Why PEO Renewal Averages Need Independent Verification

Want the plain English version? Check out the article Do Some PEOs Stay Cheap on Health Renewals, or Does Everyone Revert to the Mean? [27] It goes through the same evidence for employers and brokers and also has a published client quote comparison.

Here's how to read the labels. Direct means a source we opened states it. Interpretation is our own reading of what a source actually means. Inference is our own conclusion, drawn from several facts or from data we have. Benefitra analysis marks a result from Benefitra's own calculations on public data, not yet published in full. And PROPOSAL marks a rule that is not final yet, pending counsel and a decision from Sam Newland.

Terms. A PEO (professional employer organization) is a company that coemploys a small business's workers and handles payroll, human resources and often health benefits. WSE stands for worksite employee, meaning a client's employee whom the PEO coemploys. A renewal increase is the percentage change in what a client pays for health coverage when its contract renews. Form 5500 is the annual report most ERISA benefit plans file with the U.S. Department of Labor (DOL) through EFAST2. Schedule A is the Form 5500 attachment for each insurance contract, and it lists the carrier, persons covered, premiums and broker pay. MEWA means multiple employer welfare arrangement, and Form M-1 is its annual DOL report. ASO is administrative services only, where the plan pays its own claims (self funding). EIN / PN are the employer identification number and plan number, which together identify one plan. A wrap plan is one plan document that wraps several insurance policies. A 10-K is a public company's annual SEC report. ESAC is the Employer Services Assurance Corporation, a voluntary PEO accreditation body. And NHE refers to the CMS National Health Expenditure accounts.

1 An “average renewal” has no standard definition

1.1 What PEOs publish

Some PEOs do release renewal statistics and one we opened is Rippling's 2026 PEO Trust and Transparency Report [15]. It states that “84% will see a medical rate increase at or below our 9% base renewal, averaging just 6.5% across all benefits carriers,” and also that “another 5% will land under 17%,” and that “only 11% will see an increase higher than that.” A footnote explains that “the 6.5% average increase reflects renewal outcomes for the cohort ... who received an increase at or below our 9% base rate,” for customers “who've joined the PEO on or after 12/31/2024.” Direct

Rippling reveals those definitions and that's the kind of disclosure this paper argues for. But the example also shows why definitions matter. The 6.5% figure is based only on the 84% at or below the base rate and doesn't reflect every renewing customer. Using just the bands that Rippling published, the average for all renewing customers has to be at least about 7.8% (0.84 × 6.5% + 0.05 × 9% + 0.11 × 17%, using the lowest possible value in each upper band). The real number might be higher. Why? The report does not specify how far over 17% the top 11% went. There are also two caveats. The bands talk about medical rate increases but the 6.5% is averaged “across all benefits carriers”. And a group joining on or after December 31, 2024 will include clients who have under 12 months of tenure at the 12/1/2026 renewal. Inference (our arithmetic) Rippling does not claim that 6.5% is the average for all customers.

1.2 The choices behind any “average”

1.3 What rules exist, and what they do not cover

• Federal ERISA reporting. Neither Form 5500 nor Form M-1 has a field for a renewal increase or a rate method [1][2]. Inference
• IRS certification of PEOs. Rules for certified PEOs cover financial and employment tax requirements. Benefit pricing is not addressed in them (prior Benefitra review of 26 CFR 301.7705 to 2 [20]. Not reopened for this draft).
• State PEO laws. We looked at the PEO statutes in Texas and Florida. They require certain benefit disclosures but do not specify a method for a renewal average [20]. We did not survey all 50 states.
• Industry accreditation. Here's what ESAC's standards say. “information provided to Clients and prospective Clients by Accredited PEOs shall not contain incorrect or misleading information,” and “PEO participation in ESAC programs is voluntary” [14]. Direct What they don't do is define how a renewal average should be computed. Inference
• Securities law for public PEOs. A public PEO has to discuss known trends and uncertainties as part of the management discussion in its 10-K (Regulation S-K Item 303 [13]). Direct Each company chooses its own metric for health insurance costs. Insperity reports cost “per covered employee” [11] and TriNet uses an “insurance cost ratio” [12]. Neither of these is a distribution of client renewals. Inference
• General consumer protection law. State laws that forbid unfair or deceptive practices deal with misleading statements in general. Interpretation We haven't reviewed specific statutes though.

Inference PEO renewal marketing does fall under general rules against misleading statements. But we haven't found any rule that tells a PEO how to calculate or disclose an “average renewal.” Having a standard and voluntary way of disclosure would fill that gap.

2 Do PEOs file Form 5500? Per plan, not per carrier

2.1 Yes, for their sponsored health plans

• Direct According to the DOL's MEWA guide, “a PEO plan or arrangement that offers or provides health coverage to employees of two or more client employers ... is a MEWA,” and being a coemployer under other laws does not resolve the question for a PEO [6].
• Direct Plans that need to file Form M-1 “are required to file the Form 5500 regardless of the plan size or type of funding” [1]. The small plan exemption in 29 CFR 2520.104-20 does not apply to them [4]. Form M-1 is due on March 1 each year [5].
• Direct Form 5500 filings use EFAST2 and those are publicly available [1]. The DOL also releases bulk data for each form year [7].
• Direct (data) In their 2024 filings, Insperity, CoAdvantage and Paychex Business Solutions said “yes” to whether they were subject to Form M-1. ADP TotalSource, Justworks and TriNet said “no” [7]. In its 10-K, Insperity calls its plan “a single employer plan” [11]. Inference That Form M-1 answer is the sponsor's own legal view rather than a DOL finding. A “no” shouldn't be read as wrongdoing.

