Forty-three per cent (43%) of US workers say they are never sure they are choosing the right benefits during enrollment, according to The Hartford's 2026 Future of Benefits Study. Seventeen percent used AI to help with benefits decisions at their most recent open enrollment, with more than half of those users in the Gen Z cohort. The enrollment uncertainty is not new. What is new is the data infrastructure that can address it - and a growing body of evidence that the problem has never been one of product availability.
"What we saw after the pandemic is that employers had to step into the healthcare gap," said Nicholas McMenemy, region leader and managing director, UK, at Marsh Health and Benefits. "Employees' expectations of what their employer can deliver from a healthcare perspective have massively increased, and employers are struggling to keep up with that."
The observations come from Episode 6 of Marsh Health and Benefits' Transforming Benefits with Technology vodcast series, in which McMenemy and Andrew Owens, chief technology officer at Marsh Health and Benefits, made the case that the benefits relevance problem is a data problem rather than a design one. The benefits exist. They are not reaching the right people.
The standard approach to benefits design groups employees by demographic category - age band, life stage, family status - and builds programmes around those groupings. The problem is variation within categories. Two employees in the same age band and family structure may have materially different health priorities, geographic constraints, and financial situations. A programme designed for the average of the group serves the average poorly.
Owens framed the AI opportunity in terms of personalisation that adjusts to context rather than category. "If I live in the middle of the countryside, I don't want to be offered a gym," he said. "There are no gyms around. I want something alternative. So that's about understanding the context of the employee, what's relevant, and bringing that overall picture and offering to them something that feels personalized and relevant."
McMenemy pointed to an equity dimension in the same argument. Certain demographic groups are significantly more likely to access healthcare remotely than through traditional in-person routes. AI-powered benefits that meet employees where they already are may improve health outcomes for groups that conventional programmes have historically underserved - not by adding new products, but by changing how existing ones are surfaced and explained.
The Hartford's study found that 95% of employers are already taking steps to improve their open enrollment experience, with many deploying AI-driven recommendation engines to help employees choose appropriate coverage. The gap the data reveals is not between employers who care about benefits relevance and those who do not. It is between those who have the claims data and behavioural signals to act on that intention and those who do not.
The personalisation argument depends entirely on data quality. Owens was direct about the tension. The predictive output is only as useful as the personal information employees are willing to share - and employee trust in how that data will be used is not automatic.
He said the technical safeguards for keeping health and personal data within an organisation are well established. What matters equally is transparency with employees about why data is being collected and what it will be used for. "Organizations are pretty well versed now in the utilization of personally identified information and personal health information," he said. "The foundation is there. All organizations need to understand how to share that information in a responsible way."
On measurement, Owens pushed back against purely metric-driven approaches to evaluating benefit programme success. Utilisation rates capture whether an employee accessed a benefit. They do not capture whether it made a difference. "In the end, you want it to be the employee who feels that it's actually bringing benefit to them," he said. "The best thing to do is you ask them."
That distinction matters for brokers helping employers evaluate their programmes. A benefit with low utilisation may be poorly designed - or it may be well-designed but invisible to the employees it was built for. Those are different problems with different solutions, and the data required to distinguish them is different in each case.
The most concrete illustration in the conversation came from the intersection of fleet monitoring and occupational health. McMenemy described how telematics data - originally deployed to track driver compliance in logistics and automotive fleets - is now being used to identify health trends across a workforce. Patterns that once read as safety signals can point to underlying conditions: back problems, sleep difficulties, physical strain from repetitive motion. That reframing turns monitoring data into a healthcare signal.
"We can then have a really intelligent conversation around risk reduction for insurers," McMenemy said, "which then benefits the employer because it adjusts their premium and makes sure they're only paying for the risks that are really affecting their workforce."
The example illustrates what genuinely integrated benefits measurement looks like - where the data informing plan design is drawn from actual workforce behaviour rather than demographic proxies, and where the output of that analysis connects directly to carrier pricing rather than sitting in a separate HR analytics function. For brokers working with employers on benefits programme design, that integration is the practical standard the market is moving toward. The question for each employer client is how far their current data infrastructure sits from it - and what closing that gap would require.
