Artificial intelligence is transforming how financial institutions manage customer service, streamline operations, and make complex business decisions. For insurance companies, the challenge is finding ways to improve efficiency without compromising accuracy, customer trust, or financial risk management.
Massachusetts Mutual Life Insurance Company (MassMutual) offers a compelling example of how AI can support these goals. Serving more than four million policyowners and customers, the company reported record statutory operating earnings of $3.9 billion in 2025, total adjusted capital of $34.4 billion, and nearly $1.1 trillion in life insurance protection in force.
According to its 2025 annual report, MassMutual’s decade of technology investment has helped establish a foundation for generative and agentic AI initiatives. In 2025, approximately 51% of the company used broad-based generative AI productivity tools, contributing to an estimated 237,000 hours saved throughout the year.
Two applications demonstrate how the company is using AI to improve its operations. The first involves a generative AI assistant that helps service representatives and financial professionals retrieve information about complex insurance policies. The second uses machine learning to support life insurance underwriting by evaluating mortality risk and automating decisions for eligible applicants.
Together, these examples illustrate how AI can reduce repetitive work, accelerate decision-making, and help employees concentrate on tasks that require human judgment.
Life insurance policies can remain active for decades, creating challenges for service representatives and financial professionals who need to answer questions about older contracts. Policy terms, product conditions, and historical documentation can make finding the correct information a time-consuming process.
To address this challenge, MassMutual developed a generative AI assistant that allows employees to search for policy information using a conversational interface similar to ChatGPT. Fortune reported on the initiative in February 2025, highlighting its potential to help employees retrieve information about policies issued many years earlier.
The goal is to shorten the time between a customer’s question and the employee’s ability to provide a useful answer. Rather than manually searching through extensive documentation or contacting the service center for every question, representatives can query the assistant to locate relevant policy information.
Financial professionals can also use the tool to support their interactions with clients, reducing their dependence on the contact center for certain inquiries. This creates a more efficient service process while allowing employees to remain responsible for the information they share with policyowners.
A significant part of the implementation involved preparing the company’s data infrastructure. MassMutual CIO Sears Merritt’s team spent 36 months developing a cloud-based data platform designed to collect, store, and organize company information. This foundation enabled employees to access relevant information through the AI assistant.
The project demonstrates that implementing generative AI involves more than selecting a model or building a conversational interface. Organizations must also invest in data preparation, information accessibility, and the infrastructure needed to make AI useful within existing workflows.
Keeping Human Oversight at the Center
Although generative AI can accelerate information retrieval, it can also produce inaccurate or misleading answers. MassMutual addresses this concern through employee training and established governance practices.
Employees must complete formal training before gaining access to the company’s generative AI tools. The training covers the technology’s capabilities, limitations, and tendency to make mistakes, helping employees understand when AI-generated information requires additional scrutiny.
The company’s published governance practices also emphasize data transparency, privacy protections, human review of AI-initiated decisions, and ongoing monitoring of automated decisions. Its AI systems disclosure describes responsibilities throughout the system life cycle and risk-based approaches to model review.
These company-wide safeguards provide important context, although the published disclosures do not specify exactly how every response generated by the policy assistant is verified.
For organizations considering similar implementations, establishing clear review procedures is essential. Employees need to understand who is responsible for checking uncertain answers, when a question should be escalated to a specialist, and how errors should be documented. AI can accelerate the retrieval of information, but employees must still recognize when additional review is necessary.
Measuring the Impact of AI on Service Efficiency
MassMutual’s early results suggest that generative AI can deliver measurable operational improvements.
In February 2025, Fortune reported that the assistant had reduced response times by 10% and that financial advisors were making approximately 5% fewer calls to the contact center. These results indicate improvements in both the speed of handling questions and the ability of financial professionals to find information independently.
The assistant had reportedly been developed and deployed in parts of the business within nine months. At that time, the company was taking a cautious approach to introducing AI directly to policyholders, focusing instead on supporting employees internally.
MassMutual’s 2025 annual report confirms its continued investment in AI tools designed to support financial professionals and improve customer experience, alongside employee training.
The broader lesson is that organizations can begin realizing value from AI by improving the internal processes that support customer interactions. Employees who can find reliable information faster may be better positioned to provide timely service without sacrificing accountability.
Applying Machine Learning to Life Insurance Underwriting
Life insurance underwriting is another area where AI can improve efficiency while supporting complex decisions.
Traditional underwriting can involve medical examinations, blood tests, urine analysis, health questionnaires, and manual assessments of an applicant’s risk. Depending on the circumstances, gathering information and evaluating an application can take considerable time, delaying access to coverage.