2.2 The rule. One Form 5500 per plan, one Schedule A per contract

Say a PEO uses three different insurance carriers like UnitedHealthcare, Aetna and Cigna. Does that mean it files three Form 5500s? Not because of the carriers, no. The rules tie filings to plans.

1. For each plan there is one filing. “One Form 5500 is generally filed for each plan or entity” [1].
2. For each insurance contract there is one Schedule A. “Provide information for each contract on a separate Schedule A. Individual contracts grouped as a unit in Parts II and III can be reported on a single Schedule A” [2].
3. Self funded administration gets no Schedule A at all. “Do not file Schedule A for a contract that is an Administrative Services Only (ASO) contract” [1].
4. Separate policies do not mean separate plans. “The fact that you have separate insurance policies for each different welfare benefit does not necessarily mean that you have separate plans”. Sponsors can use a “wrap” document, and to figure it out “you must review the governing documents and actual operations” [1].
5. Each plan gets its own number and for welfare plans this numbering starts with 501 [1].

Interpretation Take a PEO that has three carriers under one wrap plan. It files one Form 5500 which has at least three Schedule As attached, and often many more, because each regional carrier entity or policy is usually a separate contract. It files three Form 5500s only if there are three plans set up or if different legal entities sponsor different plans.

2.3 What real PEO filings show

We checked bulk EFAST2 data from the DOL for form year 2024 (plan years starting in 2024). We downloaded this data on August 25, 2026 [7]. The counts refer to the Schedule As that go with each filing and, among those, the contracts flagged as medical (health, HMO, PPO or indemnity). For insurance carriers we counted distinct carrier legal names as they were filed.

• Per plan rather than per carrier. Direct Justworks files three carrier families in a single filing. ADP includes 34 carrier entities in one.
• Splitting by benefit type. Direct Insperity puts medical, dental and vision in plan 502, and life and disability in plan 504. If you're comparing “plan participants” across PEOs you have to know what benefits each plan actually covers.
• Splitting by legal entity. Direct TriNet uses coemployment through several subsidiaries and each subsidiary files its own plans. You need all of them to get a full picture of TriNet. TriNet HR III plans 503 and 504 report medical benefits paid out of general assets, and there is no Schedule A. Why? We have not identified that yet. Inference
• Splitting by carrier or region. Direct Most of the 14 medical plans of G&A Partners use just one carrier. This is G&A's plan design and not a filing rule. Interpretation
• Cafeteria and FSA plans are separate filings, and they carry no premium data. Direct TriNet, G&A, Vensure, CoAdvantage, Insperity and Paychex each file a separate Section 125 or FSA plan. The IRS suspended the old Schedule F requirement in Notice 2002 to 24 [10]. Inference With no insurance premiums in them, these should be left out of any premium analysis.

Inference Here is a practical guideline for analysts. Find every EIN used by the PEO group and list every welfare plan number under each one. Keep the plans that have medical coverage and then sum up their medical Schedule As. Counting filings would be wrong for most PEOs listed in Table 1, and assuming one filing per carrier would be wrong for them too.

3 Why Form 5500 cannot tell you the renewal increase

Form 5500 provides useful context. But it isn't a renewal measure.

