwriting capacity is scaled.

Why the traditional questionnaire is starting to show its age

Most insurance professionals understand the problem instinctively. Static questionnaires are blunt instruments.

They are long because they must cover many scenarios. They are repetitive because they cannot distinguish between what is relevant and what is not. And they depend heavily on self-reported information, which creates well-known issues around omission, error and, in some cases, fraud.

The paper argues that this approach struggles to capture individual nuance. Two applicants may receive the same questionnaire even if one has a highly straightforward risk profile and the other has a far more complex set of indicators spread across health, lifestyle and environment.

That is not just inefficient. It can also lead to weaker risk segmentation and less satisfactory customer journeys.

What ARQuest is trying to change

The framework proposed in the paper is called ARQuest, short for Adaptive Risk Questioning.

Its central idea is simple: instead of presenting a static form, use AI to build a questionnaire around the applicant.

The system does this in four stages:

1. User profiling
   It begins with basic personal information and, where permitted, external data sources such as health records, fitness data, social media content and geographic indicators.

2. Response forecasting
   An AI model analyses the available information, identifies likely risk factors, and predicts likely answers to some questions in advance.

3. Dynamic questioning
   Rather than asking everything, the model chooses the next most useful question based on what it already knows and what remains uncertain.

4. Risk assessment
   A final score is calculated using the completed profile, while discrepancies between AI-predicted answers and user corrections can also flag possible model weakness or potential misrepresentation.

That combination is important. ARQuest is not just asking fewer questions. It is trying to ask better questions.

What makes this especially relevant for insurers

There are three reasons this matters commercially.

1. It reduces friction at the point of sale

In the study, dynamic questionnaires required far fewer questions than the traditional approach. On average, applicants answered about half as many questions, and sometimes even fewer.

For any insurer focused on digital conversion, this matters. Every additional screen, field and interruption creates drop-off risk. A smoother underwriting journey can translate into higher completion rates and a better customer impression.

2. It may improve the quality of disclosures

The paper’s approach lets the system pre-fill likely answers and ask the customer to confirm or correct them. That changes the interaction from pure form-filling to guided review.

For the applicant, this feels more conversational and intuitive.
For the insurer, it creates a more structured way to test consistency between inferred context and declared responses.

That does not eliminate non-disclosure risk, but it may create a more robust signal layer around it.

3. It creates a pathway to more intelligent underwriting operations

The broader significance is operational. If underwriting can move from static forms to adaptive information gathering, insurers may be able to shorten processing times, reduce manual review loads, and target human intervention where it matters most.

That is especially attractive in product lines where underwriting cost and speed are in constant tension.

The interesting twist: users liked it more, but it was not yet more accurate

This is where the paper becomes especially useful for insurance leaders, because it avoids hype.

The researchers tested both traditional and adaptive approaches in life insurance. They ran one experiment using 85 synthetic applicants and another with 10 real users inside a mobile app.

The results were mixed in an important way.

• Traditional questionnaires were still slightly better at risk accuracy
• Dynamic questionnaires asked fewer questions
• Users clearly preferred the dynamic experience

That is the key takeaway.

In the synthetic-user tests, the traditional questionnaire achieved better alignment with the “true” risk scores. The adaptive model, even using GPT-4.1, underperformed somewhat, largely because it did not explore family history sufficiently. In other words, the model became more efficient, but sometimes at the cost of missing relevant signals.

But in the real-user study, participants strongly preferred the adaptive journey. They found it more engaging, less tedious and better tailored to their circumstances. Seventy percent preferred it outright.

For insurers, that creates a very familiar trade-off: customer experience versus technical completeness.

The long-term opportunity is not choosing one over the other. It is closing the gap.

The real innovation may be in question selection, not just risk scoring

Much of the insurance AI conversation focuses on predictive models: better pricing, better claims triage, better fraud detection.

This paper highlights something more foundational: the quality of underwriting depends on the quality of the questions asked in the first place.

That sounds obvious, but it is often overlooked.

A static form assumes the same information is equally valuable for every applicant. An adaptive questionnaire assumes the opposite: the value of the next question depends on what is already known.

That shift could have major implications beyond life insurance. In health, motor, travel and commercial lines, the ability to sequence questions intelligently could improve both applicant experience and data efficiency.

But there is a catch: external data changes the risk conversation

ARQuest depends in part on using external signals. In the study, those included:

• health records
• fitness data
• Instagram posts
• geographic indicators tied to municipalities

This is where insurance professionals should pause.

The technology may be impressive, but the governance burden rises sharply once underwriting begins incorporating alternative data. Questions quickly emerge:

• What data should be considered fair game?
• What is the legal basis for using it?
• How is customer consent obtained and explained?
• How do you avoid proxy discrimination?
• How do you challenge or audit a risk judgment built from messy external signals?

The paper is refreshingly direct about these issues. It highlights privacy, fairness, explainability and regulatory compliance as central concerns, not side notes.

That matters, particularly in Europe, where GDPR and the AI Act create a much stricter operating environment for high-impact decision systems.

For insurers, this suggests that adaptive underwriting will not just be a data science problem. It will be a joint design problem involving underwriting, compliance, legal, risk, product and customer experience teams.

Explainability is not optional here

One of the smarter features in the ARQuest design is that it does not only predict answers. It also provides explanations.

That is crucial.

If an AI system pre-fills a response suggesting elevated risk, both the customer and the insurer need some basis for understanding why. Otherwise, the process becomes opaque very quickly, and trust evaporates.

The paper also points to confidence scoring as an important control. If the AI is highly confident in a predicted answer and the applicant sharply corrects it, that discrepancy can be meaningful. It may indicate model weakness, poor data quality, or deliberate misrepresentation.

In practice, this is the kind of mechanism insurers will likely need if they want adaptive underwriting systems to pass internal governance thresholds.

What should insurance executives take from this now?

The message is not that traditional underwriting questionnaires are about to disappear next quarter.

It is that the questionnaire is becoming a strategic interface.

For years, insurers have treated the form as a necessary administrative object. This research suggests it may become an intelligent decision layer: one that shapes conversion, risk capture, fraud detection and operational efficiency all at once.

That should interest anyone responsible for:

• digital distribution
• underwriting transformation
• customer journey redesign
• AI governance
• risk operations

The likely near-term future is not full replacement of traditional methods, but hybrid underwriting journeys.

That might mean:

• using adaptive questioning for simpler or digitally acquired business
• falling back to traditional forms for complex or high-sum-assured cases
• combining external data with explicit customer review
• routing ambiguous or high-impact cases to human underwriters

In other words, the most realistic path is augmentation before automation.

Where this could go next

The paper suggests several future directions that are highly relevant commercially:

• testing the model across other lines of insurance
• improving performance with newer LLMs
• fine-tuning models on domain-specific insurance data
• introducing more agentic workflows that can interact with tools and external data sources automatically

If those capabilities mature, insurers could move toward underwriting journeys that are not only adaptive, but genuinely orchestrated: gathering evidence, testing for inconsistency, requesting clarification and escalating exceptions in near real time.

That would not just improve questionnaires. It would start to reshape underwriting itself.

Final thought

The most striking thing about this research is not that AI can generate insurance questions. That part is almost expected now.

The striking part is that it reframes underwriting as a live dialogue rather than a static declaration.

That is a profound change.

The insurers who win with AI may not simply be the ones with the best models. They may be the ones that become best at asking.

And in underwriting, better questions have always been where better decisions begin.