%20(1).jpg)
You likely don't need another person alerting you to the reality that companies are quickly and eagerly bringing AI to the workplace. An area often cited for exploration is Human Resources (HR), where internal chatbots can respond to employee benefits inquiries. In my decades of working in employee benefits, never has the pace of change – and need – been greater.
Pulling together an AI chatbot powered by Large Language Models (LLMs) and loaded with your benefits documentation such as Summary Plan Descriptions (SPDs) doesn’t seem like that much of a leap, and it’s not; that capability is available today.
But for HR and Benefits leaders, the worry should not merely be about technology concerns such as buggy user experiences. Using AI to answer employee benefits questions can be like stepping into a minefield governed by the Employee Retirement Income Security Act (ERISA). The stakes may involve direct financial liability for you as a plan sponsor.
Under ERISA, fiduciary status isn’t defined by your job title; it’s about the functions you actually perform. While there is currently no confirmed legal position that a chatbot can reach the point of providing discretion, there is a risk that if a system takes ambiguous plan language, interprets it, and tells an employee what is covered, how much it will cost, or which plan to choose, it may be seen as exercising discretionary authority. When you then deploy a chatbot that gives this kind of guidance, it is possible that everyone who authorized it — the HR team, the benefits committee, the HR leaders, you — could potentially become "functional fiduciaries" over that tool’s outputs.
Many internal teams believe they can bypass this potential liability by simply slapping a disclaimer on the chatbot that says, "This is for informational purposes only. Consult your official plan documents." While disclaimers and appropriate warning are important, and tools should always redirect participants back to a human, they are likely not sufficient on their own. Under ERISA, courts have been skeptical that you can waive your fiduciary duties to plan participants with a footnote.
The key risk is not simply whether your tool is labeled as education, but whether your tool is designed and operating in a way that avoids exceeding its intended authority. If your AI tool begins to function as a guidance mechanism, you may become legally accountable for the guidance it provides.
To understand the risk of AI producing false, illogical or unsupported responses (aka "AI hallucinations") specifically in benefits, we don't need to look at tech law; we can already look at existing ERISA case law regarding misrepresentation.
Courts have consistently held employers and plan administrators liable when they provide incorrect benefits information to employees. Take the precedent set in cases like Winkelspecht and similar ERISA misrepresentation suits: when a company representative (even a temporary HR worker) erroneously assures an employee about their coverage, the court can potentially find the employer liable for the misrepresentation. In such cases, because the employer distributed an SPD that didn't perfectly clarify the issue, and the employee reasonably relied on the representative's bad information, the employer was on the hook for the financial fallout.
Now, replace that HR professional with a generic LLM. Imagine your internal AI tool misreads a coordination of benefits clause, or “hallucinates” that a specific out-of-network surgery is covered. If the employee relies on that answer and incurs a substantial medical bill, fiduciaries who breach their duties may be held liable to make the plan whole under ERISA Section 409.
If a claim were to arise and the Department of Labor (or a plaintiff's attorney) were to investigate, they are likely going to ask the Plan Sponsor: What benefits expertise did the developers possess? How did you test and validate the accuracy of the AI? Where was your operational oversight? An internal tech project built on a generic LLM may not provide the strongest defense.
Right now, most internal AI tools operate in a low-stakes environment. If a chatbot gives a slightly wrong answer about your company travel policy, someone can simply submit an expense exception. Benefits are fundamentally different. They dictate family budgets, healthcare access, and financial wellness, and carry potentially enormous financial consequences. Furthermore, the data environment is notoriously hostile to generic LLMs:
ERISA’s prudence standard requires fiduciaries to act with the care, skill, and diligence of a knowledgeable expert. When you manage your company’s 401(k), you don’t let your internal engineering team build a custom trading platform; you hire a vetted financial institution. The same standard should apply to benefits AI.
When evaluating these technologies, the goal should be to have the safest possible tool that avoids inadvertently assuming fiduciary status by staying clearly within its intended role. As fiduciaries, the questions you should be asking yourselves and your AI partners are:
As a plan sponsor, being able to address these areas confidently allows you to regain control in this complex space and build reliable solutions to help your people thrive.