1. There's no field for renewals and no client level data either. What Schedule A reports instead are plan level premiums, claims for experience rated contracts, the number of persons covered and payments to brokers [2]. There is nothing on it that says what any one client was charged or how rates for that client moved. Inference
2. Mix and survivorship. When the client mix shifts, so does a plan's average cost. Inference And turnover is big enough to matter. Insperity reported client retention of about 83% in 2025 and 81% in 2024, and it lists “client nonrenewal due to price or service factors” as one of the causes of attrition [11]. TriNet said total worksite employees fell 10% in 2025, “primarily due to declines in our Technology, Professional Services, and Main Street verticals” and “partially attributable to necessary repricing of our health benefits services”. Average worksite employees dropped 5% [12]. Direct We go through this in Section 4.
3. Plan buy downs. Say clients move to plans with higher deductibles. Then premiums per person can go up less than the rates did, and Form 5500 does not report changes in plan design. Inference
4. There are self funded and partially self funded portions. ASO contracts do not appear on Schedule A [1]. Insperity says it accounts for its UnitedHealthcare contract “using a partially self funded insurance accounting model” [11]. And according to TriNet, under its insurance policies “we bear claims costs up to a defined deductible amount” [12]. Direct The Schedule A “premium” may not match the economic cost. And it isn't what clients actually paid to the PEO. Inference
5. Headcounts that don't match up. Form 5500 counts participants, and “covered dependents are not counted as participants” [1]. Schedule A though gives the “approximate number of persons covered” for each contract, which is a number the administrator “should estimate” [1][2]. Direct Examples Inference from 7. For the 2024 plan year, ADP's medical contracts report about 1.73 million persons covered compared to about 531,000 active participants. That's a ratio of about 3.3. For plan year 2024, TriNet HR III's medical Schedule A premium went down 56% (from about $1,249 million to about $549 million), while the number of persons covered dropped about 12%. Why? We haven't identified the reason. And for Justworks in plan year 2021, premium per covered person went up 19% while premium per active participant went up 57%, because the two headcounts moved in opposite directions.
6. Lag. A filing has to be in by “the last day of the 7th calendar month after the end of the plan” year, and extensions can't go beyond “a total of 9½ months beyond the close of the plan year” [1]. The filing for Insperity's plan year 2024 was received on October 14, 2025. CoAdvantage's plan year starting in October 2024 was received on July 15, 2026 [7]. Direct PEOs are quoting 2027 renewals, while for most PEO plans the most recent complete public year is still plan year 2024. With an extension, plans for calendar year 2025 must be filed by October 15, 2026, so Insperity's plan year 2025 filing may appear soon after this preview.
7. Plan year versus calendar year. Plan years start at different times. Insperity's starts in January, ADP's in June, CoAdvantage's and TriNet HR III's in October, and Justworks' in November [7]. A year over year change covers different calendar periods from one PEO to the next, and the CMS NHE benchmark runs on the calendar year.
8. What “premium” includes. Schedule A separates experience rated contracts from the ones that aren't experience rated, and it lists broker commissions separately [2]. You won't see the PEO's own markup or service fee there. Choose a different premium definition and you get a different answer. Inference
9. Multiple plans, entities and acquisitions. After an acquisition or consolidation, enrollment can be moved from one filing to another. And several of the “latest” filings in the data are actually amendments, filed one to three years after the original [7].

3.1 Units used in this paper

3.2 An illustration. Three series

There are three series that hold up well enough in Form 5500 to show a cost trend. Those are Insperity and CoAdvantage from plan year 2019 and Justworks from plan year 2021. Justworks' 2019 to 2021 values fail the data quality rules so we flag them. ADP, TriNet and Paychex? They have the headcount or completeness issues described above.

How we computed it. We take experience rated earned premium and add non experience rated total charges on Schedule A contracts flagged health, HMO, PPO or indemnity, then divide by the total persons covered on those same contracts. One flag. In plan year 2020 Justworks persons covered went up 65%. That breaks the headcount and premium jump rules in Appendix A (rule D3), so its 2019 to 2021 changes are not price changes.

Figure 1. Insured medical premium per covered person by plan year, taken from Table 2. It is not the renewal increase clients were offered. Justworks 2019 to 2021 is drawn dashed since it fails the data quality rules. Source. EFAST2 data 7. Benefitra calculation 9.

Inference How should we read this? CoAdvantage's 6.5% drop in 2023 could be due to mix, buy downs or changes in counting, and we can't tell from the filing which one it is. Insperity's 5.0% for plan year 2024 is close to the 4.3% rise in “cost per covered employee” reported in its 10-K for 2024 [11]. But they are different measures. The 10-K number is per covered employee and it includes claim costs that Insperity itself bears under its partially self funded arrangement. For 2025, Insperity's 10-K reports an increase of 9.2% in cost per covered employee, along with “higher average pricing of 3%” in revenue per worksite employee [11] Direct. The 10-K report is newer, and the Form 5500 has more detail. Neither one shows how client renewals are distributed.

4 Does Form 5500 control for clients a PEO loses?

No. Form 5500 is a snapshot of the whole plan. For PEO plans there is no detail at the client level and no place for groups that have left. So it can't account for attrition among clients and it can't show which clients left or why. Inference from 1, 2, 21, 22

4.1 What the form actually records

4.2 Why this matters. Survivorship

Inference When a PEO reprices at renewal, the clients hit with the biggest increases are the ones most likely to shop around or leave. They may also be the groups with higher claims. What's left in a plan level filing is then mostly the clients who stayed, and that has two effects.

1. Survivors can look cheaper. Say higher cost groups drop out. Next year's premium per covered person is then calculated on a lower cost pool, so the trend in Table 2 can rise less than the increases clients were offered.
2. Timing adds noise and it can work either way. Schedule A premium covers the whole contract year, but people covered are counted at the end of it [1]. A group that leaves part way through the year adds premium but no year end persons. Groups that join midyear do the opposite.

You can't measure or remove either effect based only on Form 5500 data.

4.3 Indirect proxies and their limits

• Participant numbers at the beginning and end of the year. They show only the net change. If you lose 20% of participants and gain 20%, the plan shows no change.
• Year over year growth in headcount. Same issue, it's only net. It also mixes hiring and layoffs inside clients who stayed with whole clients joining or leaving.
• Lists of participating employers (multiple employer filers only). You could tally employers that joined or left from these, but only by reading the PDF of each filing, and they give no rates and no reasons.
• 10-K disclosures (public PEOs only). These report client retention or WSE counts across the whole company. Nothing at the plan or cohort level.

