Where AI Actually Works in Workers’ Comp — And Where It Shouldn’t 

By James Benham, Co-Founder and CEO, JBK 

In some claims offices, AI now writes the first claim note, flags high-risk files, and even summarizes doctor reports, before the adjuster picks up the phone. AI in workers’ comp is no longer experimental; it’s operational. 

The recent AI boom in workers' comp has not gone unnoticed, and industry surveys back that up. EY’s recent global research on insurance found that most carriers are now actively piloting or deploying generative AI use cases, with the early “frontrunners” focusing on operational workflows rather than side experiments. 

In my experience, the biggest, most reliable gains with AI deployment in workers’ comp show up in three very practical places: 

• Intake and First Notice of Loss (FNOL) 
   

• Early Triage and Reserve Support 
   

• Medical Record Summarization 

When those three areas improve, adjusters get hours back in their week. They can call injured workers sooner, explain next steps more clearly, and coordinate return-to-work instead of fighting their inbox. That’s where AI starts to move the needle on cost, duration, and claimant experience at the same time. 

Intake & First Notice of Loss (FNOL) 

FNOL has always been a fragile moment in workers’ comp. Information arrives through phone calls, emails, portals, and even paper. Details are incomplete, terminology varies, and people are trying to capture everything while a supervisor is at their desk or a worker is in pain. 

Small errors at this stage can follow a claim for its entire life. For instance, a mis-coded body part, a missing employer contact, or even an unclear mechanism of injury can later surface as a disagreement about what treatment is “related” or whether the claim was reported correctly. AI is genuinely useful at this front door. 

Modern FNOL tools can now turn unstructured conversations into structured data. They listen to recorded calls or parse email and web submissions, then extract key elements—such as employer information, injury details, location, witnesses, wage information, and push them straight into the claim system. That means fewer rekeying errors, fewer blank fields, and fewer downstream corrections. 

The second job is real-time validation and routing. Claims AI “agents” and workflow engines can check coverage, suggest the appropriate team or adjuster, and flag claims that may warrant early nurse involvement or telehealth based on what’s captured at FNOL. That speed isn’t just nice to have. A Monash University study of workers’ compensation systems has shown that longer claim-processing times are associated with longer disability duration and poorer return-to-work outcomes, even after adjusting for injury severity. 

From what I’ve seen, stabilizing intake is one of the fastest ways to feel AI’s impact. You get cleaner data from day one and fewer surprises later. Adjusters begin with files that already make sense, rather than spending their first 20 minutes recreating what should have been captured at FNOL. 

What AI should not replace here is the first human contact. 

Research summarized by Workers Compensation Research Institute (WCRI) and others shows that injured workers often seek attorneys not because they’re afraid of being fired. They believe their employer questions the legitimacy of the injury or thinks their claim has been denied. Those are communication failures, not legal inevitabilities. AI can help capture what happened. The adjuster helps the worker understand what happens next. 

Triage & Reserve Support: Better Signals, Not Auto-Decisions 

Once a claim is set up, the problem changes to “Which files need attention first?” 

Predictive analytics has been part of workers’ comp for years, but pairing richer data with modern ML models is making it much more useful. Carriers and TPAs are utilizing models to estimate the likelihood of complexity, prolonged duration, or attorney involvement based on factors such as injury type, jurisdiction, comorbidities, and employer history. 

A daily worklist that was previously sorted by “oldest first” can now be ordered by risk. A claim that looks routine on the surface can be flagged for early nurse case management. A seemingly minor shoulder injury may be highlighted because it matches the pattern of prior claims that have produced high indemnity and a prolonged time off work. 

This is where AI can help: 

• Point experienced people to the files that matter most 
   

• Suggest more realistic reserve ranges based on cohorts of similar claims 
   

• Highlight “drift” when a claim starts deviating from expected recovery patterns 
   

Where human judgment stays non-negotiable is what you do with that information. Decisions about compensability, benefit levels, settlement strategy, significant procedures, or legal escalation affect people’s livelihoods. Regulators and industry frameworks for “responsible AI” are clear: high-stakes insurance decisions need transparency, audit trails, and accountable human oversight. 

So the way I think about triage AI is simple: it’s a sharp second set of eyes on every file. It helps you see patterns and priorities you’d otherwise miss. It does not get the final word on benefits or coverage. 

Medical Record Summaries: The Quiet Workhorse 

If there’s one area where AI is already changing the workday for adjusters and nurses, it’s documentation: medical records, claim notes, emails, and diaries. 

Complex workers’ comp claims generate an enormous amount of text. Clinic notes, imaging reports, therapy updates, nurse calls, employer emails, and adjuster diaries all pile up. Manually turning that into a clear story is slow, inconsistent, and exhausting. It’s also exactly the kind of pattern-recognition work AI is good at. 

That’s why we built Terra’s AI Note Summarization to sit directly inside the claims system. Terra ingests long-form case notes, email threads, and adjuster diaries and condenses them into concise, actionable summaries, using models trained specifically on workers’ comp and P&C terminology. The goal is simple: give adjusters, supervisors, and managers a fast way to see what’s happened on a claim without reading pages of free text.  

