Industries / Finance and business services / Payroll / PEOs
HR support
This is the work of helping clients with the rest of being an employer: answering managers' questions, writing handbooks, running required training, handling hiring and firing, and insuring against claims by employees.
6
Milestones achieved, from 1998 to 2024.
2
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1998
An HR department off site
Administaff, founded in 1986 by Paul Sarvadi and Gerald McIntosh and now called Insperity, offered "personnel policy and employee handbook drafting and reviewing" and employed "attorneys specializing in several areas of employment law." It had 1,893 clients and 29,725 worksite employees at the end of 1997, and in 1998 it released its first web service, "Administaff Assistant." Its 2009 annual report describes employment practices liability insurance (EPLI), which covers claims such as discrimination and wrongful dismissal, "for ourselves and our clients"; TriNet's EPLI generally excludes overtime and minimum-wage claims.
Credit
Paul Sarvadi, Gerald McIntosh and Administaff
2003
The client still decides
The US Court of Appeals for the Fifth Circuit, in an opinion by Judge Edith H. Jones, held that Administaff was not liable under the WARN Act, which requires 60 days' notice of a plant closing, for its client TCS's closing of a New Jersey plant. The court described Administaff as "essentially operating as an off-site human resources department" that co-employed TCS's workers so they could receive benefits and workers' comp, and found that it "did not participate in TCS's decision."
LessonCo-employment shares paperwork, pay and benefits, but the client still answers for its own business decisions.
Credit
US Court of Appeals for the Fifth Circuit
2018
Training required by law
California's SB 1343, signed on 30 September 2018, extended required harassment training to employers with five or more employees. Every two years supervisors must receive 2 hours and other employees 1 hour, and the state's civil rights agency gives the next deadline as 1 January 2027. New York's 2018 law requires harassment training for all employees "on an annual basis." By the Congressional Research Service's count in April 2026, 18 states, counting the District of Columbia, also require private employers to provide paid sick leave.
LessonEach new state rule is a recurring task that a small business can hand to its PEO.
cooley.comcalcivilrights.ca.govnysenate.goveverycrsreport.com
Credit
California and New York legislatures
2022
A PEO's records exposed
TechCrunch reported that an attacker "may have accessed a cloud storage system that contained personal information" at the PEO Sequoia One between 22 September and 6 October 2022, including Social Security numbers and COVID-19 test results; the company would not say how many people were affected.
LessonA PEO holds identity, pay and health records for every client's workers, so one breach reaches many employers at once.
Credit
Sequoia One (a failure); TechCrunch (Carly Page), which reported it
2024
A city bot's unlawful advice
New York City's MyCity chatbot, built on Microsoft's Azure AI services, told business owners "Yes, you can take a cut of your worker's tips," which New York law forbids, as Colin Lecher of The Markup found in March 2024. In January 2026 Mayor Zohran Mamdani called the bot "functionally unusable," and the city ended its test.
LessonAn AI that answers employment-law questions for small businesses has to be tested against the law before anyone relies on it.
Credit
The Markup, which tested it; New York City (a failure)
2024
Legal AI tools measured
Researchers at Stanford's RegLab and its Institute for Human-Centered AI put over 200 legal questions to commercial AI research tools in May 2024: Westlaw's AI-Assisted Research "hallucinated more than 34% of the time," and Lexis+ AI and Ask Practical Law AI gave wrong answers "more than 17% of the time." A 2026 study by Afane, Hariri, Ouyang and Ho found 58% (Westlaw AI) and 64% (Lexis+ AI) accuracy on questions about state unemployment insurance law.
LessonTools built for lawyers still miss state-by-state questions often enough to need a known answer key.
Credit
Varun Magesh, Faiz Surani, Matthew Dahl, Mirac Suzgun, Christopher Manning and Daniel Ho (Stanford); the 2026 study's authors
Watch for
PEO assistants answer HR questions
TriNet said its TriNet Assistant would "deliver fast personalized answers to thousands of HR, payroll, and benefits questions." ADP said its HR agents give "personalized, curated answers for frequently asked questions" drawn from each client's handbooks. Insperity said its HR360 AI Agent "handles everyday HR administration questions," with "a direct path to Insperity specialists." None has published an accuracy figure.