Inference from 7 Direct from 11 Here's an example of the gap. Insperity's Group Health Plan had 207,449 active participants at the end of plan year 2023 and 202,013 at the end of plan year 2024. That's a net drop of around 2.6%. Yet in its 10-K for 2024, Insperity put client retention at about 81%. So roughly one client in five left. Clients and participants are not the same unit, so we can't reconcile the figures exactly. But they do show how much gross turnover can hide behind a small net change.

4.4 Evidence that exits are material

• Direct Insperity. Client retention was about 83% in 2025 and 81% in 2024, and its 10-K names “client nonrenewal due to price or service factors” as one of the causes of client attrition [11].
• Direct TriNet. Total worksite employees fell by 10% in 2025, “primarily due to declines in our Technology, Professional Services, and Main Street verticals” and “partially attributable to necessary repricing of our health benefits services” [12]. For TriNet HR III's main plan (plan year starting October 2024) there were 245,531 participants at the beginning of the year and 220,810 active participants by the end [7]. Because that filing's medical premium moved very differently from its headcount (Section 3), we don't compute a premium per person trend from it.

Inference Exits are large enough to move a plan level average. That is why the gold standard in Section 6 keeps lost clients in the cohort at the rate they were offered. And it puts retention and the actual premium change right next to the headline.

5 What 20 years of filings show, and what they can't

Benefitra's companion analysis [23] goes back as far as the public data allow. It read every Form 5500 main filing for filing years 1999 through 2025, plus Schedule A and, if one was filed, Schedule H. The search looked at 33 named PEOs (17 on Benefitra's initial list plus 16 from the ESAC accredited directory), checking name variants and EINs. The companion report also calls this “about 45 PEO groups”. We haven't yet reconciled those two counts, and the full search list will be published along with the data. The analysis looked at two things. Do PEO health insurance premiums go up and down in a pattern? And how much of a long term trend could come from churn, client mix or self funding rather than price?

Three caveats first. All figures are insured premium per covered person or per participant as reported on Form 5500. None of them is a renewal increase. Most tests use 5 to 22 yearly observations for each PEO, so take the p values as just indicative, and a real effect of 1 percentage point a year would often go undetected. And the sample itself is survivor biased. Only PEOs that keep filing clean data make it into the long windows. All of this is Benefitra analysis of public data, and the analysis isn't fully published yet.

5.1 Which PEOs have 10, 15 or 20 clean years

A clean run means plan years in a row with no break. What counts as a break varies by metric. For per covered person, it's a missing year, persons covered ÷ participants shifting more than 30%, or premium per person jumping more than 30%. For per participant, it's a missing year, a jump of more than 30%, or participant numbers stuck at the 32,767 legacy field cap. The per participant rule has no headcount ratio test, so a per participant run can contain headcount restatements (see ADP below). How long is long enough? For a 10 year window you need at least 11 clean yearly levels, 15 years need 16, and 20 years need 21.

Not one per person series reaches 20 years. AdvanStaff HR (ProService group) qualifies per person only for 2009 to 2020, at 6.2% a year, and per participant for 2011 to 2024. Most large PEOs do file a health plan under their own EIN, but their series fall short for other reasons.

There are some other patterns too. Some plans are very small (Emplicity, Employer Solutions Group, AmCheck). Under the names and EINs we searched, the only plans found for Asure were for Asure's own employees, and no PEO sponsored health plan came up for Key HR or Lone Star PEO. A PEO may offer coverage through plans sponsored by insurance carriers or clients, and those plans never show up under the PEO's own EIN. So “not found” does not mean a PEO offers no health coverage. Because there is no bulk dataset for Form M-1, MEWA reports couldn't be used.

5.2 No detectable high low pattern, and why that doesn't settle mean reversion

People often wonder if a PEO that had a bad year is “due” for a good one. The pooled filings don't show that.

• Over 10 series and 76 yearly changes pooled together, the correlation between one year's change in premium per person and the next is −0.23 (lag 1 autocorrelation). But short samples alone would give about −0.15, and a permutation test yields p = 0.35.
• Over 1 to 5 years, the spread of growth rates across PEOs shrinks about as fast as simple averaging would predict (standard deviation 10.6, 6.0 and 4.7 points at 1, 3 and 5 years, compared to 10.6, 6.1 and 4.7 predicted). It narrows faster at 7 and 10 years (3.1 and 2.8 versus 4.0 and 3.4). However only about two PEOs get to those windows.
• Two single PEO tests do show reversion type results. AdvanStaff HR's regression slope for 3 years ahead is −0.56 (p = 0.004), but it relies on just 5 nonoverlapping observations. In the first version of the analysis, the per participant series for ADP TotalSource shows a slope of about −0.41 (p < 0.01). This is fragile, because it's driven largely by the 2012 headcount jump.
• One off counting errors also give a mild negative tilt. A headcount error makes one year look higher and the next year look lower. Interpretation

Inference Public filings can't prove or disprove mean reversion in PEO health insurance renewals. Why not? Premium per participant moves with churn, buy downs and mix, so “no detectable pattern” in the filings doesn't tell you much about how a particular client's renewals behave. Only offered renewals for each client could settle it, and that includes the clients who left. The companion article goes through this for employers.

