Workers’ compensation has never been a low-stakes industry. But as AI tools proliferate across carriers, employers, and pharmacy benefit managers, industry leaders Paul King and Cliff Belliveau worry that the industry is accumulating a new kind of risk: one hiding in plain sight.

Belliveau, Chief Innovation Officer at MyMatrixx by Evernorth, has made it his business to understand what’s happening beneath the AI marketing surface — and what he finds consistently surprises him. At industry conferences, he poses two simple questions to his audiences. First: how many of you are aware of your organization’s formal AI governance frameworks? Almost no hands go up. Second: how many of you use AI tools? Nearly every hand rises.

“The confidence people feel is largely fueled by the B2C side of AI,” Belliveau said, referring to tools like ChatGPT or Gemini that anyone can pick up and use. But the consumer familiarity with those tools does not equal a competency or readiness for using AI to make high-stakes decisions about pharmacy recommendations, claims management, and patient safety.

“We hear about this every day,” said Paul King, President of MyMatrixx by Evernorth. “The industry is being influenced by buzzwords and promises of AI solutions to simplify claim management processes. But AI is not simple. There is a real gap between talking about AI and implementing an AI solution,” King continued.

For workers’ comp carriers and employers, that gap between perceived AI understanding and actual governance isn’t just an efficiency concern. It’s a risk management problem, and MyMatrixx has a framework for addressing it.

Understanding the Real State of AI Adoption

Cliff Belliveau, Chief Innovation Officer, MyMatrixx by Evernorth

The confidence gap stems from multiple sources. The technology industry’s famous hype cycle has created unrealistic expectations, while FOMO — the fear of missing out — drives decision-makers to adopt AI solutions without fully understanding them.

These days, every company has become an AI company, whether their product is rubber washers or analytic solutions. The workers’ compensation industry is no exception, said Belliveau. “That really captures where things stand right now,” he said.

The result is an industry full of organizations deploying technology they may not fully comprehend, governed by frameworks that may not exist, and managed by leaders who may not know the right questions to ask.

Knowing What You’re Actually Buying

Paul King, President, MyMatrixx by Evernorth

For procurement professionals and technology leaders evaluating AI vendors, three critical questions separate genuine AI solutions from impressive-sounding marketing claims.

First, understand the company’s foundation. Ask vendors directly: How did you get into AI? Was the founder a technologist, or someone influenced by science fiction? Belliveau emphasized the importance of distinguishing between companies with genuine technical depth and those making opportunistic pivots. “Pure-play analytics companies have made a very natural soft pivot into AI,” he explained. “If you already have a strong data management foundation, the transition is quite seamless. The challenge arises when a company never had a real data program to begin with.”

Belliveau also recommended bringing technical expertise into vendor meetings. “Bring your expert to the meeting, someone who really knows the technology,” he said. “Because it can get overwhelming if you’re not versed in this space.”

Second, assess their customer maturity. Ask whether your organization would be a beta tester, and what their existing customer base looks like. “Even if they can’t share a logo list, get a sense of the industries they serve and the type of work they do,” Belliveau said. Being among the first customers has trade-offs — potentially lower costs and influence on the product road map — but also means bearing the risk of an immature solution.

Third, verify security and compliance credentials. Ask about ISO certifications and SOC 2 compliance. For organizations handling sensitive healthcare and claims data, these aren’t optional. “For a company like ours, owned by a large publicly held entity like Cigna, that’s a non-negotiable checkbox,” Belliveau said. “But you’d be surprised how few startups prioritize it during their MVP cycle.”

“Another point to assess with any vendor is the depth of their data. Where are they pulling it from? Do they own it? Do they audit it? How often are they planning to expand it?” added King as AI solutions depend on a robust data set for appropriate analysis and accurate output.

Generative vs. Non-Generative AI: A Critical Distinction Being Missed

One of the most dangerous misunderstandings in workers’ compensation involves conflating all AI technologies. Generative and non-generative AI serve fundamentally different purposes, and deployment decisions should reflect these differences — especially in safety-sensitive applications like pharmacy recommendations.

