> ## Content Index
> Fetch the complete content index at: https://www.thefringetheory.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# $250 million, 8% attrition, and the case study nobody has run
- URL: https://www.thefringetheory.com/250-million-and-a-case-study/
- Published: 2026-08-08T04:08:53.000Z
- Updated: 2026-08-08T04:26:54.000Z
- Description: Bank of America covers GLP-1s for 132,000 employees and calls it an investment rather than a saving. Its filings go back fourteen years, its turnover runs at a third of its sector's rate, and whether those two facts are connected is nobody outside the company can answer.
- Author: Derek Winn
- Tags: Tail

Three questions

1. Did Brian Moynihan just open the door to the best long-term case study anyone could run on large employers covering GLP-1s for weight loss?
2. If a health budget sits flat for ten years, does that mean the plan is being run well?
3. Bank of America keeps its people far longer than its industry does. Is the benefits spending part of the reason?

## What he said

On August 5, Brian Moynihan [told CNBC's Squawk Box](https://www.cnbc.com/2026/08/05/bank-of-america-ceo-glp-1-drugs-cost.html?ref=thefringetheory.com) that Bank of America spends "about $250 million or more on GLPs, and that's up from zero," meaning up from zero four or five years ago. He put total employee healthcare spend at more than $2 billion. He said the bank pairs the drugs with health coaching. He said they were negotiating hard on price. He said the medications were lowering near-term cardiac events among people taking them, including people who "don't have all the attributes." He raised, unprompted, that employees leave and the bank may never see the long-term savings. He said it was worth doing anyway. He called it a good investment.

Those are the facts of it. Everything after this is what I think they mean.

The short version, three parts. I think this is a defensible thing to spend money on. I think the way he framed it is more honest than the way most employers talk about this category. And I think the company is sitting on the most useful dataset in American benefits right now, and nobody appears to have asked for it.

## Where a plan this size files its numbers

The $250 million itself cannot be checked anywhere, and that is worth establishing before building on it.

Bank of America's health plan is self-funded, like essentially every employer its size. That means the company pays the claims itself rather than buying insurance and letting a carrier pay them.

Every benefit plan files a Form 5500 each year with the Department of Labor, and for cost work that filing does not help much here. It shows roughly $477 million of insured coverage, mostly dental, life, disability and the regional Kaiser plans. It shows nothing at all about the self-funded medical and pharmacy spending, which is most of the program. That is where the trail usually goes cold.

It does not go cold here, and the reason is how the money is held. Bank of America funds its benefits through a voluntary employees' beneficiary association, which is simply a trust set up to hold benefit money and pay it out. Theirs is called the Bank of America Group Benefits Program Trust. Trusts like this are tax-exempt, tax-exempt organizations file a Form 990 every year, and the Form 990 is public. So the trust has to report what it spent, and anybody can read it.

For 2024 that figure was $2,823,914,831\. The trust paid $86,365 in administration against that $2.8 billion, which tells you it is a pipe rather than a business. Money goes in and benefits come out, and almost nothing sticks to the sides.

Two things to know before using that number. It covers the whole benefits program, not just health: life insurance, disability, dental and legal are in there too. And it includes what employees pay in through payroll, so it is the full cost of the program rather than the company's share of it.

What it does not include is retirement money. That is worth saying plainly, because it is the first thing anyone sensible asks. A trust like this is not allowed to hold it. The rule that governs them, 26 CFR 1.501(c)(9)-3, says they cannot pay anything resembling a pension, an annuity at retirement, or a profit-sharing plan. A 401(k) is a profit-sharing plan, so it is out.

The filings back that up. Bank of America's 401(k) and its pension are separate plans with their own trusts and their own Form 5500s, holding roughly $62.9 billion and $19.0 billion at the end of 2024\. None of that money is in the benefits trust. The trust's own description of itself on the 990 says what is in it: welfare benefit plans covering life, health, and disability.