5.3 Insperity. Close to the national trend from 2006, below it over 10 and 15 years

Insperity has the best long series. But what it shows depends on which window you pick.

Figure 2. The left panel shows long series indexed to 2008 = 100. These are Insperity premium per covered person and per participant, ADP TotalSource premium per participant, CMS NHE employer sponsored insurance per enrollee, and AHRQ MEPS-IC single and family premiums. The right panel shows premium per covered person for clean runs only, indexed to 2019 = 100, for Insperity, ADP TotalSource, TriNet HR III, CoAdvantage, G&A Partners plan 504, Sequoia One and CMS NHE. All are insured premium per covered person or participant, not the renewal increase clients were offered. Source. Benefitra companion analysis 23, chart 01.

Benefitra analysis The things the filings can measure account for well under 1 percentage point a year of Insperity's trend.

• Mix. More dependents per covered employee (1.97 to 2.13) add about 0.4 points a year to growth per participant. They add none to growth per person.
• Timing of headcounts. This comes to about zero.
• Carrier margin. Claims outgrew premium by about 0.2 points a year.
• Changes in carriers and acquisitions. There's no detectable effect, and Insperity did not acquire any PEO client books.
• Coverage selection. The share of worksite employees enrolled dropped from 72.7% to 64.0%, and we're not sure which way that pushes the trend.

From 2006 to 2024, Insperity's 10-K cost per covered employee rose 4.0% a year, compared to 4.8% per participant in the Form 5500 data. Yearly changes in the two sources correlate at r = 0.52 (p = 0.02). Client retention rates at Insperity ranged from 71% to 86%. And years that saw a bigger 10-K cost increase were followed by lower client retention the following year (R = −0.42, p = 0.05, borderline. Spearman −0.54, p = 0.009. N = 22). Note this is one of several low power tests, and no correction has been made for multiple comparisons. It fits with clients leaving after big increases. But we can't measure from filings how much that moves the premium trend. Inference

5.4 ADP TotalSource. A gap the filings only partly explain

From plan years 2005 to 2024, premium per participant for ADP TotalSource rose about 2.1% a year. NHE rose about 4.4%, leaving a gap of about 2.3 points a year. Change the starting year, though, and the results vary quite a bit. It's 2.7% a year for 2006 to 2024 (gap about 1.6 points), 2.3% for 2014 to 2024 (gap about 2.6) and 0.6% for 2009 to 2024 (gap about 3.4). ADP's plan year runs from June to May, so plan year t falls mostly in ADP fiscal year t+1. We haven't aligned NHE (calendar year) for this. Over 19 years this doesn't matter much. Benefitra analysis

• Participant counting accounts for about 1 point a year. Form 5500 participants rose from about 59% to about 75% of average worksite employees in the 10-K. Premium per 10-K worksite employee rose 2.8% a year over fiscal 2006 to 2024, compared to 1.8% per participant over those same fiscal years. Most of that increase in the enrollment share came in one jump, from fiscal 2012 to 2013 (plan year 2012. Participants +42%). Then in fiscal 2025 the share dropped back to about 71%. Was it a change in who gets counted, or actual growth in enrollment? The filings can't tell us.
• The remaining 1.3 to 1.5 points a year are unexplained. Treat this as approximate, since it mixes the plan year 2005 to 2024 gap with the fiscal 2006 to 2024 headcount effect. Some possible reasons include plan buy downs, a younger or healthier client base, carrier contracting and churn driven selection. None of these can be measured without ADP level data.
• Self funding and timing explain about nothing here. Benefits paid directly dropped from 3.2% to 1.3% of the plan's benefits.
• ADP's own accounting also tells a similar story. Its 10-K “zero margin benefits pass through” per average worksite employee rose 2.7% a year from fiscal 2017 to 2025. So the low trend is not just an artifact of Form 5500. But that measure is per worksite employee too, so it also depends on the enrollment share. We haven't rechecked this figure against each 10-K.
• Client retention. ADP doesn't publish an annual PEO client retention rate. Its fiscal 2025 and 2026 10-Ks estimate PEO client retention at “approximately 6 years,” down from “approximately 7 years” in fiscal 2016 and 2021 [25][26] Direct . Six years comes to roughly one client in six leaving each year (our arithmetic), which is about the same as Insperity's attrition. For ADP, though, churn can't be tested year by year.

Figure 3. Left panel. ADP TotalSource Form 5500 premium per participant, Form 5500 premium per 10-K average worksite employee, and 10-K benefits pass through per average worksite employee, indexed to fiscal 2017 = 100 (fiscal year = plan year + 1). The fiscal 2026 point is the pass through series only, from the fiscal 2026 10-K. There's no matching Form 5500 year yet. The spike and drop in premium per participant in fiscal 2011 to 2013 lines up with the 2012 headcount restatement. Right panel. Form 5500 participants as a percentage of 10-K average worksite employees (worksite employee levels are chained from one stated figure, so small errors compound). These are not renewal increases. Source. Benefitra companion analysis 23, chart v2_02.

