A June 2026 NBER working paper examines whether employee reviews can reveal information about product and service quality before problems become visible externally. Wei Cai, Dennis Campbell, Yaxuan Chen, Yufei Chen and Andrea Prat use more than 4.3 million Glassdoor reviews to construct firm-level quality measures for S&P 1500 companies covering 2008 to 2023.

The authors train machine-learning models using customer satisfaction information for a subset of firms and assess predictive performance outside the training sample. Their measures are associated with subsequent quality provision and outcomes including product recalls, brand value and profitability. The research concerns predictive information. It does not demonstrate that improving benefits causes higher product quality, or that a company’s average Glassdoor score alone supplies an equivalent measure.

Relevance for HR and benefits governance

For HR teams, the study raises a useful question about how employee observations enter management decisions. Employees may describe operational processes, service pressures and obstacles that are difficult to capture in a satisfaction score. An employer can consider whether its internal listening arrangements make those observations accessible to the people responsible for reviewing them.

This is an editorial application of the research, rather than an intervention tested by the paper. It does not require a benefits team to adopt the authors’ model or to treat anonymous reviews as verified evidence. The appropriate response is to distinguish a signal that warrants investigation from a fact that has been independently established.

Applying the distinction to suppliers

Benefits procurement can make the same distinction when assessing providers. Employee or customer feedback may suggest a question about service quality, but a purchasing decision can also require contract information, service records and a documented response from the provider. Several sources of information can be considered together without pretending that they measure the same thing.

The transfer from large US-listed companies to a specialist benefits provider also needs care. Coverage, review volume and the nature of the service may differ. A predictive relationship in the research sample does not automatically establish the same relationship for a smaller organization or another country.

The practical lesson is about the design of a review process. Teams can record what an observation alleges, how it was checked, and what action followed. That approach preserves the value of employee voices while making uncertainty visible. The paper offers a reason to take those voices seriously as a source of information, with conclusions kept within the limits of the evidence.

Sources: Source de référence