Of that $477 million of insured coverage, roughly $355 million is not health: the life, dental, disability and legal lines. Take that out of the trust total and about $2.47 billion of health-side cost is left, which sits comfortably around the "more than $2 billion" Moynihan described. That subtraction carries an assumption worth naming. It works only if those premiums were paid out of the trust rather than directly by the company, and the filings do not say which. Read the other way, the health-side figure is the full $2.82 billion. Spread across the roughly 132,000 employees enrolled in a national medical plan, or a somewhat larger number once you fold in the regional HMO enrollees the bank counts separately, the readings put gross cost somewhere between about $17,000 and $21,400 per enrolled employee per year. The low end is the likeliest, since premiums for a plan are ordinarily paid as an expense of that plan. Mercer put the national average at $16,501 for 2024\. Every version of this calculation lands above that, which is where a large bank with a generous plan should land, and that is the best evidence I have that this is reading the filings rather than inventing a number. A separate retiree trust spent another $228 million. What none of it contains is a drug line, so the $250 million stays where he left it.

## Fourteen years, filed

### The filed record: what the benefits trust actually spent, 2011 to 2024

Total annual expenses of the Bank of America Group Benefits Program Trust, the VEBA that pays the welfare program, from its public Form 990 filings. Every bar is a filed figure. Nothing here is modeled.

Trust expenses, filed (Form 990) Year-over-year change, filed 

**Every figure is filed.** Bars are total annual expenses of the Bank of America Group Benefits Program Trust, EIN 95-7055654, a 501(c)(9) VEBA ("comprehensive welfare benefit plans for employees and their families covering life, health, and disability"), from its Forms 990 via IRS/ProPublica; participant counts from the plan's Form 5500 filings (EIN 56-0906609, plan 501, DOL EFAST2). Read the scope carefully: this is the whole welfare program, health plus life, disability, dental and legal, and it includes employee contributions, so it is gross program cost, not the company's share alone. The bridge to the CEO's August 2026 "more than $2 billion on healthcare": the 2024 filed total is $2.82B; the latest filed Schedule A carries roughly $355M of non-health insured premiums (life, dental, vision, disability, legal), leaving about $2.47B of gross health-side cost, and a company share above $2B is consistent with employee contributions covering the difference. A separate retiree VEBA (EIN 30-6599079) spent a further $228M in 2024 and is not in these bars. Anything paid outside the trusts would not appear here; the scale against the CEO's figure suggests substantially all of the program flows through. **Read the gross line with the denominator in hand.** Participants fell 33% over this span, from 257,525 (2012) to 172,475 (2024), so gross dollars grew just 0.2% a year while cost per participant grew 4.0% a year. Flat here means shrinking, not controlled. Recent years converge as participation stabilized: 2019 to 2024, gross +3.6% against +4.1% per participant, or $13,412 to $16,372 each. Independent and unaffiliated; not advice.

Data table 

| Year | Trust expenses | YoY    | Plan participants | Per participant |
| ---- | -------------- | ------ | ----------------- | --------------- |
| 2011 | $2,736,880,074 | –      | –                 | –               |
| 2012 | $2,646,670,886 | −3.3%  | 257,525           | $10,277         |
| 2013 | $2,470,926,341 | −6.6%  | 242,895           | $10,173         |
| 2014 | $2,442,262,897 | −1.2%  | –                 | –               |
| 2015 | $2,478,142,522 | +1.5%  | –                 | –               |
| 2016 | $2,317,825,749 | −6.5%  | 190,415           | $12,173         |
| 2017 | $2,301,254,276 | −0.7%  | 183,749           | $12,524         |
| 2018 | $2,385,339,593 | +3.7%  | 182,896           | $13,042         |
| 2019 | $2,370,867,732 | −0.6%  | 176,772           | $13,412         |
| 2020 | $2,273,303,967 | −4.1%  | 178,711           | $12,721         |
| 2021 | $2,546,883,219 | +12.0% | 179,672           | $14,175         |
| 2022 | $2,546,210,273 | 0.0%   | 172,451           | $14,765         |
| 2023 | $2,721,116,134 | +6.9%  | 177,189           | $15,358         |
| 2024 | $2,823,914,831 | +3.8%  | 172,475           | $16,372         |

Read on its own, that chart says a plan that solved something. Total expense was $2.74 billion in 2011 and $2.82 billion in 2024\. In between it fell for most of a decade, bottomed at $2.27 billion in 2020, and has climbed since. Across the whole span it grew two tenths of one percent a year, against employer trend rates running five and six percent through the same period. With this chart and nothing else, you would reasonably conclude the program was well run.