Benefitra client case (published) The article's ADP section adds one published side by side quote. It's for an 11 employee Boston law firm. There, a Prestige PEO quote with a Cigna PPO came in 29.7% below ADP TotalSource's comparable HMO quote for year one [28]. That's one group and one quoting cycle. It isn't evidence about ADP's book, and Benefitra was the broker on the case. The article has the details.

5.5 TriNet, Paychex and the pooled data

• TriNet. From plan year 2020 on, Schedule H classifies 80% to 82% of the main plan's benefits as paid directly. This may be at least partly due to an accounting reclassification. Either way, TriNet's Schedule A premium understates its total cost, and its premium only series can't be compared to Insperity's. Total plan cost per participant rose 4.0% a year (2016 to 2021). TriNet's 10-K insurance costs per average worksite employee rose 4.2% a year (2019 to 2023), whereas Schedule A premium per person rose 2.5% a year. There are two caveats though. TriNet's “insurance costs” include workers' compensation, so they are not just health. And starting in 2024, its worksite employee count includes COBRA and platform users. Benefitra analysis
• Paychex. It only discloses company wide client retention (81% to 82% for fiscal 2015 to 2019, mostly payroll clients), and no PEO worksite employee counts. Its Form 5500 series also breaks too often to decompose.
• All PEOs pooled (13 per person series and 100 yearly changes, a different sample from the autocorrelation test). Premium per person moves strongly opposite to changes in persons per participant (coefficient −0.70, p < 0.001). It also moves against participant growth (−0.36, p < 0.001). Much of the year to year variation comes from how headcount is defined and timing rather than price. Changes of carrier show no effect (p = 0.25). Benefitra analysis

5.6 What the filings cannot measure, and the question that would

A tier mix question (enrollment by employee only, plus spouse, plus children, family) is a candidate for the next cycle. Right now it can't be added, because questions stay frozen during a cycle.

5.7 What this means

• Take the one clean long series (Insperity). Public filings back up “roughly the market trend, plus or minus mix and funding, depending on the window.” They do not back up “PEOs beat the market over time.” Inference
• Part of ADP's lower trend comes from a counting effect. The rest remains unexplained. Inference
• Just because there's a lower premium trend over the long run does not mean renewals are lower. It might be due to buy downs, or a changing client base, or survivors (Section 4). Only PEO data from the gold standard cohort can distinguish between them.

6 What the Index method aims to do

The Benefitra PEO Renewal Transparency Index is a proposed, voluntary disclosure method. Each in scope PEO gets asked the same questions. Then we publish whatever each one provided, laid out in the same way for all of them. What gets rated is how fully a PEO discloses its renewal figures and supports them. Price isn't rated, and neither is service quality or whether a PEO fits a particular employer. Appendix A contains the full rules and this section goes through them.

6.1 The gold standard. One cohort, four parts

Benefitra proposal This isn't a regulatory requirement.

• Main cohort. This is every client that had 12 months of experience or more on its renewal date and was offered a renewal during the period. Clients that got an offer and then left are included too, at the rate they were offered. We never compute any part on accepted rates alone.
• Newer clients (under 12 months at renewal) have their own separate figure. They are never mixed in.

Part 1. Headline. This is the like for like renewal at book level.

Headline = (sum of premium at offered renewal rates ÷ sum of premium at prior-year rates) − 1, for the same enrollment on the same plans, networks and deductibles, for every client in the main cohort.

Inference (arithmetic) This gives you the premium weighted mean of like for like increases at the client level, where each client counts in proportion to its share of premium from the prior year. Churn can't change it, because clients who left stay in at the rate they were offered (as long as the PEO can produce the offered rate for every leaver. See Appendix A, E10). Buy downs can't change it either, since plan, network and deductible stay constant. And the choice of average can't change it, because there is just one defined calculation and, once enrollment stays fixed, you get the same result per covered member as per enrolled employee.

Part 2. Distribution. It's the median like for like increase per client along with the shares of clients over 10% and over 20%. Why bother? So large clients can't hide what small groups go through.

Part 3. Claims versus premium (the loss ratio), included for context. It's incurred claims divided by premium for the same cohort and plan year, with definitions given. It is not the renewal measure.

In Part 4 we show retention and actual premium change next to the headline. We report the share of main cohort clients (and of their employees and premium) that renewed, along with the premium change actually observed after exits and plan changes. Then compare that with the headline. The difference shows how much churn and buy downs can make an advertised average look lower than what clients were offered.

6.2 Evidence tiers

• Tier 1 (best). Data at the client level is anonymized and under NDA, with enough detail for Benefitra to recompute every part by itself (the record layout is in Appendix A, E8). It doesn't contain any protected health information. Guidance from HHS says that aggregate figures which don't identify individuals fall outside of PHI [19]. Small cells are suppressed.
• Tier 2. Signed written answers. At this level the PEO does the calculations and a named officer signs off on them. Benefitra can't recompute these numbers, so we publish them as the PEO's own statements.
• Tier 3 (context only). Public filings, which means Form 5500 (all EINs and plan numbers, and only medical Schedule As) and 10-K disclosures for public PEOs. This layer can raise a question but it can't confirm or refute a renewal figure, and on its own it never raises a PEO's level.