Achieved whenOne of them publishes accuracy on a fixed set of state-specific questions graded by people outside the company.
Claimed by
TriNet, October 2025; ADP, January 2026; Insperity, October 2026
A global HR benchmark
Remote, an employer-of-record firm, published HRmark, 2,500 HR questions across 50 countries, on which "every answer is graded against that reference by a model grader." It does not say who wrote the reference answers or whether the questions are public.
Achieved whenThe questions and reference answers are published, and people outside Remote check a sample of the model's grades.
Claimed by
Remote, October 2026
Open
01
HR answers with a key
A public set of at least 300 small-business HR questions, each with the correct answer for a named US state or city and the statute or regulation it rests on, written and checked by two people, and a published score for the HR assistants PEOs offer and for general AI models.
The briefCredit · unclaimed02
Handbooks with planted errors
A public set of invented employee handbooks for several states, with planted conflicts with state law (final pay deadlines, paid sick leave, meal breaks, harassment training), and a published score for handbook-checking tools and AI models on finding them.
The briefCredit · unclaimed03
Required training as data
A free, machine-readable table of required workplace trainings by state and city: who must take each one, how many hours, how often, the employer size it starts at, and the law that requires it.
The briefCredit · unclaimed
Open 01 · the brief
HR answers with a key
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of at least 300 small-business HR questions, each with the correct answer for a named US state or city and the statute or regulation it rests on, written and checked by two people, and a published score for the HR assistants PEOs offer and for general AI models.
Why it is still open
LegalBench (2023), 162 legal reasoning tasks, is not built around state employment law. LaborBench (2025) covers state unemployment insurance only, and the Vals legal research report of October 2025 is not HR-specific. Remote's HRmark grades with a model, covers countries rather than US states, and does not say whether its questions are public. None of the PEO assistants above publishes accuracy.
Built with what exists
State labor agency pages and statutes; AI models that cite sources, such as Claude Opus 5.5; questions written the way managers ask them, since Ai2's XYBench found models steered people to the goal behind a question less often than the best human answers (see the Note "Answering the question behind the question"); a grading script for answers with fixed choices.
First step
Write 30 questions on New York wage and tip rules, including the tip question MyCity got wrong, with answers taken from state law.
Done looks like
300 questions across at least 10 states and 3 cities, the two writers' rate of agreement reported, each answer citing its source; at least four systems scored three times each; all questions and answers public.
Open 02 · the brief
Handbooks with planted errors
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of invented employee handbooks for several states, with planted conflicts with state law (final pay deadlines, paid sick leave, meal breaks, harassment training), and a published score for handbook-checking tools and AI models on finding them.
Why it is still open
This research found no such set. The nearest, HANDBOOK.md (July 2026), tests whether AI agents follow a handbook, and the best model "passes 36.2% of trials"; it does not test whether they find errors in one. Handbook-audit tools found in searches publish no accuracy figures.
Built with what exists
State statutes, such as California's training rules above; Python to plant errors; an AI model that reviews long documents, such as Claude Opus 5.5 (see the Note "Seven more flagged").
First step
Write one California handbook with ten planted errors, including harassment training hours below what SB 1343 requires.
Done looks like
40 handbooks for 10 states with 200 planted errors; a tool that finds at least 90% of them with no more than one false flag per handbook, scored by someone other than its builder.
Open 03 · the brief
Required training as data
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A free, machine-readable table of required workplace trainings by state and city: who must take each one, how many hours, how often, the employer size it starts at, and the law that requires it.
Why it is still open
A guide from Emtrain (December 2025) covers 7 states and 3 cities as a web page, with few citations and no download. SixFifty's Employment Law Informatics Project covers "over 100 employment law topics across every state and major locality," but is free only to legal academic institutions.
Built with what exists
State and city codes; a web agent to collect them; Meta's Muse Ember to pull fields from the texts (see the Note "The jobs agents repeat"); a CSV file on GitHub.
First step
Code California and New York, whose laws this research read, then the states Emtrain lists, and publish the first version.
Done looks like
All 50 states, the District of Columbia and the 20 largest cities, every field cited; two people agree on at least 95% of fields; refreshed twice a year.