BRENNEN

Industries / Finance and business services / Payroll / Benefits administration

Claims and questions

This is the work of deciding and paying what members claim, from a hospital bill to an advance from a provident fund, handling their appeals, and answering their questions about what they are owed.

6

Milestones achieved, from 2000 to 2025.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 2000

    Rules for claim denials

    The US Labor Department's claims procedure rule, published on 21 November 2000 and applying to claims filed on or after 1 January 2002, sets time limits for deciding claims and appeals in ERISA plans and lists what a denial must say, "in a manner calculated to be understood by the claimant": the specific reasons, the plan provisions it rests on, what more is needed to complete the claim, and the plan's review procedures and time limits, including the right to sue under section 502(a) of ERISA. A group health plan's denial must also give any internal rule relied on and, for a denial on medical necessity, an explanation of the clinical judgment, or say that these are free on request.

    govinfo.govlaw.cornell.edu

    Credit

    US Department of Labor, Employee Benefits Security Administration

  2. 2016

    Vague denials lose deference

    In at least one notice, the only explanation Yale Health Plan gave the member Tiffany Halo was "SERVICE NOT AUTHORIZED". On 12 April 2016 the US Court of Appeals for the Second Circuit held that when a plan fails to comply with the claims procedure rule, the court reviews the claim de novo, deciding it afresh instead of deferring to the plan, unless the plan has procedures in full conformity with the rule and can show the failure was inadvertent and harmless.

    LessonA plan that skips the required explanation loses the deference courts usually give its decisions.

    govinfo.gov

    Credit

    The Second Circuit (Chief Judge Katzmann, Judge Lynch and Judge Arterton); Tiffany L. Halo, who sued

  3. 2023

    Denials signed in batches

    ProPublica and The Capitol Forum reported on 25 March 2023 that Cigna, which covers or administers health plans for 18 million people, including employers' plans, used a system called PXDX, "procedure-to-diagnosis", that flags for denial any claim whose treatment is not on a list approved for the diagnosis. Over two months in 2022 Cigna doctors denied more than 300,000 payment requests this way, spending an average of 1.2 seconds on each, company documents showed. Cigna called the reporting "biased and incomplete" and said the reviews take place after care is given and do not deny care.

    LessonWhen a doctor signs automated denials in bulk, the individual review the rules expect may not take place.

    propublica.org

    Credit

    Patrick Rucker (The Capitol Forum), Maya Miller and David Armstrong (ProPublica)

  4. 2023

    Disputes swamp arbitration

    Under the No Surprises Act, a provider and a health plan that disagree over payment for certain out-of-network care can send the dispute to a certified independent entity to decide. GAO reported on 12 December 2023 that the federal departments running the process had expected about 22,000 disputes in 2022 but received nearly 490,000 through June 2023, and that about 61% were unresolved in June 2023; officials named as the main cause the complexity of deciding whether each dispute was eligible.

    LessonA dispute process sized to a forecast, not to the volume that arrives, stalls at its slowest step, here the eligibility check.

    gao.gov

    Credit

    US Government Accountability Office

  5. 2024

    A clearinghouse held hostage

    A ransomware attack on Change Healthcare, part of UnitedHealth Group, whose systems handle claims, payments to providers, prior authorizations and eligibility checks, was detected on 21 February 2024 and held up claims and payments across US health care for months. UnitedHealth's chief executive, Andrew Witty, testified that attackers used stolen credentials on a remote-access portal that did not have multifactor authentication turned on, and the company paid a $22 million ransom. Change Healthcare later put the number of people affected at 192.7 million, the figure on the HHS breach portal by August 2025.

