A framework for AI in benefits administration – getting it right and the competitive opening for organizations that do.
The pace of AI adoption in our industry is unlike anything we’ve seen before. Workflows that used to take days now take minutes. Patterns that might have been missed are identified in moments. Most importantly, teams are reclaiming countless hours to focus on what truly drives the business: clients, strategy, and growth.
“Intelligent automation” is what happens when process automation, machine learning, and language models work together to handle cognitive work (like reading a document, classifying a claim, spotting an anomaly, or drafting a response) that once required a human at every step. And on that front, AI is delivering.
But there is a meaningful difference between “AI can assist us” and “AI can decide for us,” and crossing that line carries serious consequences.
The Cases Worth Watching
If you’ve followed the headlines over the last couple of years, you know the stories. It’s worth pulling a few together here because the pattern matters more than any single case.
A major health insurer’s automated claims-review system. The company used an algorithm to flag and deny medical claims, and reports surfaced that during batch reviews, physicians spent only one to two seconds per claim. Plaintiffs argued that a signature attached to that kind of “review” doesn’t mean much, and the court agreed, letting a class-action case proceed on claims of ERISA fiduciary breach and violations of state laws requiring licensed clinician review. The litigation continues, and every time the story resurfaces, the reputational damage deepens.
A leading Medicare Advantage carrier and its post-acute care subsidiary. AI was used to predict how long patients would need post-hospital care, driving decisions about when patients were discharged and how much was reimbursed. The trouble started when the model began overruling treating physicians. Most of the denied claims were later reversed on appeal, which tells you everything you need to know. Courts allowed breach-of-contract and good-faith claims to proceed. CMS started paying attention, and so did Congress.
The industry-wide prior authorization wave. Payer after payer began using AI not just to sort requests but to make coverage calls. Members couldn’t see how decisions were made or easily figure out how to appeal. The fallout? Lawsuits against multiple insurers, new state and federal laws mandating human review of denials, and a number of companies forced to tear out and rebuild systems they’d just finished installing.
The Real Issue Isn’t AI – It’s Accountability
The common thread in these cases has little to do with AI itself. It’s accountability. In each one, AI was placed in a seat the law, a contract, or a fiduciary duty had reserved for a human. Decisions were made at massive scale with little meaningful review. And when someone asked how a particular outcome was reached, no one could give a clear answer.
Companies don’t get sued for using AI. They get sued for yielding accountability to it.
That distinction changes everything.
Three Non-Negotiables
Three principles separate the companies that get this right from those that put themselves at serious risk.
First, AI does not get to make the final call.
When the decision carries legal, contractual, or fiduciary weight, you cannot rely on AI to make the decision. No matter how good the model is, decisions such as denying a complex claim or making a medical-necessity determination must be made by a human.
Second, the human review has to be real.
A clinician spending two seconds per claim isn’t reviewing anything. They’re being a rubber stamp, and the courts have started treating them that way. If your workflow doesn’t give qualified people the time and information to make genuine judgments, you’re asking for real trouble.
Third, every AI-assisted outcome has to be explainable and auditable.
When you are asked by a member, regulator, or attorney how a decision was reached, “the system said so” is not an answer. You need to be able to walk them through the decision logic, inputs, and human review trail, possibly months or years after the fact.
That’s how DataPath builds it. Every AI-assisted decision in our systems is traceable from input to outcome and ready to withstand scrutiny, whether from a member, a regulator, or a courtroom.
Where AI Belongs and Where It Doesn’t
Done right, intelligent automation doesn’t put the brake on innovation; it makes innovation defensible. The guiding principle is simple: AI can recommend, surface, and prioritize, but humans decide.
The table below illustrates this principle. None of the functions on the left are decisions; they clear the path so humans can make the decisions on the right more quickly and effectively.
| Use AI for … | Don’t use AI for … |
| Document intake and classification | Final claims denial |
| Eligibility validation checks | Eligibility determinations |
| Claims routing and prioritization | Medical necessity decisions |
| Fraud and anomaly detection | Automatic fraud denials |
| SLA monitoring and forecasting | Auto-closing cases to hit SLAs |
| Drafting communications for human review | Sending binding decisions |
| Quoting plan language as written | Interpreting ambiguous policy language |
The biggest productivity wins in our industry don’t come from AI replacing adjudicators, examiners, or service reps. They come from AI clearing the busy work so those people can focus on the parts of the job only humans can do well.
What Responsible Deployment of AI in Benefits Administration Looks Like
Building intelligent automation is a discipline, not a sprint. It starts with not deploying AI just because everyone else is. At DataPath, the rule of thumb is that every feature must be directly linked to a specific operational problem worth solving.
From there, it requires real domain expertise. The regulatory and operational nuances of benefits administration can’t be reverse-engineered from a generic large language model. It demands rigorous testing in controlled environments before anything touches live data or live participants. And it depends on workflows designed with humans actually in the loop, supported by AI that gives them what they need to make sound decisions.
The market is splitting fast. On one side are platforms trying to use AI to remove humans from the equation to reduce costs. On the other are platforms – DataPath among them – using AI to make humans dramatically more efficient and effective. The first group is carrying the legal exposure you’ve been reading about. The second is quietly building a durable advantage.
The Competitive Opening
For TPAs, brokers, banks, and any organization in the business of fiduciary money or regulated decisions, the next few years will be defining. The technology is moving fast, regulations can’t keep up, and members and clients are paying closer attention than ever to how their data is used and how decisions are made.
Organizations that race to adopt AI irresponsibly are at risk of becoming the next case study. The winners will be those who understand that AI’s value lies not in replacing human judgment but in maximizing the time, capacity, and clarity that good judgment requires.
There’s also a real competitive opening here. In a market increasingly skeptical of opaque automation, “explainable” and “human-led” have shifted from minimum requirements to active differentiators – and the business case follows. Organizations that govern AI well process claims and transactions faster because they aren’t constantly having to stop to defend their decisions. They carry lower compliance and litigation risks because their outputs hold up under regulatory and audit scrutiny. They win more RFPs because procurement teams are now asking pointed questions about how AI is used and overseen. And they earn the kind of durable client trust that drives renewals, referrals, and growth.
AI expands what’s possible. Humans protect what’s at stake. Build for both, and the rest takes care of itself.
Bo Armstrong is the Chief Product and Marketing Officer for DataPath. He joined the company in 2015 and has held multiple leadership positions at companies ranging in size from Fortune 500 to start-ups.
