AI can't be a fiduciary, but it can leave a paper trail

ERISA's mandate to act solely in the interest of plan participants and beneficiaries is deceptively simple, a column in 401(k) Specialist argued on Sept. 21, because the fiduciaries who carry it out bring cognitive bias, relational pressure, self-interest, and the comfort of calling a convenient decision reasonable. The industry has spent two decades closing an information gap that, the column contends, was never about information. The missing ingredient, in its framing, is moral and ethical efficacy — a fiduciary's operating belief in his capacity to identify the right course of action, take it, and defend it before a regulator, a plaintiff's attorney, or a room of participants whose retirement depends on the judgment.

Two decades of fiduciary education have produced sponsors who understand prudent process, fee benchmarking, and investment policy compliance better than at any point in ERISA's history, yet the breach docket stays busy with conflicts of interest, undocumented decision-making, and provider selections shaped by familiarity rather than evidence. The breach claim arrives without enforcement data attached. The column leans on research from Dr. Sean Hannah, a co-founder of the Behavioral Governance Institute, describing ethical leadership as among the strongest predictors of ethical behavior under pressure — stronger than stated values, training certifications, or codes of conduct standing alone. If that holds, the fiduciary who knows what ERISA requires and the fiduciary who acts on it when a long-standing service-provider relationship or a board's budget pressure points the other way are frequently different people.

AI enters the argument as an accountability instrument rather than a decision-maker; the column concedes that a model holds no moral agency, cannot feel the weight of a conflict, and cannot exercise conscience in a committee meeting, then claims the technology can surface the moral architecture of a fiduciary decision in real time, before consequences harden. It closes on a branded construct, Special Purpose Avatars, whose mechanism the published excerpt does not explain.

What a sponsor can test is narrower than the thesis. Undocumented decision-making is one of the column's own named drivers of breach, and a contemporaneous record is the one thing software genuinely changes; whether the committee reading that record grows braver is not a claim anyone can verify, and the piece offers a research construct and a product name rather than a standard to audit against. This publication has argued that AI has saturated the paperwork layer and that the next moat is governance and certification a sponsor can hold up when it is examined; the column points the same direction and stops short of the certification question, where the liability sits. The first real test will be a deposition in which a sponsor is asked to produce the trail.