AI Enters ERISA’s Document-Heavy World, Promising Benefits and Revealing Risks

Few can dispute the omnipresence of artificial intelligence.

In a podcast interview several months ago, highly respected finance journalist Andrew Ross Sorkin said he completed a simple legal task using AI that would have cost him a few hundred dollars in fees if he had sent it to his lawyer for review.

Sometimes, however, the use of AI does not go as smoothly.

Sullivan & Cromwell LLP recently admitted to and apologized for AI hallucinations included in a motion the firm filed in an April bankruptcy filing. Hallucinations are common errors or inventions of information AI tools may make when responding to a prompt. The mistakes can include generating false or fictional information.

For Sam Widdoes, an entertainment attorney based in Los Angeles, Sorkin’s comments struck a nerve. Widdoes went on to write a Substack post laying out how the technology may shape the legal profession.

“My industry is changing as every industry is going to be changing,” he says.

Impact on ERISA Space

In the document-heavy world of retirement and benefits plans, attorneys tell of how AI has changed—some for the better, some for the worse, some to be determined—the way some legal work is done. Legal experts say plan sponsors can best use the technology through AI legal review, such as reviewing documents, which allows plan sponsors to be more informed when they meet with counsel. They caution, however, that they could violate their fiduciary duties by over-relying on the technology.

Legal work done to comply with the Employee Retirement Income Security Act is filled with paper—plan documents, investment committee hearing minutes, claim files, discovery in litigation. But as ERISA attorneys, fiduciaries and others use AI to draft plan documents, review claims, analyze investment data and sift through sprawling litigation records more efficiently, the tools also raise profound questions about prudence, privacy and the limits of automation in a field built on fiduciary judgment.

“There is no blanket prohibition on using AI,” says David Tracey, a partner in Sanford Heisler Sharp McKnight LLP who represents plaintiffs in fiduciary breach litigation. “But there are imprudent uses of AI.”

The American Bar Association reported that 31% of legal professionals said they used generative AI for work purposes in 2024, up from 27% in 2023, and 21% of law firms reported that they used the technology—down from 24% a year prior. However, according to the ABA, larger firms were more likely to use the technology, and approximately 15% of respondents said they were unsure if they would use the technology for work purposes.

Amanda Amert, chair of the ERISA litigation group at Willkie Farr & Gallagher LLP, says plaintiffs’ firms, for example, can use AI to more effectively review plan documents such as Form 5500s to compare investment performance or fee structures that could become the basis of a legal complaint.

Last year, 94 excessive fee and imprudent investment cases were filed in the U.S., according to Encore Fiduciary, and while plaintiffs’ firms have always had to manually draw the comparisons, AI can now assist with the process.

Meanwhile, Amert says she believes that plan participants also can use the technology to scour plan documents to learn about benefits to which they are entitled.

In both instances, Amert says plan sponsors must do what they should already be doing: follow a prudent process and document their work when selecting investments or negotiating fees.

Prudence, Loyalty

Under ERISA, many lawyers say the process is the important thing to document, because the law itself is rooted in fiduciary duties of prudence and loyalty. Courts routinely examine not only what retirement plan fiduciaries decided, but how they reached those decisions.

That standard, practitioners say, may become the defining legal test for AI adoption in the benefits industry.

Marcia Wagner, a longtime employee benefits attorney who founded the Wagner Law Group, says AI could substantially improve plan administration by helping generate legally compliant documents and making technical disclosures easier for plan participants to understand. The technology could also refine investment recommendations for retirement savers, she says, though it remains vulnerable to hallucinations.

Wagner warns that fiduciaries using AI systems would still be required to understand and monitor the technology, rather than blindly relying on it.

“Reliance upon a particular tool to assist in satisfying the duty of prudence cannot be an assurance that the duty of prudence is satisfied,” she says.

That concern has become especially acute as large language models gain traction inside law firms and consulting practices that manage retirement plans worth billions of dollars.

Sanford’s Tracey compares AI systems to outside investment advisers—useful, but incapable of relieving fiduciaries of responsibility.

“A fiduciary can’t just delegate their functions to AI and not question and evaluate the recommendation they are getting,” he says. “There is a real temptation to just let AI do it and not pay attention.”

When a plan sponsor hires a 3(38) fiduciary, that fiduciary assumes discretion over investment selection and monitoring. While the plan sponsor then, in theory, has fewer fiduciary obligations, the plan sponsor must still thoroughly monitor that fiduciary and evaluate their fulfillment of their fiduciary obligations.

A Tighter Embrace

The legal profession’s embrace of AI has accelerated rapidly over the past two years, particularly in litigation discovery, a stage at which lawyers must review enormous collections of emails, committee materials and financial records, experts say.

ERISA litigation is particularly suited to using AI tools because the cases often involve years of investment data and extensive administrative records.

Tracey says litigators increasingly use AI to summarize documents, compare records and identify patterns across massive productions—work once handled primarily through keyword searches and by armies of junior associates.

The upcoming Supreme Court review in Winston R. Anderson et al. v. Intel Corp. is expected to include a decision on whether plaintiffs must identify a “meaningful benchmark” when claiming that retirement funds underperformed. For some observers, the ruling could increase the importance of using AI for benchmarking.

“If the Supreme Court determines next term that meaningful benchmarks are a necessary pleading requirement,” Wagner says, “AI could [identify] those investments that best satisfy that standard.”

Cybersecurity Risks

While the benefits and risks of AI tools become clearer, and even as the technology advances, lawyers say some of the most immediate concerns are cybersecurity, confidentiality and bias.

ERISA plans contain highly sensitive financial, medical and personally identifiable information. If AI systems are used in claims administration—particularly in health plan reviews—experts say companies will need safeguards to ensure that algorithms do not improperly discriminate against participants or expose confidential data.

“Contracts with service providers will need to be reviewed to the extent that they are using AI,” Wagner says.

But as the technology progresses— Anthropic’s CEO Dario Amodei warns 50% of entry level white collar jobs could be disrupted in the next five years—how AI is used in the ERISA benefits world could quickly change.

Willkie’s Amert remembers that when she began practicing, the topic that worried colleagues as the movement to daily pricing of plan investments from monthly. By now, that is simply the norm.

“It did not end retirement plans, but this is another wave of technological change, and we’ve weathered a lot of them,” she says. “We will find the right way to use AI and incorporate