AI in Open Enrollment: 7 Proven Wins and 3 Costly Mistakes to Avoid
AI in open enrollment has gone from a pilot project to a line item, and most employers still cannot tell you whether it earned its keep. Your benefits administration vendor added a chatbot. Your enrollment platform added a recommendation engine. Somewhere in the demo there was a slide about reducing HR ticket volume by 40 percent. Then enrollment ended, and nobody measured anything.
That is a shame, because the underlying problem is real and expensive. Employees are making five-figure decisions in under an hour with almost no support. If you deploy AI in open enrollment carefully, it can close part of that gap. If you deploy it carelessly, you have handed plan interpretation to a system that no one at your company owns, no one audited, and no one can explain to an employee who chose wrong.
Key Takeaways
- Employees are not lazy about benefits. They are under-supported, and the data on plan selection errors proves it.
- AI in open enrollment delivers the most value on volume work: repetitive questions, cost modeling, and personalized reminders.
- Recommendation engines carry fiduciary and privacy exposure that most vendor contracts do not address.
- Governance is not optional. Nineteen of the most populous states now regulate employment-related AI.
- Every answer from AI in open enrollment should trace back to a plan document a human can point to.
Open Enrollment Was Already Broken Before AI Arrived
Start with what employees are actually deciding. According to KFF, annual premiums for family coverage reached $26,993 in 2025, a 6 percent increase, with workers contributing an average of $6,850. The average single-coverage deductible sat at $1,886. This is not a small purchase buried in a benefits portal. For many households it is the second-largest annual expense after housing.
Now look at how well people navigate it. In a study of nearly 24,000 employees published in the Quarterly Journal of Economics, researchers found that the majority chose financially dominated plans, producing excess spending equal to 24 percent of their chosen plan premiums. A dominated plan is one that costs more no matter what happens medically. There is no scenario where it wins. Lower-income employees picked them at higher rates, and most did not correct the mistake the following year.
The researchers tested whether menu size or a preference for low deductibles explained the pattern. Neither did. The driver was a straightforward lack of understanding of how health insurance works.
Recent survey data says the experience has not improved. The 2026 Selerix Employee Benefits Survey found that 35 percent of employees regret their last open enrollment choices, and only about a quarter say they understand their benefits well. Roughly seven in ten spend less than an hour on the whole process, and only 16 percent use decision-support tools at all.
Read those three numbers together and the case for AI in open enrollment becomes obvious. The money at stake is enormous, the error rate is high, and the support tools that already exist go mostly unused. Something has to change about how guidance reaches people.
What AI in Open Enrollment Actually Does Right Now
Strip away the marketing and there are four practical jobs that AI in open enrollment performs today. Each has a different risk profile, which matters more than the feature list.
1. Chatbots that answer questions at 11 p.m.
This is the most visible face of AI in open enrollment, and the easiest to justify. Benefits questions do not respect business hours. Employees compare plans after the kids are asleep, on a phone, in a browser tab they will close in four minutes. A well-built chatbot indexed to your actual plan documents answers those questions instantly and privately, which matters for the questions people are embarrassed to ask a human. Is my therapist in network. Does this cover fertility treatment. What happens to my HSA if I leave.
2. Decision-support tools that model total cost
The single most useful thing software can do during enrollment is stop the conversation about premium and start the conversation about total annual cost. A decision-support tool asks a handful of questions about expected utilization, prescriptions, and family situation, then projects out-of-pocket exposure across each plan option. Done well, this is precisely the intervention the dominated-plan research called for.
3. Personalized nudges instead of mass emails
Generic all-staff emails get ignored. AI in open enrollment can segment your population by enrollment status, life event, plan history, and language, then time reminders to the people who actually need them. A single employee who has not logged in gets a different message than a family of four re-electing the same plan for the fourth year.
4. Back-office work employees never see
This is the least glamorous and least risky use case. Reconciling eligibility files, flagging dependent data gaps, drafting first-pass communications, summarizing question volume by topic so HR knows what confused people. None of it touches an employee decision, and all of it saves your team hours.
The Adoption Numbers Behind AI in Open Enrollment
Before you assume your competitors are far ahead, look at the actual adoption data. SHRM’s State of AI in HR 2026 report found that fewer than half of organizations will use AI in HR in 2026, and adoption is concentrated in a handful of practice areas. Recruiting leads at 27 percent of organizations. Employee experience, the category that covers benefits communication, sits at 14 percent.
Meanwhile 92 percent of CHROs expect AI to be further integrated into the workforce this year. That gap between executive expectation and operational reality is where bad purchasing decisions live. Someone gets asked what the company is doing about AI, and a chatbot appears in the enrollment portal three weeks later with no owner, no testing protocol, and no plan for what happens when it is wrong.
Where AI in Open Enrollment Goes Wrong
Three failure modes come up repeatedly when AI in open enrollment is rolled out fast, and none of them show up in a product demo.
Confident answers that are simply wrong
A general-purpose model that has not been grounded in your specific plan documents will still answer the question. It will do so fluently. It will describe a typical PPO, or last year’s deductible, or a rule that applies to a different plan entirely. The employee has no way to detect the error and acts on it. Any AI in open enrollment deployment needs to be retrieval-grounded in your current summary plan descriptions, and it needs to say plainly when it does not know.
Protected health information and the vendor nobody vetted
The moment an employee tells a benefits chatbot about a diagnosis, a medication, or a pregnancy, you are handling protected health information. Under HIPAA, business associates are directly liable for compliance with certain requirements of the HIPAA Rules, which means the AI vendor and its subcontractors need a business associate agreement in place before launch, not after. Ask where prompts are stored, how long they are retained, and whether any of it is used to train a model. Get the answer in writing.
The fiduciary question nobody asks the vendor
Here is the one that keeps benefits attorneys up at night. If a recommendation engine steers employees toward a specific plan, what is it optimizing for? Employee out-of-pocket cost? Employer premium spend? Whatever the vendor’s carrier partner prefers? Plan sponsors carry fiduciary duties, and outsourcing a recommendation to a black box does not outsource the duty. You still own the outcome. If the vendor cannot explain the logic in plain language, that is your answer.
A Governance Checklist for AI in Open Enrollment
You do not need a forty-page policy. You need a short, documented set of decisions made before enrollment opens. The NIST AI Risk Management Framework organizes this work into four functions: govern, map, measure, and manage, and it maps cleanly onto benefits.
- Name an owner. One person accountable for what the tool says. Not a committee, and not the vendor.
- Define the scope in writing. What the tool may answer, what it must escalate to a human, and what it must never do.
- Ground it in current documents. This year’s summary plan descriptions, this year’s rates, this year’s networks.
- Test before launch. Run 50 real questions from last year’s ticket log and check every answer against the plan document.
- Disclose it. Employees should know they are talking to software and how to reach a person instead.
- Log and review. Every conversation is data about what your communications failed to explain.
- Check state law. SHRM’s research notes that as of February 2026, 19 of the most populous states had enacted AI laws or regulations touching employer AI use.
That checklist takes a few hours to complete. Skipping it is what turns AI in open enrollment from a support tool into a liability.
Questions to Ask Before You Buy
Bring these to the next demo before you sign anything. The quality of the answers will tell you more about a vendor’s readiness for AI in open enrollment than any product tour will.
- What documents does the model retrieve from, and who updates them when our plan changes mid-year?
- What does the tool do when it does not know? Show us that response.
- Will you sign a business associate agreement, and are your subprocessors covered by it?
- Is any employee input used to train or fine-tune your models? Can we opt out?
- If your tool recommends a plan, what variable is it optimizing? Show us the logic.
- What accuracy rate did you measure in production, and how did you measure it?
- What reporting do we get on question volume, topics, escalations, and unresolved conversations?
AI in Open Enrollment Works Best When a Human Owns the Outcome
Here is the honest version. AI in open enrollment is very good at scale and speed. It is not good at judgment, and it does not carry accountability. An employee whose spouse just got a cancer diagnosis does not need a faster answer about coinsurance tiers. They need someone who understands the plan, the provider network, and what happens next.
The right model is layered. Software handles the high-volume, low-stakes questions so your team stops answering the same deductible question 200 times. Humans handle the moments that matter, backed by a plan that was designed to be explainable in the first place. Complexity is the root problem here, and no chatbot fixes a plan design that requires a chatbot to understand.