## The denominator

The plan's own Form 5500 reports how many participants it covers, on the same filings, for the same years. In 2012 the program covered 257,525\. In 2024 it covered 172,475\. The plan lost a third of the people in it.

Put the filed dollars over the filed participants and the picture turns over. Cost per participant went from $10,277 in 2012 to $16,373 in 2024, a rise of 59%, compounding at 4.0% a year. Over the same twelve years, off the identical filings, gross dollars grew 0.5% a year.

### Normalized: the same filed dollars, per plan participant

Trust expenses divided by the participant count on the plan's own Form 5500\. The gross line looked flat for a decade because the plan lost a third of its participants. Per person, it never was.

Cost per participant, filed Projection range Participants (the denominator) 

**Normalizing reverses the story, and it matters in both directions.** Numerator: total expenses of the Bank of America Group Benefits Program Trust (Form 990, EIN 95-7055654). Denominator: total participants reported on the plan's Form 5500 (EIN 56-0906609, plan 501, DOL EFAST2), which counts participants covered by any benefit in the program, not covered lives. On the filed record, 2012 to 2024, gross dollars grew **0.5% a year** while cost per participant grew **4.0% a year**, because participants fell 33%, from 257,525 to 172,475\. So the earlier "flat for a decade" reading was an enrollment artifact, and the gross line understates the underlying trend by about three and a half points a year. Over the recent window the two nearly converge, since participation stabilized: 2019 to 2024, gross +3.6% against +4.1% per participant. Forward, the projection compounds the per-person figure at 6.5% (upper) and 4% with a roughly four-point 2027 step-down for a GLP-1 net-price reset (lower), and holds participants flat at the 2024 count, which is the assumption to argue with: BofA headcount has been flat near 213,000 for three years, but plan participation still drifted about half a point a year over 2019 to 2024\. Carry that drift and the gross equivalents fall roughly $95 million below the flat-participation case by 2029, so a gross projection at these rates is mildly steep while a per-person one is not. Vertical axes are truncated for resolution. Independent and unaffiliated; not advice.

Data table 

| Year | Trust expenses | Participants | Per participant    | Per-participant YoY |
| ---- | -------------- | ------------ | ------------------ | ------------------- |
| 2012 | $2.647B        | 257,525      | $10,277            | –                   |
| 2013 | $2.471B        | 242,895      | $10,173            | −1.0%               |
| 2014 | $2.442B        | 219,321      | $11,136            | +9.5%               |
| 2015 | $2.478B        | 202,030      | $12,266            | +10.2%              |
| 2016 | $2.318B        | 190,415      | $12,172            | −0.8%               |
| 2017 | $2.301B        | 183,749      | $12,524            | +2.9%               |
| 2018 | $2.385B        | 182,896      | $13,042            | +4.1%               |
| 2019 | $2.371B        | 176,772      | $13,412            | +2.8%               |
| 2020 | $2.273B        | 178,711      | $12,721            | −5.2%               |
| 2021 | $2.547B        | 179,672      | $14,175            | +11.4%              |
| 2022 | $2.546B        | 172,451      | $14,765            | +4.2%               |
| 2023 | $2.721B        | 177,189      | $15,357            | +4.0%               |
| 2024 | $2.824B        | 172,475      | $16,373            | +6.6%               |
| 2025 | projected      | held flat    | $17,028 to $17,437 | +4.0% to +6.5%      |
| 2026 | projected      | held flat    | $17,706 to $18,570 | +4.0% to +6.5%      |
| 2027 | projected      | held flat    | $17,677 to $19,777 | −0.2% to +6.5%      |
| 2028 | projected      | held flat    | $18,385 to $21,062 | +4.0% to +6.5%      |
| 2029 | projected      | held flat    | $19,122 to $22,431 | +4.0% to +6.5%      |

Worth saying plainly, because 4% a year sounds worse than it is: that is below the market. Mercer's national average rose 4.81% a year over the same five years against Bank of America's 4.07%. The levels are not directly comparable, since Mercer counts per covered employee and the Form 5500 counts participants, but on growth the plan ran under the benchmark it is usually measured against.

