Industries / Finance and business services / Payroll / Benefits administration
Enrollment
Enrollment is the work of offering employees their benefit choices and signing them and their families up: when they are hired, once a year, and when their lives change.
4
Milestones achieved, from 1978 to 2017.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1978
Benefit menus enter law
A US law of 6 November 1978, Public Law 95–600, added section 125 to the tax code. It defines a "cafeteria plan" as a written plan in which all participants are employees who "may choose among 2 or more benefits consisting of cash and qualified benefits", which gave employers a legal form for letting each worker choose between pay and a set of benefits.
Credit
US Congress
2000
The default decides
In a paper issued in May 2000, Brigitte Madrian and Dennis Shea studied a large US company that began enrolling new hires in its 401(k) retirement plan automatically unless they declined. They found participation "significantly higher under automatic enrollment", and that the contribution rate and investment fund the company chose as defaults had "a strong influence" on how much people saved, because a substantial share of those enrolled stayed with them.
LessonThe default an administrator sets at enrollment becomes the choice of many of the people enrolled.
Credit
Brigitte C. Madrian and Dennis F. Shea (National Bureau of Economic Research)
2014
Small-business exchange stalls
The Affordable Care Act set up exchanges, called SHOPs, where small employers could buy group health cover. The US Government Accountability Office found on 13 November 2014 that the 18 state-based SHOPs had enrolled about 76,000 people through nearly 12,000 small employers as of 1 June 2014, "significantly lower than expected". Stakeholders blamed a small-business tax credit that "may be too small and administratively complex", employers' freedom to renew their existing plans (allowed until 1 October 2016) and employer misconceptions about whether SHOPs were available.
LessonA new enrollment channel draws few employers unless it beats the arrangement they already have.
Credit
US Government Accountability Office (reporting)
2017
Most picked a losing plan
Saurabh Bhargava, George Loewenstein and Justin Sydnor studied 23,894 employees at a US firm who built their own health plan from 48 options that differed only in deductible, copayment, coinsurance, out-of-pocket limit and premium. Nearly every low-deductible plan was "financially dominated": an otherwise identical high-deductible plan cost less whatever medical bills arrived. The majority chose dominated plans, with excess spending equal to 24% of the premiums they chose; low-income employees were more likely to, and most did not switch the next year. The authors trace this to "employees' lack of understanding of health insurance", not to the size of the menu.
LessonMore plan choices help employees only if they can compare what each one would cost them.
Credit
Saurabh Bhargava, George Loewenstein and Justin Sydnor (Quarterly Journal of Economics, 2017)
Watch for
Benefits answers from an agent
Workday said on 24 September 2026 that Benefits Guidance, part of its new Total Benefits product, "answers employees' benefits questions in plain language, through open enrollment and major life moments" through its Self-Service Agent, and expected it to be generally available in October 2026. The release gives no figures for the accuracy or use of the guidance.
Achieved whenA customer or outside reviewer publishes how often the agent's answers matched the plan documents over a full open enrollment period.
Claimed by
Workday, September 2026
Open
01
A plan-choice test for AI
A public test of AI helpers that recommend health plans at enrollment. It holds at least 200 invented households (ages, expected doctor visits, prescriptions, income) and plan menus built the way the firm in the 2017 study built its 48 plans, so that some plans are dominated. A script works out each household's expected cost under every plan and which plans are dominated; each AI helper is asked to recommend a plan, and the test reports how often it picks a dominated one and what that costs the household.
The briefCredit · unclaimed02
Mid-year changes, answered
A public answer key for mid-year changes to benefit elections in US cafeteria plans: at least 300 cases, each an employee, a life event and a requested change, answered "allowed" or "not allowed" with the paragraph of Treasury regulation 1.125-4 it rests on. Then a rules engine that passes the key, and published scores for benefits software and AI models.
The briefCredit · unclaimed
Open 01 · the brief
A plan-choice test for AI
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of AI helpers that recommend health plans at enrollment. It holds at least 200 invented households (ages, expected doctor visits, prescriptions, income) and plan menus built the way the firm in the 2017 study built its 48 plans, so that some plans are dominated. A script works out each household's expected cost under every plan and which plans are dominated; each AI helper is asked to recommend a plan, and the test reports how often it picks a dominated one and what that costs the household.
Why it is still open
In the 2017 study most employees chose dominated plans. The closest AI test found, by Mallika Mainali and colleagues (arXiv, October 2025), measures whether GPT-4 and GPT-5 match decision-makers with set levels of risk tolerance on health insurance choices; it does not test whether they avoid dominated plans. Workday's benefits agent of September 2026 publishes no accuracy figures. This research found no public test of AI helpers against dominated plans.
Built with what exists
Python for the cost and dominance calculations; AI models such as Claude Opus 5.5 and Mistral Large 4, whose weights an employer can run on its own machines (see the Note "A trillion open parameters"), as systems under test; prompts scored on whether the helper first asks about the household's needs, the habit Ai2's XYBench measures (see the Note "Answering the question behind the question").
First step
Rebuild a 48-plan menu from the four cost-sharing parts the paper describes, with invented prices that make the low-deductible plans dominated, and ask one model to advise 20 households.
Done looks like
At least 200 households across five menus, each with the dominated plans worked out by script and checked by two people; a published rate of dominated picks and the average excess cost in dollars for at least three AI models, each run three times.
Open 02 · the brief
Mid-year changes, answered
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public answer key for mid-year changes to benefit elections in US cafeteria plans: at least 300 cases, each an employee, a life event and a requested change, answered "allowed" or "not allowed" with the paragraph of Treasury regulation 1.125-4 it rests on. Then a rules engine that passes the key, and published scores for benefits software and AI models.
Why it is still open
An employee may change elections during the year only after the events the regulation lists, and its paragraphs (b) to (d) work through them in at least 13 numbered examples, covering marriage, adoption, divorce, a child aging out, starting or ending a job, a move to a new worksite and special enrollment rights; one begins, "Employee A's marriage to B is a change in status". A search for this research found vendor guides but no public answer key built from the regulation.
Built with what exists
The regulation's text from Cornell's Legal Information Institute; Python; an AI coding agent that writes the rules engine and is checked only against the answer key, as AWS checked an emulator its agents wrote (see the Note "Code no person read").
First step
Turn the 13 numbered examples into 13 cases, then write three variations of each that change one fact.
Done looks like
At least 300 cases, each citing its paragraph and signed off by two reviewers, released under an open license; a rules engine that passes all of them; scores for at least three AI models, with every disagreement listed.