We pair that with Terra’s broader automation and OCR-driven ingestion, so the same platform that centralizes claims workflows also structures the underlying documentation. In our partnership with Gradient AI, those summarized notes and extracted data become even more powerful: Gradient’s risk models plug into Terra’s native tools (OCR, claim-note summarization, tasking) to help identify high-risk cases earlier and support faster, better-informed decisions.  

The pattern is the same whether you’re looking at medical records, IME reports, or long claim diaries: let AI handle the heavy lifting of sorting, extracting, and summarizing, and let humans handle judgment and strategy. Terra’s own content has repeatedly made this point: AI can automate data entry from medical records and other documents, but the real value comes when that automation is integrated into the workflow, allowing adjusters actually to regain time.  

In practice, this kind of summarization changes the nurse's starting point. Instead of opening a file and facing dozens of unstructured notes, they start with a clear timeline and key issues, then verify, correct, and decide. That shift can turn hours of document review into minutes—and that reclaimed time is what allows people to focus on conversations with providers, employers, and injured workers, where outcomes are truly shaped. 

Bandwidth, Attorneys, and Return to Work 

Why does all of this matter beyond a productivity slide? 

First, because duration drives cost. Studies of workers’ comp and related systems show that longer claim processing times and more adversarial claim experiences are associated with longer disability and worse return-to-work outcomes. 

Second, because attorney involvement is expensive. A recent WCRI study of over 950,000 lost-time claims found that approximately 34% of workers absent for more than seven days had an attorney, and their total costs were roughly $7,700 to $12,400 higher on average than those of similar non-represented claims. Again, WCRI’s “Avoiding Litigation” work ties much of that representation back to fear, confusion, and perceived threats, not just severity. 

AI alone doesn’t make the system warmer. But by stripping away hours of clerical work, it gives adjusters and nurses bandwidth to call early, explain clearly, and follow up consistently, the exact behaviors that reduce fear and confusion. 

There’s even some early evidence outside of computer science that AI-assisted communication can be more consistently empathetic and straightforward. Allstate, for example, reported that AI-drafted emails, checked by humans for accuracy, were perceived as clearer and less jargony than those written from scratch by reps.  

Put simply: good AI makes it easier for humans to show up well. 

Where Human-in-the-Loop Is Non-Negotiable 

It’s tempting, especially under cost pressure, to look at a strong model and wonder if it can “just handle” whole classes of decisions. In workers’ comp, it’s misaligned with what claimants, employers, and regulators expect. 

Certain decisions need to remain firmly human-owned: 

• Whether an injury is work-related and compensable 
   

• How benefits are structured and adjusted over time 
   

• When to authorize significant procedures or IMEs 
   

• When to involve legal counsel or pursue alternative resolutions 
   

They involve questions of fairness, trust, and long-term impact that demand human judgment. Responsible AI frameworks in insurance consistently emphasize transparency, governance, and clear human accountability at these decision points. 

A useful mindset is “AI + HI” artificial intelligence plus human intelligence. AI highlights patterns, consolidates information, and suggests where to look next. Humans decide what is appropriate, ethical, and aligned with both policy and values. 

Any AI roadmap in workers’ comp should make that division of labor explicit. 

A Practical Way to Start 

For claims and risk leaders who want to move from theory to practice, the most successful paths I’ve seen are focused and incremental. 

A simple sequence looks like this: 

1. Stabilize intake: Utilize AI-enabled FNOL capture across your primary channels to minimize manual rekeying and data errors, and track the time elapsed from injury to the first meaningful contact. 
   ﷟HYPERLINK "https://www.moxo.com/blog/fnol-workflow-intake-triage" 

2. Pilot medical summarization on heavy claims: Start with high-volume, complex injuries where record review is a bottleneck. Let AI compile, and have nurses or specialists validate. Track time saved and the impact on reserve accuracy and RTW planning.¹⁰ 
    

3. Layer in triage signals: Introduce predictive models to help prioritize worklists and flag outliers. Keep the final decisions with human owners, and give adjusters a way to provide feedback so the models improve over time.⁹ 
    

4. Document human-in-the-loop points: Be explicit about which decisions AI may support and which require human sign-off. That clarity is essential for good governance and effective communication, to both your own teams and regulators. 
    

None of this requires a moonshot. It’s about making specific parts of the claims journey smoother and more informed, then giving adjusters and nurses the time to do what only they can do. 

AI in workers’ comp is ultimately a set of tools that, when applied to intake, triage, and medical summarization, can quietly transform the experience of a claim for everyone involved. 

For adjusters and nurses, it means more time on meaningful work and less time on repetitive tasks. For injured workers, it means faster contact, clearer explanations, and fewer unpleasant surprises. For employers and carriers, it means less leakage, better reserve accuracy, and a path to lower total cost of risk that doesn’t depend on burning out the people doing the work. 

Used that way, AI will complement human intervention and not replace it. 

James Benham 

James Benham is the Co-Founder and CEO of JBK, a technology and consulting firm serving the insurance and risk management industry. He also hosts The InsurTech Geek podcast and is a frequent speaker on insurance innovation. 

Share This Article: