Massachusetts Mutual Life Insurance Company (MassMutual) is a mutual life insurer serving more than 4 million policyowners and customers. For 2025, the company <a href="https://www.massmutual.com/about-us/news-and-press-releases/press-releases/2026/03/massmutual-delivers-excellent-2025-financial-results” rel=”nofollow noopener” target=”_blank”>reported record statutory operating earnings of $3.9 billion, total adjusted capital of $34.4 billion, and nearly $1.1 trillion in life insurance protection in force.
The company’s 2025 annual report attributes its generative and agentic AI work to a decade of technology investment. It reports that 51% of the company uses broad-based generative AI productivity tools, saving an estimated 237,000 hours in 2025.
For executives, MassMutual’s scale highlights the challenge of improving service speed and operational efficiency while maintaining accuracy and control over financial risk.
This article examines two internal AI use cases at MassMutual:
- Answering policy questions for service teams with generative AI: Service representatives and financial professionals retrieve information about complex, long-dated policies, helping reduce response times and calls to the contact center.
- Scoring mortality risk to automate life insurance underwriting: A machine learning model supports automated decisions on eligible applications, allowing underwriters to focus on cases that need closer review.
We begin by examining how MassMutual applies generative AI to the questions its service representatives and financial professionals handle every day.
Answering policy questions for service teams with generative AI
Life insurance policies can remain in force for decades, leaving service representatives and financial professionals answering questions about contracts written under older product terms. MassMutual built a generative AI assistant to help employees retrieve policy information, including information about policies issued decades earlier, Fortune reported in February 2025.
The business problem is the time spent finding that information and escalating questions to the contact center. Giving employees a way to query policy material can shorten the path from a client’s question to an answer.
In a November 2024 company blog, Adam Fox, then head of distribution technology and data science, described generative AI initiatives aimed at improving customer service and productivity to help financial professionals serve policyowners.
The assistant runs on a cloud-based data platform that MassMutual CIO Sears Merritt’s team spent 36 months building to collect, store, and organize company data. Fortune describes an interface similar to ChatGPT through which employees can ask questions. The preparation of company information was therefore a substantial part of the implementation.
The assistant shifts policy-information searches toward employee self-service, helping representatives and advisors answer questions faster while retaining responsibility for what reaches the client:
- Policy lookup: Representatives query the assistant about policy terms, including those associated with older contracts.
- Advisor self-service: Financial professionals can obtain information to support client service, reducing some of their need to call the contact center.
- Human review: The assistant supports employees, who remain involved in decisions and responsible for the information they communicate to clients.
Employees must complete formal training before accessing the generative AI tools. That training covers the technology’s capabilities, limitations, and tendency to make mistakes. Access therefore depends on preparation to use and assess the output.
MassMutual’s video below places virtual assistance within its broader support for financial advisors. It describes day-and-night assistance and the use of operational data to provide insights.
The company’s published governance practices add detail to the human-review requirement. Its 2024 annual report describes data transparency, privacy protections, human review of AI-initiated decisions, and continued monitoring of automated decisions.
MassMutual’s AI systems disclosure also assigns responsibilities throughout a system’s life cycle and describes model reviews prioritized according to risk. These are company-wide controls; the disclosure does not explain how individual answers from the policy assistant are checked.
For leaders introducing a similar tool, the practical step is to define who checks uncertain answers, when a question returns to a specialist, and how errors are recorded before expanding access. Faster retrieval creates value only if employees can recognize when the answer needs further review.
In February 2025, Fortune reported that the assistant had reduced response time by 10% and that financial advisors were placing approximately 5% fewer calls to the contact center, indicating improvements in service speed and advisor self-service.
The two measures address different parts of the workflow: response time captures how quickly questions are handled, while call volume indicates whether advisors need less assistance from the contact center.
At the time of Fortune’s report, the assistant had been built and deployed in parts of the business within nine months. The tools were internal, and Merritt described a cautious approach to putting AI directly in front of policyholders.
MassMutual’s 2025 annual report confirms continued investment in AI tools supporting financial professionals and customer experience, alongside employee training.
Scoring mortality risk to automate life insurance underwriting
Traditional life insurance underwriting can involve a physical examination, blood work, urine analysis, and a manual assessment of an applicant’s risk. The National Association of Insurance Commissioners notes that issuing a policy can take months, creating an obstacle for people seeking coverage.
A faster process still has to estimate mortality risk accurately enough to price each policy and manage risk across the insurer’s portfolio. For MassMutual, the workflow change was to use a machine learning model and business rules to evaluate eligible applications, with underwriters reviewing exceptions.
MassMutual’s data scientists documented this approach in a peer-reviewed paper presented at the 2019 Innovative Applications of Artificial Intelligence conference. The paper describes a mortality model underpinning the MassMutual Mortality Score (M3S) and LifeScore360.
The historical model provides detailed evidence about implementation and results. MassMutual’s current disclosures describe the underwriting programs available today, but do not establish whether those programs use the same model or a replacement.