“The distinction between generative and non-generative AI is a big deal, and I’m not convinced it’s fully understood yet,” Belliveau said. He used a retail analogy to illustrate the point: “A Chief Marketing Officer and a cashier both work in retail, but their roles are vastly different. AI is exactly like that.”

Non-generative AI, like predictive analytics, classification systems, and measurement engines, has proven highly effective in workers’ compensation. Several companies successfully use extraction-based summarization to condense 50 pages of claim notes into single paragraphs, reducing reviewer burden while maintaining accuracy. This represents practical, proven AI deployment.

Generative AI, systems that create new content by predicting the next word in a sequence, presents fundamentally different risks and capabilities. When you ask ChatGPT, Copilot, Claude, or Gemini the same complex question five times, you receive five different answers. This variability might be acceptable in marketing copy. It’s unacceptable in pharmacy.

“You simply cannot apply that kind of variability to pharmacy data,” Belliveau said. “You can’t afford to be wrong or even inconsistent when someone’s life is on the line, or when you’re trying to get someone back to work sooner because they’re addicted to a particular medication or receiving the wrong mix of medications.”

At MyMatrixx by Evernorth, this distinction drives critical product decisions. The organization does not use generative AI in its pharmacy and therapeutic recommendation engines. “We don’t consider using GenAI for pharmacy therapeutic recommendations — at least not in the foreseeable future,” Belliveau said.

The reasoning is straightforward: hallucinations — plausible-sounding but false outputs — cannot be tolerated when clinical decisions affect patient safety and treatment outcomes. Even inconsistency erodes the trust of adjusters, nurses, and doctors managing complex pharmaceutical regimens.

The Real Cost of AI Adoption

Beyond technology selection, organizations must confront the economics of AI. A recent Fortune article documented CEOs frustrated after investing millions in AI tools only to see expected productivity gains and ROI fail to materialize. This pattern echoes a similar phenomenon from the 1980s and 1990s, when personal computers first arrived on office desks.

“Productivity actually decreased globally,” Belliveau said. “Email, word processors, connectivity, shortcut keys in WordPerfect, Lotus Notes — there was simply too much to learn, and people couldn’t focus on their actual jobs. It wasn’t until the World Wide Web came into broader use that productivity began to recover, and that lag lasted nearly a decade.”

Belliveau referenced research from a Nobel Prize-winning economist showing that AI follows a similar J-curve pattern: productivity declines initially before rising exponentially. “Right now, we’re somewhere in that downward portion of the curve, and productivity may be declining in some ways and not in others,” he said. “The core question is: where are you on that J-curve, and can you afford to be there?”

This timing consideration creates a critical test for any proposed AI investment: If AI had never existed — if your only frame of reference was a 1918 Fritz Lang film or 2001: A Space Odyssey, would you still pursue this business opportunity? If the answer is no, AI isn’t solving a real problem. If yes, you have a legitimate use case worth the investment.

A Path Forward for All Organization Sizes

The workers’ compensation industry runs largely on paper, faxes, and phone calls, a reality that reflects the complex, document-intensive nature of claims. Interoperability challenges, inconsistent electronic standards, and entrenched processes create friction that no technology vendor can eliminate alone.

Yet solutions exist for organizations of all sizes. “The capabilities are there, and the big vendors — Microsoft, Google, and Anthropic — know they are going to have to serve the small caps along with the big players,” Belliveau said. “Some of this technology is direct to consumer and commoditized in a way where even frontier technology can still be leveraged by smaller organizations.”

The key is disciplined evaluation and governance. Ask the right questions of vendors. Understand the technology you’re buying. Ensure valid use cases drive decisions, not FOMO. Establish governance frameworks before deploying solutions at scale.

“And most importantly,” added King, “remember that AI is a tool designed to solve specific problems, not a magic solution applicable to every challenge.”

As the industry continues navigating AI adoption, organizations that combine technical rigor with practical caution will emerge as leaders. Those that chase shiny objects without governance frameworks will face growing risk.

To learn more about responsible AI deployment in workers’ compensation, visit https://www.mymatrixx.com/.

This article was produced by the R&I Brand Studio, a unit of the advertising department of Risk & Insurance, in collaboration with MyMatrixx by Evernorth. The editorial staff of Risk & Insurance had no role in its preparation.