From Data to Decisions: The Claims You Can’t See Coming

By James Benham, Co-Founder, CEO, JBK, and author of Be Your Own VC

How Routine Workers’ Comp Claims Turn into Runaway Costs

When a claim crosses an adjuster’s desk on a normal Tuesday, it usually looks routine. A warehouse worker with years on the job reports aching shoulders and a sore back. There was no dramatic accident, no ambulance, no lost time yet. It looks like one of the hundred files already in the queue, so it gets handled like one.

Eleven months later, that same claim has three specialists, a disputed diagnosis, an attorney, and a reserve that has been revised upward four times. Nobody made an obvious mistake. The warning signs were there the whole time, scattered across a dozen notes, forms, and separate systems, and no one had the minutes to connect them while the claim still looked ordinary.

Why Experience Alone Can’t Keep Up

For a long time, catching those claims was a job we left to instinct. A veteran adjuster would get a feeling about a file and dig in. That still matters, but instinct alone no longer scales. Claims teams are carrying more complexity per file than ever, driven by an aging workforce, existing conditions like diabetes and hypertension, and psychosocial factors that quietly decide how long someone stays out. Few adjusters have time to cross-reference every small update against decades of history.

This is where a good predictive model earns its place. On its own, a slightly delayed report means little. So does one missed therapy appointment, or one existing health condition. Put hundreds of those small signals together, though, and a pattern appears that no one could track by hand. The model’s work is not to raise a hand and walk away. It names the drivers behind the pattern: the reporting delay, the widening gap between visits, the existing health condition working against the injury, the course of care that has quietly drifted off guideline. The adjuster opens the file already knowing what is pulling it toward lost time or litigation, and that is the difference between an alert and an explanation.

Evaluating Claims Based on Historical Patterns, Not Just Hindsight

Programs go wrong when they treat the prediction as the product. In many claims shops, analytics still means a dashboard, a rear-view mirror that reports what already happened last quarter. Predictive analytics is a different animal. It weighs a newly reported injury against patterns from thousands of past claims and estimates what it is likely to do next, whether a no-lost-time claim is heading toward lost time, or a routine file is quietly drifting toward litigation.

A score on its own is not the product, and it is nowhere near the limit of what a well-built model can tell you. The useful part is the reasoning underneath it: which factors are pushing the claim in that direction, how much weight each one carries, and what that combination has meant on files like it before. That is what turns a number into a decision someone can make this week instead of next quarter. Analytics that do not change what someone does on Monday morning are just expensive trivia.

Mitigating Severity with Early Intervention and Cost Avoidance

We need to get this right because the margin for error keeps shrinking. Cumulative trauma claims, the ones with no single accident date, accounted for 26 percent of California indemnity claims in 2024. 82 percent of those filed while the worker is still employed are litigated, and 99 percent of those filed after the worker has left, according to the Workers’ Compensation Insurance Rating Bureau of California. The friction costs follow the litigation. Medical-legal services now make up 37 percent of paid medical costs on cumulative trauma claims for accident years 2022 through 2024, up from 24 percent in the 2017 through 2019 period. The California Workers’ Compensation Institute has tracked the same pattern from the litigation side: cumulative trauma grew from 29.4 percent of all litigated claims in 2010 to 37.5 percent in 2022. These are the files that look ordinary at intake and expensive two years later.

In claims, the biggest wins are the ones that never show up on a report, because they arrive as costs you never have to spend: the litigated claim that never happened, the disability that never dragged on. Catch a no-lost-time claim that is trending toward lost time, act on it early, and you can change the entire financial trajectory of that file. The fix is rarely exotic, usually a nurse call, an ergonomic adjustment, or an early return-to-work conversation. What matters is timing. The same outreach that changes everything in week one does almost nothing in month six, once the claim has hardened, an attorney is involved, and the worker has spent months convinced no one was ever on their side.

Medicine is part of this too. Pharmacy and treatment analytics can catch an opioid regimen trending toward long-term use, or a course of care drifting off evidence-based guidelines, early enough for a clinical peer to step in. And the spend is more concentrated than most people assume. In the first half of 2024, a single drug, the topical painkiller lidocaine, accounted for 8.5 percent of all drug payments in California’s workers’ comp system, ranking first among every medication prescribed, according to the California Workers’ Compensation Institute. When a handful of drugs drives that much of the cost, spotting the outliers early is prevention, not paperwork.

Solving the Core Data Quality Bottleneck in Claims Systems

None of this runs on autopilot, and I would be skeptical of anyone who tells you it does. A model fed by half-empty intake fields and inconsistent records will produce confident nonsense at scale. Clean, reliable data is the boring prerequisite that rarely makes the sales slide, and it is usually the real reason one carrier’s models outperform another’s.

So before you expect an algorithm to fix your escalation rates, look hard at what you are feeding it. You cannot predict the future on top of a disorganized past.

Operationalizing Risk Signals in Adjuster Workflows

Every prediction needs an owner and a next step. If a model flags a claim as high risk, explains exactly why, and nothing in the workflow says who acts, how, and by when, you have not bought insight. You have bought anxiety. What you need is a system: each signal attached to a specific action, a nurse assignment, a supervisor review, a call to the worker that same day.

And you have to measure the intervention, not the algorithm. A model’s accuracy score is the vendor’s concern. Yours is whether acting early actually shortened durations, reduced litigation, and got people back to work, and whether your team trusts the reasoning enough to act on it at all. Adjusters do not act on a number they cannot interrogate. They act on a case the system can make.

Optimizing Adjuster Bandwidth for High-Value Claims

In any claims operation, attention is the most valuable resource you have, and complexity is rising faster than you can add headcount. The scarcest thing in the building is an experienced adjuster’s focus. Used honestly, predictive analytics is how you point your best people at the handful of claims where showing up early changes the outcome, instead of spreading them evenly across a stack of files, most of which would have been fine either way.