6.3 Disclosure levels

Each PEO gets one level, from 0 to 4. This isn't a score. There aren't any weights, points or stars either. Levels 1 and 2 are cumulative. Levels 3 and 4 both build on Level 2, but Level 4 does not need Level 3. So a PEO never has to hand client level data to a broker to get to the highest level.

Inference Some PEOs may not track rates offered to clients who have left, or premium at prior year rates for the same enrollment, unless they take extra steps. So in the first cycle most PEOs may end up at Level 0 or 1. That shows what is disclosed today, but it says nothing about pricing.

6.4 Separate performance panel

Each PEO page shows the level along with the date it was assigned and the questions answered. In a visibly separate panel it shows the reported figures, each tagged with its evidence tier, and if a figure is missing it says “not provided.” Price never affects the level. Consider two PEOs with headlines of 4% and 14%. If both disclose the same things with the same support, they get the same level. Interpretation Grading price would punish candor (counting lost clients at offered rates tends to push the headline up), and it would look like steering by a broker who gets paid commissions and sells alternatives to PEOs.

6.5 Reconciliation note

Benefitra compares the actual premium change a PEO reports against public comparators for the same PEO and time period. If the difference exceeds the published tolerance, we add a note. There is never a deduction and never a level change, and the PEO's reply is published right next to it. PROPOSAL The v0.4 draft used a tolerance of 5 points, which is tighter than the noise in the comparators themselves. Take a look at the numbers. The cross PEO spread of one year Form 5500 changes is about 10.6 points, Table 2 shows single year swings ranging from 6 to 19 points, and Insperity's yearly Form 5500 and 10-K changes correlate at only r = 0.52. A 5 point rule would flag honest figures often. This version widens the tolerance to 10 points. It also requires that the difference show up against every available comparator, and a back test of the false flag rate on Insperity's 10-K and Form 5500 history before launch. See Appendix A, E13 for the details and the unit matching rules.

6.6 Annual cycle and conduct

Before sending out the survey we publish the questions, level criteria, tolerance, tier labels and the in scope PEO list, with a version number and a SHA-256 fingerprint. Then they are frozen for the cycle. Every PEO gets the same questions, the same deadline and one follow up. We publish answers exactly as given and state declines as facts. Each PEO sees its draft entry at least 10 business days before publication and can respond or dispute its level (see Appendix A, sections A and F).

7 Methods and data

• Sections 2 through 4 use bulk data from DOL EFAST2 (the f_5500 and Schedule A files) for form years 2019 to 2025, downloaded on August 25, 2026 [7]. The form and instruction text comes from both the 2024 and 2025 editions [1][2][3][21][22].
• Section 5 uses all Form 5500 main filings for filing years 1999 through 2025, along with Schedule A and, if one was filed, Schedule H. Files from 2009 and later are DOL's “Latest” datasets (organized by form year). Earlier files are arranged by plan year end date. We read the SEC 10-K numbers for Insperity, ADP, TriNet and Paychex by hand. Our benchmark is CMS NHE Table 21, employer sponsored private health insurance per enrollee, for calendar years 1987 to 2024 [24].
• Filters. We retained welfare plans with health coverage (benefit code 4A) and Schedule A contracts flagged health, HMO, PPO or indemnity. We did not include cafeteria, FSA or nonmedical plans. For each EIN, plan number and plan year start, we kept only the most recent filing. Premium is the experience rated earned premium plus the non experience rated total charges.
• Tests. We used compound annual growth rates over stated time periods, and a pooled lag 1 autocorrelation of year over year changes that we checked with a permutation test. We compared cross PEO dispersion with what averaging predicts (SD ÷ √years). We also ran 3 year ahead regressions with HAC standard errors, along with a pooled regression with PEO fixed effects. We did not compute confidence intervals and instead show ranges across windows.
• The plan level CSV lists the EFAST2 acknowledgment IDs for each filing used (for filings from 2009 on. Earlier filings carry DOL's pre-EFAST2 filing ID). See Data availability.

8 Limitations

• Our Form 5500 analysis covers plans we could identify by sponsor name and EIN. We might have missed plans filed under unrelated entity names, and PEOs whose clients sponsor their own plans.
• Premium per covered person depends on how we define premium and on carriers' estimates of persons covered. If you pick another reasonable definition you will end up with different numbers.
• We did not open individual filing PDFs (these contain broker names and lists of participating employers), and we also did not query the Form M-1 database [8].
• The explanations we give for unusual data points (the ADP headcount ratio, TriNet HR III plan year 2024, TriNet HR III plans 503 and 504) are based on inference. We haven't actually asked the companies.
• Long run results are sensitive to the start and end years (Tables 5 and Section 5.4). The sample is also biased towards survivors, because only PEOs with long, clean filing records make it into the longer windows. With 5 to 22 yearly points the tests have low power. We did not compute any confidence intervals, and the results show associations rather than causes.
• Plan year 2024 is the newest Form 5500 year. For calendar year plans, plan year 2025 filings are due by October 15, 2026.
• We didn't survey state law comprehensively.
• The level criteria, question map, reconciliation tolerance, scope rules and suppression thresholds are all proposals. The gold standard is Benefitra's proposal. This is not an industry or regulatory definition.
• We go into detail about one published renewal statistic (Rippling). We did not review every PEO's marketing materials.
• Benefitra is not independent of the PEO market (see Disclosures). We design the method so that levels don't rely on Benefitra's judgment of price. And still, the author and owner is Sam Newland, and there is no external reviewer yet.