The authors emphasized collaboration with actuaries, medical doctors, underwriters, and reinsurers. For leaders automating a regulated decision, the practical lesson is to involve the people responsible for assessing and managing risk before setting the model’s decision boundaries.
The team trained the model on roughly 908,000 historical applications from 1999 to 2014 that included lab testing. The dataset covered 9.16 million exposure years and approximately 15,700 observed deaths. Much of the questionnaire data first had to be digitized from paper archives.
The model’s inputs depended on which applicant records could be linked to mortality outcomes, making historical data coverage a constraint on what the team could use:
- Application data: Demographics, policy details, and responses to health and behavioral questionnaires.
- Lab results: Biophysical measurements, cholesterol, liver and kidney function, blood sugar, and screening indicators.
- Excluded third-party data: Prescription histories and motor vehicle records informed underwriting but were excluded from the model because sufficient historical coverage was unavailable.
The model uses a machine learning method called a random survival forest to estimate mortality risk over 10 years from nearly 60 inputs. It converts that estimate into a life score from 0 to 100, with higher scores indicating lower relative risk within groups defined by five-year age bands, sex, and smoking status.
The life score ranks applicants against others in the same age, sex, and smoking group. Two applicants with the same score can therefore have different mortality risks if they belong to different groups. For underwriting leaders, this distinction matters when setting offer thresholds: the score supports comparison within a defined group, while business rules and exception review determine how that assessment translates into an offer.
For eligible applications, the system combines mortality scoring with business rules to determine which cases can proceed and which require an underwriter’s review:
- Application and testing: The applicant completes a digital application and lab tests, generally through a paramedic visit.
- Risk scoring: The model converts questionnaire responses and lab results into a life score.
- Business-rule checks: Thousands of automated rules check the application and can assign a worse rating than the score alone suggests.
- Exception review: Underwriters assess flagged cases, request more information, override a rule, or confirm the suggested class. Approximate contribution factors help explain the model’s assessment.
- Offer determination: The score, rules, and any required underwriter review inform the final offer.
Underwriters therefore review exceptions instead of manually assessing every eligible file. The model supports the decision within a wider process that retains business rules and human review.
MassMutual began with a 1,000-case pilot running alongside human underwriters. It then automated best-class offers for policies up to $1 million for applicants aged 17 to 40, expanded to $3 million and ages up to 59, and subsequently included all standard-and-above risk classes.
At the time of publication, the authors reported that algorithmic underwriting applied to 90% of applications within those parameters.
The rollout offers a practical rule for leaders: expand automation in stages, comparing decisions against an established baseline before widening eligibility.
The paper reports gains in risk selection, processing speed, policy take-up, and operational efficiency, although the findings combine historical simulation with deployment results and do not apply uniformly to all applicants:
- Risk selection: A simulation using historical data produced 6% fewer deaths in the model-selected best non-tobacco class than in the underwriter-selected class over a 15-year horizon. This measures risk selection; it does not mean the model prevented deaths.
- Speed: Time to issue a policy fell by more than 25% for certain applicants, rather than across all applications.
- Take-up: Applicants receiving a model-based best-class decision were more than 30% more likely to purchase their policies than those receiving a traditional best-class decision. The comparison does not establish that the model alone caused the increase.
- Operational savings: The authors reported millions of dollars in efficiency savings over two years, without a precise figure or calculation method.
The technology also reached the external 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 in its Magnum underwriting system. This establishes historical commercialization, rather than the status of the product today.
MassMutual’s current underwriting disclosure describes how applicant eligibility determines access to algorithmic assessment with or without medical examinations and lab tests:
- Algorithmic underwriting: Available for eligible individual term and whole life applications up to $3 million for applicants aged 17 to 59. An examination and lab tests remain required.
- MassExpress Fluidless: Uses AI and external consumer data to assess eligibility for underwriting without an examination or lab tests, for applicants aged 17 to 50 seeking up to $3 million in coverage.
Both programs have additional eligibility restrictions and product exclusions. The disclosure also identifies electronic health information, medical claims, pharmacy records, and credit-related information among the data available to its AI systems.
Meeting the published criteria does not guarantee a fluidless offer. Some applicants still require an examination and lab tests, and some otherwise eligible MassExpress applicants are selected at random for control and audit purposes.
For leaders, the continuing lesson is to preserve a way to evaluate automated decisions after deployment. MassExpress’s random holdouts illustrate one control that can support that evaluation, alongside defined eligibility rules and review of exceptions.
For leaders pursuing similar gains, clear decision ownership, staged automation, and investment in usable data help improve efficiency while keeping operational risk in view:
- Clear decision ownership — Assign responsibility for checking AI outputs and escalating uncertain cases so faster service preserves accountability.
- Staged automation — Compare automated decisions with an established human baseline before widening eligibility, so problems can be identified while the scope remains manageable.
- Data readiness — Budget for digitization and data organization before model deployment, because reliable inputs determine which workflows AI can support.