For insurers, the challenge is to accelerate this process without weakening the assessment of mortality risk. Underwriting decisions must account for the likelihood of future claims while supporting appropriate pricing and responsible risk management.
MassMutual has explored this challenge through machine learning models that evaluate mortality risk and work alongside business rules to determine which applications can be processed automatically and which require additional review.
The company’s data scientists documented an earlier approach in a peer-reviewed paper presented at the 2019 Innovative Applications of Artificial Intelligence conference. The research described the MassMutual Mortality Score, or M3S, and its relationship to LifeScore360, a system designed to support life insurance risk assessment.
Although this research provides detailed evidence about the historical model and its implementation, MassMutual’s current disclosures do not establish whether its present-day underwriting programs use that same model or a replacement.
An important feature of the original project was collaboration among data scientists, actuaries, medical doctors, underwriters, and reinsurers. This multidisciplinary approach helped connect the technical development of the model with the practical requirements of insurance risk assessment.
For organizations introducing AI into regulated decision-making, involving the people responsible for evaluating and managing risk is essential. Technical performance alone is not enough; automated assessments must also fit the organization’s underwriting standards and decision-making processes.
Building a Mortality Risk Model from Historical Data
Developing the mortality model required substantial historical data preparation.
The research team trained its model using approximately 908,000 historical applications submitted between 1999 and 2014 that included laboratory testing. The dataset covered approximately 9.16 million exposure years and recorded around 15,700 deaths.
Much of the questionnaire information had originally been stored in paper archives, meaning the team first needed to digitize historical records before using them for analysis.
The model incorporated information such as applicant demographics, policy details, responses to health and behavioral questionnaires, and laboratory measurements. These measurements included cholesterol levels, liver and kidney function, blood sugar, and other screening indicators.
Some third-party information, including prescription histories and motor vehicle records, was excluded from the model because the historical dataset did not provide sufficient coverage to support its use.
This highlights a common challenge in AI implementation: a company may possess large amounts of information, but not all of it is sufficiently complete, consistent, or accessible to support reliable modeling.
The researchers used a machine learning technique called a random survival forest to estimate mortality risk over ten years based on nearly 60 inputs. The model then converted its assessment into a life score ranging from zero to 100, with higher scores representing lower relative risk among applicants within defined groups based on five-year age bands, sex, and smoking status.
The score is a relative comparison rather than a universal measure of mortality risk. Two applicants with the same score may have different absolute risk levels if they belong to different demographic groups.
This distinction matters when translating model predictions into underwriting decisions. The score helps assess an applicant relative to comparable individuals, while business rules and human review determine how the assessment influences the final offer.
Combining AI Scoring with Business Rules and Human Review
MassMutual’s historical underwriting approach combined machine learning predictions with a broader system of automated checks and underwriter oversight.
For eligible applicants, the process begins with a digital application and laboratory testing, generally completed through a paramedic visit. The model evaluates questionnaire responses and laboratory results to generate a life score that helps estimate the applicant’s relative mortality risk.
The score is then evaluated alongside thousands of automated business rules. These rules can identify additional risk factors and assign a less favorable rating than the model score alone might suggest.
Applications that trigger exceptions can be referred to underwriters, who may request additional information, review the circumstances, override a rule where appropriate, or confirm the suggested risk classification. Approximate contribution factors can also help explain how different inputs influence the model’s assessment.
The final insurance offer therefore depends on the combined results of the model, the applicable business rules, and any required human review.
This approach allows underwriters to focus their attention on applications that require closer examination rather than manually assessing every eligible file. AI supports the decision-making process without eliminating the role of professional judgment.
The company introduced the system gradually, beginning with a pilot involving 1,000 cases that ran alongside human underwriters. It subsequently automated best-class offers for policies worth up to $1 million for applicants aged 17 to 40, expanded eligibility to coverage of up to $3 million and applicants aged up to 59, and later included additional standard-and-above risk classes.
At the time the research was published, the authors reported that algorithmic underwriting covered 90% of applications within the specified parameters.
The staged rollout demonstrates the value of validating automated decisions against established human processes before expanding their scope. Beginning with a limited group of applications allows an organization to assess performance, identify potential issues, and refine the system before increasing automation.
The Results of AI-Assisted Underwriting
The historical research reported improvements in risk selection, processing speed, policy take-up, and operational efficiency. However, the findings came from a combination of historical simulations and deployment results, so they should not be interpreted as uniform outcomes for every applicant.
In a simulation using historical data, the model-selected best non-tobacco risk class had 6% fewer deaths over a 15-year horizon than the class selected by human underwriters. This result reflects differences in risk selection; it does not mean that the model prevented deaths.