9 Next steps

1. Complete the EIN and plan number map for each in scope PEO, with EFAST2 acknowledgment IDs, and then compare it against the filing PDFs.
2. When plan year 2025 filings arrive, refresh Tables 2 and 5.
3. Back test the reconciliation tolerance and then publish the false flag rate.
4. Pilot the questionnaire with a few PEOs, then refine the wording and the deadline.
5. Publish the method, questions and in scope list with a SHA-256 fingerprint. Then send out the survey to every in scope PEO on the same day.

10 Disclosures, license, citation and changelog

License

© 2026 Benefitra. Text and tables derived by Benefitra are licensed under CC BY 4.0 (creativecommons.org/licenses/by/4.0/). You may share and adapt them as long as you give credit to Benefitra and a link to this page. Third party data and quotations (DOL, SEC filings, CMS, AHRQ, Rippling, ESAC and PEO documents) stay subject to the terms of their owners. Benefitra's analysis code is not published.

How to cite

Newland, S. (2026). Why PEO Renewal Averages Need Independent Verification. The Benefitra PEO Renewal Transparency Index (Version 0.6). Benefitra. https://benefitra.com/peo-renewal-transparency-index/. Accessed .

Data availability

All inputs are public. They include DOL Form 5500 bulk data [7], SEC 10-K filings and CMS NHE tables. The derived plan level data behind Tables 2 to 5 and Figures 1 to 3 is published on benefitra.com along with this paper, as a CSV (peo-plan-level.csv) with a MANIFEST of SHA-256 hashes. Any other requests should go to the corrections contact below.

Corrections contact. info@benefitra.com. Each correction is logged below along with the date, what was incorrect, what changed and which numbers moved.

Changelog and errata

v0.1 to v0.4 were internal drafts written on the same day and never published. Errata. The Table 3 and Table 4 corrections in v0.6.

11 References

A Appendix A. How the Index rates a PEO (Scorecard Standards)

Status. These rules carry the version number of this paper (v0.6), and they will become v1.0.0 at launch. They replace the standalone Scorecard Standards drafts, v0.1 to v0.3.1 (September 26, 2026). The last standalone file had a mismatch (v0.3.1 in its header, v0.3 in its status line), and that file is now retired. The paper, these rules and the questionnaire now share one version number. Each rule has an ID and is written so that a reviewer can mark it pass or fail. A page or release ships only if every applicable rule passes (section G). Anything marked PROPOSAL needs to get a decision from Sam Newland and counsel before it is first used.

A. Conduct and safety

B. Reader trust and search

Google's guidelines ask whether authorship is “self-evident,” whether readers can see “how the tests were conducted,” and whether there is disclosure about automation use. They also say that review snippets for organizations need ratings “sourced directly from users” [16][17][18]. Direct

C. Site structure on benefitra.com

D. Data handling

E. Index rules. Gold standard, tiers, questionnaire and levels

In the standalone drafts these rules used to be numbered G1 to G15. That conflicted with a Sources section also labelled G. They are now renumbered E1 to E15.

F. Scope. Which PEOs are in the Index

G. Prepublish checklist

Tick every box. A single “no” blocks publication.

H. Standards changelog

• v0.5-preview (2026-09-29). Folded into the paper as Appendix A. Changes. Scope rules (F) and dispute path (A15) added. Score language removed and listing order alphabetical only. A9 scoped to allowed uses. Level 3 renamed “Recomputed by Benefitra” and Level 4 “Independently attested,” with Level 4 reachable without Level 3. Q20 split into Q20a and Q20b. Leavers' offered rate rule (E10). Tolerance widened from 5 to 10 points with reasoning, unit matching and a required back test (E13). D3 flags harmonized with the research rules. D8 and the annual cycle made consistent. B1 disclosure made two way. Right of reply extended to research papers. The MIT code license, code downloads and repository citation file removed. G rules renumbered E. An internal placeholder for the decision maker replaced with Sam Newland. Drafting folder references removed.
• v0.3.1 (2026-09-26). Cross references updated to paper v0.4 (no rule changes). Its status line wrongly read v0.3.
• v0.3 (2026-09-26). Weighted rubric replaced by cumulative levels 0 to 4 and a question map. Separate performance panel. Reconciliation note. Annual cycle.
• v0.2 (2026-09-26). Gold standard, evidence tiers, Tier 1 record and 20 question list added.
• v0.1 (2026-09-26). First draft.