That is where the work starts. Our health benefits services cover benefits administration technology, integrations, eligibility, auditing, compliance, and employee communications, so the tools your employees touch are actually connected to the plan behind them. And because plan sponsor obligations are only getting sharper, our fiduciary framework exists to make sure decisions like these are documented and defensible.
The Bottom Line on AI in Open Enrollment
Employees are overspending by measurable amounts because nobody helped them decide. That is the problem worth solving, and AI in open enrollment is a legitimate part of the solution when it is grounded, governed, and owned by a named human being.
What it is not is a substitute for strategy. If your enrollment experience only works when a chatbot translates it, the plan itself is the thing to fix. Get the design right, then let the technology do what it is good at. That order matters, and it is the difference between a tool that empowers your people and one that quietly makes your exposure worse.
Frequently Asked Questions About AI in Open Enrollment
What is AI in open enrollment?
AI in open enrollment refers to software that uses artificial intelligence to support benefits selection. In practice that means three things: chatbots that answer plan questions in plain language, decision-support tools that project total annual cost across plan options, and systems that personalize reminders and communications by employee situation. It also covers back-office work like eligibility reconciliation and question-volume reporting that employees never see.
Are benefits chatbots HIPAA compliant?
Not automatically. Compliance depends on the vendor and your contract, not on the technology. The moment an employee mentions a diagnosis, medication, or pregnancy, protected health information is in play, and HIPAA holds business associates directly liable for certain requirements. Before launch you need a signed business associate agreement that covers the vendor and its subprocessors, plus written answers on data retention, storage location, and whether employee inputs are used to train models.
Do decision-support tools actually help employees choose better plans?
They help when employees use them, and most do not. Recent survey data puts usage at roughly 16 percent. The tools that move the needle are the ones that shift the conversation from monthly premium to projected annual out-of-pocket cost, because premium-only comparison is exactly what drives people into plans that cost them more no matter what happens medically. Adoption is a communication problem, not a software problem.
Can AI replace a benefits advisor during open enrollment?
No. AI in open enrollment handles volume and speed well, and it carries no accountability. An employee facing a serious diagnosis, a complicated provider network question, or a mid-year life event needs judgment, plan-specific context, and a person who owns the answer. The workable model is layered: software absorbs the repetitive questions, and humans handle the decisions that carry real consequences.
What are the biggest risks of using AI in open enrollment?
Three stand out. First, fluent answers that are factually wrong because the tool was never grounded in your current plan documents. Second, privacy exposure from an unvetted vendor handling protected health information. Third, fiduciary risk when a recommendation engine steers employees toward a plan without disclosing what it is optimizing for. All three are manageable, and all three are invisible in a product demo.
How do we measure whether AI in open enrollment worked?
Set the baseline before enrollment opens. Track question volume and topic mix, escalation rate to a human, unresolved conversation rate, answer accuracy against plan documents on a sampled basis, and plan-switching behavior compared with prior years. If your vendor cannot report on those, you have no way to know whether the tool helped or quietly misinformed people.
How much does AI in open enrollment cost?
Pricing usually arrives bundled into a benefits administration or enrollment platform fee rather than as a standalone line item, which makes the true cost hard to isolate. Ask for the AI component to be priced separately so you can evaluate it on its own merits, and factor in the internal time required for document grounding, pre-launch testing, and ongoing review. A tool that nobody maintains is more expensive than no tool at all.
Ready to Make Enrollment Something Employees Understand?
Let’s talk about the plan design, technology, and communication strategy behind a benefits program your people can actually use.
Schedule a CallDisclaimer: This article is provided by Ethos Benefits for general informational and educational purposes only. It does not constitute legal, tax, compliance, medical, or investment advice, and it does not create an advisor-client relationship. Regulations governing employee benefit plans, artificial intelligence, and health information privacy vary by jurisdiction and change frequently. Employers should consult qualified legal counsel and their own benefits advisors before making decisions about plan design, vendor selection, or the deployment of artificial intelligence tools. Ethos Benefits makes no representation or warranty regarding the accuracy or completeness of third-party data referenced here.