The researchers also reported that policy issuance time fell by more than 25% for certain applicants. This indicates that automation can accelerate selected underwriting workflows, although the improvement did not apply to every application.
Policy take-up also improved in the reported comparison. Applicants who received a model-based best-class decision were more than 30% more likely to purchase their policies than those receiving a traditional best-class decision. However, this comparison does not establish that the model alone caused the increase.
The research further reported millions of dollars in operational efficiency savings over two years, although it did not provide a precise amount or calculation method.
Beyond its internal use, the technology also reached the broader insurance market. In March 2018, Swiss Re announced a partnership with LifeScore Labs, then identified as a wholly owned MassMutual subsidiary, to offer LifeScore360 as an additional analytics option within its Magnum underwriting system. This demonstrates the technology’s historical commercialization, rather than establishing its current product status.
Taken together, these findings illustrate how machine learning can support faster underwriting and more consistent risk assessment when integrated with established business processes and appropriate oversight.
MassMutual’s Current Approach to AI-Enabled Underwriting
MassMutual’s current underwriting disclosures describe programs that use different eligibility requirements to determine whether applicants can receive algorithmic assessments and whether medical examinations or laboratory tests are necessary.
Its algorithmic underwriting program is available for eligible individual term and whole life applications involving coverage of up to $3 million for applicants aged 17 to 59. Under this program, medical examinations and laboratory tests remain required.
MassExpress Fluidless takes a different approach. It uses AI and external consumer data to assess eligibility for underwriting without a medical examination or laboratory tests. The program applies to eligible applicants aged 17 to 50 seeking up to $3 million in coverage.
Both programs have additional eligibility requirements and product exclusions. MassMutual’s disclosure also identifies electronic health information, medical claims, pharmacy records, and credit-related information among the types of data available to its AI systems.
Importantly, meeting the published criteria does not guarantee that an applicant will receive a fluidless offer. Some applicants may still need medical examinations and laboratory testing. Certain otherwise eligible MassExpress applicants may also be selected randomly for control and audit purposes.
These provisions illustrate the importance of maintaining safeguards even when a process has been substantially automated. Eligibility rules establish the boundaries of the system, while additional assessments and audit procedures help the organization evaluate how its automated processes perform.
The available disclosures describe MassMutual’s current programs but do not establish that they rely on the exact mortality model documented in the 2019 research.
Lessons for Organizations Adopting AI
MassMutual’s experience demonstrates that successful AI implementation requires more than deploying a capable model. Organizations must also establish clear responsibilities, prepare their data, measure results, and maintain appropriate controls over automated decisions.
Clear decision ownership is particularly important. Employees need to understand when they can rely on AI-generated information, when additional verification is necessary, and who is accountable for resolving uncertain or incorrect outputs. This is especially relevant in financial services, where inaccurate information or poorly assessed risk can have significant consequences.
A gradual implementation strategy is equally valuable. MassMutual’s historical underwriting project began with a limited pilot before expanding to broader groups of applicants. Comparing automated results with established human decisions provides an opportunity to identify weaknesses and evaluate performance before increasing the system’s reach.
Data readiness is another fundamental requirement. The company’s work to build a cloud-based data platform and digitize historical underwriting records illustrates how much preparation may be required before AI can deliver meaningful results. Reliable models depend on information that is accessible, sufficiently complete, and appropriate for the intended use.
Finally, organizations should measure AI’s impact using outcomes that reflect their actual business objectives. In customer service, response times and contact center call volumes can reveal whether employees are becoming more efficient. In underwriting, processing times, risk selection, policy take-up, and operational costs provide different perspectives on performance. Evaluating these measures together offers a more balanced understanding of whether an AI initiative is delivering value.
Conclusion
MassMutual’s use of generative AI and machine learning demonstrates how artificial intelligence can support both everyday service operations and complex insurance decisions.
Its generative AI assistant helps employees retrieve policy information more efficiently, reducing response times and some reliance on contact center support. Meanwhile, its historical machine learning underwriting initiative illustrates how mortality risk scoring can work alongside business rules and human review to accelerate eligible applications.
The company’s experience also highlights the importance of strong data infrastructure, employee training, staged deployment, and ongoing oversight. These foundations help organizations introduce AI without losing sight of accuracy, accountability, and risk management.
As AI adoption continues across financial services, the greatest opportunities may come from combining automation with the expertise of the people responsible for customer relationships and critical decisions. MassMutual’s experience offers a practical example of how organizations can pursue greater efficiency while keeping human judgment and responsible governance central to the process.