You Cannot Delegate Prudence to a Model, and ERISA Makes That Personal
Benefits is where AI adoption is quietly fastest — enrollment assistants, claims triage, retirement guidance chat — and it is also the one corner of the company where the people approving the tool can be personally liable for the outcome. The standard is not whether the tool works. It is whether the decision to use it was made properly and is still being watched.
Why Benefits Is Different From Every Other Department
When a marketing team adopts an AI tool badly, the company has a problem. When a benefits committee adopts one badly, the individuals on that committee may have a problem. ERISA imposes duties of loyalty and prudence on plan fiduciaries, requires them to act solely in the interest of participants and beneficiaries, and provides for personal liability for losses to the plan caused by a breach.
That structure produces a distinctive culture. Benefits committees keep minutes, run formal vendor searches, benchmark fees and revisit decisions on a calendar — not because anyone enjoys it, but because the defense to a prudence claim is the record of the process. AI adoption tends to arrive through a different door entirely: an HR technology vendor ships a new assistant feature, it gets switched on, and nobody treats it as a fiduciary decision because it did not come through a fiduciary process.
That mismatch is the actual risk. Not the model. The absence of a file.
Where the Duty Attaches
ERISA determines fiduciary status functionally. A person is a fiduciary to the extent they exercise discretionary authority or control over plan management, exercise authority or control over plan assets, render investment advice for a fee with respect to plan assets, or have discretionary authority in plan administration. None of those turn on what a contract calls someone.
Two consequences follow for AI tooling. First, selecting and monitoring a service provider is itself a fiduciary function, so bringing in a vendor does not exit the regime — it changes what the duty is about. Second, if the tool makes discretionary decisions, the question of who is exercising that discretion becomes live, and the answer is not automatically the vendor simply because their software produced the output.
- Retirement guidance chat. The education versus advice line is old, heavily regulated and genuinely contested. A tool that answers "what should I put my contribution in" for a specific participant is operating near it.
- Health claims triage. Adjudication touches the claims procedure rules, which require a full and fair review and an explanation the claimant can actually use to appeal.
- Eligibility and enrollment automation. Errors here are quiet and compounding — a participant wrongly excluded may not surface for years, and correction is expensive and retroactive.
- Fee and fund analytics. Using a model to support an investment lineup review is fine; treating its output as the review is the failure mode.
Participant Data Is Not Ordinary Vendor Data
Whether participant data constitutes a plan asset in the technical ERISA sense is a real open question that has been argued in fee litigation, particularly where a recordkeeper wanted to use participant information to cross-sell unrelated products. You do not need that question resolved to see the practical problem.
The Department of Labor has published cybersecurity guidance directed at plan sponsors, fiduciaries, recordkeepers and participants, covering vendor selection practices and program expectations. Whatever the plan-asset characterization, the protection of participant information sits within the fiduciary frame. Feeding that data into a model raises specific and answerable questions:
- Is participant data used to train or improve the vendor's models? Get this in the contract, not in a sales answer, and confirm it covers subprocessors.
- Where does inference data go? A benefits assistant built on a third-party model API means a second vendor exists that may not be in your file at all.
- What is retained, and can it be deleted? Conversation logs from a benefits chat assistant contain health and financial detail that the employer often should not hold.
- Who sees it internally? An HR-side assistant that surfaces participant queries to managers creates an employment law problem alongside the ERISA one.
- Is any of it protected health information? Group health plan data brings HIPAA obligations that run parallel to ERISA and are not satisfied by the same paperwork.
What a Defensible Adoption File Looks Like
Because prudence is assessed on process, the deliverable is a file. It does not need to be long. It needs to exist, be contemporaneous, and show that the right questions were asked by the right people.
- Minutes recording the decision. Who proposed it, what it does, what alternatives were considered, what concerns were raised and how they were resolved.
- A written scope of what the tool decides versus suggests. This single distinction drives most of the analysis, and it is remarkable how often nobody has written it down.
- Vendor diligence on the AI specifically. Not the vendor's SOC 2 alone. Model provenance, subprocessors, training use, human override, error rates and how errors are surfaced to you.
- Fee analysis including the AI component. If the feature carries a cost to the plan or the participants, it belongs in the reasonableness assessment.
- A monitoring cadence with content. Monitoring means reviewing something specific — error and override rates, complaint volume, sampled outputs — not confirming the vendor is still in business.
- A documented escalation and correction path. Benefits errors have established correction programs, and the time to understand yours is before an automated process replicates a mistake across a population.
Frequently Asked Questions
Our recordkeeper turned on an AI assistant. Do we have to do anything?
Yes — at minimum, look at it and record that you did. A feature enabled by default is still a service being provided to your plan, and monitoring service providers is a fiduciary function. The practical step is to put it on the next committee agenda, ask what it does, ask what data it uses, and minute the answers. That converts a passive default into a documented decision.
Is an AI benefits chatbot giving investment advice?
It depends on what it says and to whom. General education about plan features and investment concepts has long been treated differently from individualized recommendations to a specific participant about their own account. The boundary is fact-specific and has been the subject of extensive regulatory activity. If the tool can produce anything a participant would reasonably read as a recommendation, that characterization needs a considered answer from counsel rather than an assumption.
Can we contractually shift fiduciary liability to the AI vendor?
Partially and imperfectly. Some providers will accept a defined fiduciary role for specified functions, which is meaningful and worth pursuing where available. What cannot be shifted is the duty to prudently select and monitor that provider, and ERISA restricts provisions purporting to relieve a fiduciary of responsibility. Indemnification affects who ultimately pays; it does not remove you from the claim.
How does this interact with state AI and health insurance laws?
ERISA preemption is a genuinely complicated area and self-funded plans are positioned differently from insured products. Several states have enacted requirements around AI use in utilization review and coverage denials, and whether a given requirement reaches a particular plan arrangement is a legal question with a non-obvious answer. Treat overlapping obligations as the default assumption and get the analysis done for your structure.
What is the single highest-value thing to do this quarter?
Inventory it. Ask every benefits vendor, in writing, where AI is used in the services provided to your plan, what data it processes, and whether plan data is used for training. Most sponsors cannot currently answer that question, and the inventory is both the prerequisite for every other control and, by itself, evidence of a prudent process.
Does any of this apply to a small employer plan?
ERISA's fiduciary duties do not scale down with headcount. What scales is proportionality: a small plan is not expected to run a large plan's process, but it is expected to have one. For a small sponsor the realistic version is an annual documented review of vendors and the tools they have deployed, plus a decision file for anything new.
Selling Software Into HR and Benefits?
Buyers in this space are personally on the hook for the tools they approve, and they read vendor sites as diligence material. Claims about what your AI decides, what data it retains and who reviews its output will be quoted back to you in a committee meeting.
See every capability and data-handling claim on your site in one pass. Run a free scan before your next enterprise review.
This article is general information and not legal advice. ERISA fiduciary questions are fact-specific, several issues discussed here are actively contested or evolving, and plan structure materially changes the analysis. Consult qualified ERISA counsel about your plan.