Dollars & Sense
Predictive tools can identify expensive workers’ compensation claims. Whether the response improves recovery and reduces total costs demands a different kind of proof.
A workers’ compensation claim receives a high-risk score. The next decision may involve a nurse case manager, medical review, specialist evaluation or closer claims oversight. Each response may be appropriate and some may prevent years of avoidable disability. Each can also generate another expense in an already complex system. As claims analytics become more sophisticated, we need to ask whether the services triggered by a prediction actually change an injured worker’s outcome. A score identifies a possibility. Its financial and clinical value depends on what happens next.
What are we actually predicting?
Carriers need credible estimates of future liabilities. Third-party administrators need to allocate professional attention and employers want to identify barriers to recovery before they become more complicated. Predictive analytics can support these objectives by identifying patterns across historical claims and estimating the likelihood of future costs or disability. But predicting total incurred cost is different from predicting avoidable disability or recovery potential. A model may estimate expenditures accurately without establishing whether a particular intervention improves care. The outcome selected for prediction matters as much as the model’s statistical accuracy. If our objective is better recovery at a sustainable cost, we need evidence connecting the prediction to a meaningful change in what happens afterward.
When a Flag Becomes Revenue
The economics become more complicated when a high-risk label prompts additional services. In a 2026 Securities and Exchange Commission filing, Pacific Health Care Organization described workers’ compensation revenue from medical bill review, utilization review and case management. Its payment arrangements include fixed fees, hourly charges and certain fees calculated as percentages of savings. These are legitimate commercial structures and the services may provide substantial value. They also establish that additional claims-management activity can generate revenue for the organizations providing it. When predictive technology directs more claims toward these services, expenditures may increase before the financial or clinical benefits become apparent. That possibility deserves measurement rather than an assumption of wrongdoing.
The financial impact also depends on where these expenses appear. Some vendor services are charged directly to individual claim files as allocated loss adjustment expense (ALAE), while others are included in administrative fees or different expense categories. For employers participating in loss-sensitive insurance arrangements, the classification and allocation of these charges can influence their ultimate financial responsibility. Self-insured employers may bear these expenses directly. A service that reduces medical payments may still produce a smaller net financial benefit once its fees are included. Employers and claims leaders should therefore examine the total cost of responding to a prediction rather than evaluating each service in isolation.
Evidence After the Referral
A University of Washington evaluation provides an important caution, although it examined nurse case management rather than AI referrals. Researchers studied a pilot program serving injured workers with catastrophic injuries. Participants reported high satisfaction with nurse case-management services, but medical costs excluding case-management fees and time-loss duration did not change significantly following implementation. In a small analysis involving 15 accepted cases from one outcome-based provider, the average case-level ratio of case-management payments to medical expenses was 1.8. Other providers had substantially lower relative case-management costs. Differences in injury severity, the limited observation period and the study design prevent definitive conclusions about financial effectiveness. Nevertheless, the findings demonstrate why satisfaction, service delivery and measurable financial improvement require separate evaluation.
Washington also provides encouraging evidence. Thomas Wickizer, Gary Franklin and Deborah Fulton-Kehoe evaluated the Centers of Occupational Health and Education (COHE), which emphasized coordinated treatment, occupational best practices and return-to-work support. Their eight-year follow-up found approximately 30% lower odds of long-term work disability among injured workers treated within the COHE model compared with the study’s comparison group. This was a nonrandomized evaluation rather than an AI experiment, so the findings cannot establish that predictive algorithms deliver the same results. They do demonstrate the potential value of organizing occupational healthcare around interventions designed to reduce long-term disability. The distinction is important because identifying an injured worker who may experience prolonged disability creates an opportunity. The effectiveness of the response determines whether that opportunity becomes a better outcome.
Financial incentives provide another reason to examine what happens after a referral. Research by Dongchun Wang at the Workers Compensation Research Institute documented that physician-dispensed medications frequently cost substantially more than identical medications dispensed through retail pharmacies. The study also examined differences in pricing and the effects of reimbursement reforms on dispensing practices. These findings do not establish that predictive analytics cause unnecessary utilization or that healthcare professionals are motivated primarily by revenue. They demonstrate that payment arrangements can influence the economics of medical services. When organizations introduce predictive tools that generate referrals or additional reviews, reimbursement structures deserve consideration alongside clinical necessity and expected benefits.
Proving the Return
A credible return-on-investment calculation cannot simply subtract actual claim costs from a model’s forecast. Predictions contain uncertainty and the most complex cases are often selected for additional services. An injured worker’s final cost falling below a predicted amount does not establish that the intervention produced the difference. Comparing managed and unmanaged claims without accounting for injury severity can be equally misleading. Evaluation requires an appropriate comparison group, clearly documented intervention criteria and sufficient follow-up to capture meaningful disability outcomes. The calculation should include medical payments, indemnity and relevant administrative expenses rather than excluding the costs incurred to produce the reported savings.
The measurement problem extends beyond financial reporting. An intervention that reduces immediate spending may create additional expenses if treatment delays prolong disability or interfere with a sustainable return to work. Conversely, additional early spending may be justified when appropriate care prevents complications and improves long-term function. Claims professionals need to distinguish between expenses that contribute to recovery and activity that adds complexity without measurable benefit. Employers and carriers should evaluate outcomes over time rather than relying exclusively on short-term financial projections. The objective is to determine whether the intervention improved the trajectory of the claim and whether the resulting value justified its total cost.
Before expanding a predictive program, employers, carriers and TPAs should require clear answers about what the technology predicts and which actions its recommendations trigger. Vendors should disclose relevant affiliations, service pricing and the methodology used to calculate reported savings. Claims professionals need authority to question recommendations when their direct knowledge of an injured worker suggests another approach. Leaders should compare outcomes across similarly situated claims and distinguish improvements in reserve accuracy from reductions in actual expenditures. Financial reports should be examined alongside the injured worker’s functional progress, treatment experience and ability to return to meaningful work. These expectations provide a foundation for evaluating predictive technology without dismissing its potential or overlooking its limitations.
Prediction offers a valuable opportunity to intervene sooner and allocate resources more effectively. It can also make an expensive process more elaborate without improving the result. Both possibilities deserve rigorous evaluation rather than a predetermined verdict. The essential measure is the complete relationship between prediction, intervention, expenditure and recovery. Technology becomes valuable when organizations can demonstrate that the decisions it influenced produced meaningful improvements. The business of prediction succeeds when we can explain what improved for the injured worker and what it actually cost to achieve that improvement.
Research Referenced
Fulton-Kehoe, D., Le, V., Sears, J., Smith-Weller, T., Wickizer, T., & Wu, R. (2020). Catastrophic Care Management Evaluation: Final Report. University of Washington School of Public Health. Full Evaluation.
Pacific Health Care Organization, Inc. (2026). Form 10-Q, Quarterly Period Ended March 31, 2026. U.S. Securities and Exchange Commission. SEC Filing.
Wang, D. (2012). Physician Dispensing in Workers’ Compensation. Workers Compensation Research Institute. WCRI Research.
Wickizer, T. M., Franklin, G. M., & Fulton-Kehoe, D. (2018). Innovations in Occupational Health Care Delivery Can Prevent Entry into Permanent Disability: 8-Year Follow-up of the Washington State Centers for Occupational Health and Education. Medical Care, 56(12), 1018–1023. Full Study.