    LessonOne clearinghouse can be a single point of failure for a country's claims.

    cybersecuritydive.comhipaajournal.com

    Credit

    Reporting by Cybersecurity Dive and The HIPAA Journal

  6. 2025

    Claims settled automatically

    India's Employees' Provident Fund Organisation (EPFO), which holds the retirement savings that employers and workers pay into, settles many members' requests for advances automatically. The Tribune reported in June 2025 that 59% of advance claims in 2024–25 were settled this way, up from 31% in 2023–24, and that EPFO had raised the ceiling for automatic settlement from ₹1 lakh to ₹5 lakh (₹100,000 to ₹500,000) so that members can get funds within three days.

    tribuneindia.com

    Credit

    Employees' Provident Fund Organisation

Watch for

  1. Free appeal letters

    Counterforce Health, a nonprofit founded by Neal K. Shah, writes appeal letters against health insurance denials at no charge, based on the patient's policy and the record of successful appeals for the drug or treatment in question, NBC News reported on 18 July 2025. The article gives no success rate for the tool.

    Achieved whenAn outside study compares how often denials are overturned on appeals written with the tool against a matched group written without it.

    nbcnews.com

    Claimed by

    Counterforce Health, July 2025

Open

  1. 01

    Checking denial letters

    A checker that reads a health plan's denial letter and lists which of the items the claims procedure rule, 29 CFR 2560.503-1(g), requires are missing or too vague, with a public test set of at least 300 invented letters, some complete and some with seeded omissions, and published scores.

    The brief
    Credit · unclaimed
  2. 02

    Graded answers to benefits questions

    A public test of AI helpers that answer employees' benefits questions: at least 500 questions written against the public plan documents of ten employers (the summary plan descriptions and SBCs that public employers post), each with the correct answer and the page it comes from, including questions where the right answer depends on facts the employee has not given or where the employee has asked about the wrong benefit.

    The brief
    Credit · unclaimed

Open 01 · the brief

Checking denial letters

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A checker that reads a health plan's denial letter and lists which of the items the claims procedure rule, 29 CFR 2560.503-1(g), requires are missing or too vague, with a public test set of at least 300 invented letters, some complete and some with seeded omissions, and published scores.

Why it is still open

In the Halo case one notice said only "SERVICE NOT AUTHORIZED", and the court held such failures can cost a plan its deference. In KFF's 2023 survey, 18% of insured US adults (21% of those with job-based cover) had a claim denied in the past year; 69% of them did not know they had appeal rights and 85% never filed a formal appeal. Federal reporting of denial rates covers only HealthCare.gov plans, not group health plans, so how well employer plans explain denials is not measured. This research found AI tools that write appeals, such as Counterforce Health's, but no public test of whether denial letters meet the rule.

kff.orgkff.orglaw.cornell.edu

Built with what exists

The rule's text from Cornell's Legal Information Institute; AI models that run on the reader's own machine, so that a real letter never leaves it (see the Note "A trillion open parameters"); a second pass that lists the items a person should check, in the manner of the record reviewer in the Note "Seven more flagged".

  1. First step

    Write 30 letters from the list in paragraph (g)(1), each missing one required item, and run one model on them.

  2. Done looks like

    300 invented letters with at least 600 seeded omissions; a checker that finds at least 95% of them and flags no more than 5% of complete items wrongly; test set, checker and scores public.

Open 02 · the brief

Graded answers to benefits questions

Credit · unclaimed

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 answer employees' benefits questions: at least 500 questions written against the public plan documents of ten employers (the summary plan descriptions and SBCs that public employers post), each with the correct answer and the page it comes from, including questions where the right answer depends on facts the employee has not given or where the employee has asked about the wrong benefit.

Why it is still open

Workday's new agent "answers employees' benefits questions in plain language", by its own description, and its release gives no accuracy figures. Ai2's XYBench found that the best AI models tested steered users to their real goal 60% of the time on programming questions, against 79% for top human answers (see the Note "Answering the question behind the question"), so a benefits test needs questions of that kind. This research found no public test set for benefits questions.

investor.workday.com

Built with what exists

Plan documents that state governments and public universities post; AI models such as Claude Opus 5.5 and Mistral Large 4 as systems under test; a grading script that checks each answer against the cited page, with people grading a sample.

  1. First step

    Collect the plan documents of two public employers and write 50 questions with answers and page numbers.

  2. Done looks like

    500 questions with answers checked by two reviewers, released under an open license; scores for at least three AI helpers, each run three times, with the share of answers that are correct, the share that cite the right page and the share that ask for missing facts.