AI Is Deciding Workers' Comp Claims — and Sitting on Three Statutes at Once
Health insurance AI has taken the public heat. Workers' compensation has quietly automated further and been examined less. Predictive triage, reserve setting, return-to-work forecasting and automated utilization review now shape what an injured worker receives — decisions that land simultaneously on disability accommodation duties, state insurance regulation and the bad-faith standard that governs claim handling.
The Population Makes This Different
Most algorithmic-discrimination analysis starts by asking whether a protected class is disproportionately affected. In workers' compensation the question is inverted: the entire population consists of people with medical conditions arising from work, many of which meet the definition of a disability. A model that treats some conditions systematically worse than others is not producing a subtle disparate impact on a minority subgroup — it is sorting a protected population by exactly the characteristic the law is most attentive to.
That matters most for the conditions with the weakest objective imaging and the longest historical record of skepticism: soft-tissue injury, chronic pain, repetitive strain, and psychological claims. A model trained on how those claims were handled in the past learns the skepticism and returns it as a number, and a number is far more persuasive inside an organization than an adjuster's hunch ever was.
Where the Models Sit in the Claim Lifecycle
- •Attorney representation and prior claims used as risk features
- •Zip code and employer industry act as proxies for protected traits
- •High scores trigger surveillance and independent examinations
- •Routing decisions rarely documented as decisions at all
- •Language preference correlates with national origin
- •Medical necessity determinations need credentialed human reviewers
- •Guideline engines applied without individualized consideration
- •Batch denials produce identical reasoning across dissimilar files
- •Appeal explanations that no claimant can meaningfully contest
- •Turnaround metrics that reward speed over the record
- •Forecast duration used to pressure premature release
- •Accommodation duties are individualized; a population model is not
- •Modified-duty offers built from a score rather than a restriction
- •Age and comorbidity features carry obvious proxy risk
- •Interactive-process obligations survive any prediction
- •Settlement offers anchored to modeled willingness to accept
- •Historical settlement data encodes prior disparities
- •Represented and unrepresented claimants scored differently
- •Reserve pressure quietly shapes authorization behavior
- •Model rationale rarely reproducible six months later
Delay Is the Under-Examined Exposure
Compliance reviews concentrate on denials because denials generate letters, appeal rights and audit trails. The model's larger operational effect is usually upstream of that: which files get expedited, which get an extra layer of review, which sit in a queue awaiting a second opinion. None of those produce an adverse-determination notice, and all of them change what the injured worker actually receives. Bad-faith standards in most states reach unreasonable delay in investigating and paying, which means a routing algorithm can create liability without ever issuing a decision.
This is also where disparate impact hides best. A fairness review that measures denial rates by demographic group will report clean numbers while median time-to-authorization differs sharply across the same groups. Measure the delay distribution, not just the outcome distribution.
Three Regimes, One Deployment
Disability law imposes individualized assessment and interactive-process duties that a population-level prediction cannot satisfy on its own. State insurance regulation increasingly requires documented governance of models used in claim determinations, credentialed human review for medical necessity, and accountability that does not transfer to a vendor. Comprehensive state privacy laws add access, correction and profiling opt-out rights over automated decisions with significant effects. A single claims platform can be simultaneously non-compliant under all three while every individual configuration looks reasonable to the team that owns it.
Governance Checklist
Before Deployment
- ☐Inventory every model touching a claim, including vendor features enabled by default
- ☐Demand training-data documentation and known-limitation disclosures in the contract
- ☐Strip proxy features that add little signal and carry obvious protected-trait correlation
- ☐Confirm credentialed human reviewers own every medical-necessity determination
- ☐Write the appeal explanation template before the model goes live
In Ongoing Operation
- ☐Monitor time-to-authorization and denial rates by condition type and demographics
- ☐Track override rates — a reviewer who never overrides is documenting a rubber stamp
- ☐Log the model version and inputs behind every adverse determination
- ☐Re-audit after each vendor model update, not on an annual calendar
- ☐Give claimants a route to see and contest the data used in their file
Frequently Asked Questions
Can we let the model issue denials directly?
That is the configuration most likely to draw regulatory attention and the hardest to defend. Medical-necessity determinations generally require an appropriately credentialed reviewer, and accountability for an adverse determination stays with the licensed entity. Use models to prioritize and inform; keep the determination with a person who can explain it.
We removed race and gender from the features. Is the model neutral?
Removing the direct variables removes the easiest evidence, not the disparity. Zip code, industry, language preference, attorney representation and claim history all carry protected-trait signal. The only way to know is to measure outcomes and delays across groups after the fact, which requires retaining enough demographic data to run the test.
Does a fraud-propensity score create legal risk if it never denies anything?
Yes, because of what it triggers. A high score routes claimants toward surveillance, additional examinations and slower authorization, and unreasonable delay in investigating or paying is itself actionable in most states. Treat routing effects as decisions and monitor them as such.
Our claims platform is a third-party product. Who is responsible?
You are, for the determinations made under your license and your duty to the claimant. Vendor indemnity is worth negotiating and does not answer a market-conduct examination. Require model documentation, change notification and audit rights before signing rather than after the first complaint.
How does return-to-work prediction interact with accommodation duties?
Poorly, if the prediction drives the offer. Accommodation obligations are individualized and interactive; a population-derived duration forecast is neither. Use the forecast for capacity planning and keep the modified-duty conversation anchored to the treating provider's actual restrictions.
What does a regulator or plaintiff ask for first?
The audit trail. Which model version scored the file, what inputs it saw, who reviewed the recommendation, how long they spent, and how often that reviewer overrides. If those fields are not logged, the absence itself becomes the story. Instrument the logging before you need it.
Your Claims Portal Is Part of the Record
Injured workers file, upload documents and track decisions through your website. A portal that is hard to use, or a public page that describes a claim process differently from how it actually runs, becomes evidence in a dispute about how claimants were treated.
See how your site holds up. Run a free scan and review the pages claimants use to interact with your process.