Neither number is wrong, and neither one is the honest number by itself. They answer different questions. The total says what the trust paid, which is what a CFO funds and what a board sees. The per-person figure says what covering one person costs, which is what a benefits decision actually moves. For twelve years the total was flattered by a workforce that kept getting smaller. Nobody had to make a mistake for the two to drift this far apart.

## What changed around 2020

Most of the participant decline happened between 2012 and 2019, when the program shed roughly five percent a year. After 2019 it flattens: 176,772 in 2019 against 172,475 in 2024, about half a percent a year, and Bank of America has now reported approximately 213,000 employees for three consecutive years. The shrinking stopped.

The gross line responded almost immediately. Filed expense rose 12.0% in 2021, was flat in 2022, then rose 6.9% and 3.8%. From the 2020 low through 2024 the trust's spend rose $551 million. I would not hang that turn on GLP-1s, and the filings do not support anybody who does. 2020 was a pandemic year full of deferred care and 2021 is mostly that care arriving. GLP-1 spending in 2021 was near zero on Moynihan's own account.

What I keep turning over is how this looks from inside. Spend ten years watching a budget line that does not move and you would reasonably believe the program is well managed, because on the measure most people look at, it is. Then headcount steadies, the same underlying trend keeps compounding the way it always did, and the total climbs for the first time in a decade. From the checkbook it reads as something breaking. Nothing broke. The cushion ran out, and the cushion was a labor-market fact rather than a benefits achievement.

## An investment, not a return calculation

Here is where I think a lot of the commentary went sideways, including an earlier draft of this piece.

Moynihan did not say GLP-1s pay for themselves. He was asked about a cost and answered with an investment, and those are different propositions. An investment says you are buying something of value and the value may not come back to you as reduced claims. Nobody runs a return calculation on the 401(k) match or on parental leave. I argued in [Your mileage may vary](https://www.thefringetheory.com/your-mileage-may-vary/) that GLP-1s get ROI-tested mainly because the cost happens to arrive with a member ID attached, while the other things an employer buys for its people do not. He is the first executive I have seen put that distinction on the record, and he did it in the least convenient setting for it, live, with the number in his mouth.

It matters who said it, too. A benefits director defending a nine-figure line has to clear an internal return screen. A CEO calling it a great-place-to-work investment is making a compensation decision, and that is a decision he is entitled to make without one. Good for him for saying so plainly instead of commissioning a savings estimate nobody believes.

His clinical remark holds up better than the offhand delivery suggested. SELECT, published in the New England Journal of Medicine in 2023, randomized 17,604 adults aged 45 and over, with a BMI of 27 or above and established cardiovascular disease, none of them diabetic, to semaglutide 2.4 mg or placebo. Over an average of 39.8 months of follow-up, the combined count of cardiovascular deaths, non-fatal heart attacks and non-fatal strokes came to 6.5% of the treated group against 8.0% of the placebo group. That is a hazard ratio of 0.80, or about a fifth fewer events. That is what "even if they don't have all the attributes" was pointing at: benefit in people without diabetes. The trial required established cardiovascular disease and a minimum age of 45, so it does not describe a general employed population, and I am not the person to say how far it travels. Worth knowing alongside the benefit: 17% of the treated group stopped because of side effects, against 8% on placebo. But it is a real result, and it appeared inside about three and a half years, which is inside the tenure window everybody says is too short.

## What the calculator prices, and what it cannot

Here is the coverage calculator, opened with those numbers already in it. Every slider still moves.

Tool · loaded scenario

## Bank of America's GLP-1 line, run through the calculator

The standard coverage calculator, opened with Bank of America's own filed and disclosed numbers already in it. Every slider still moves. Drag any of them and this becomes your plan instead.

What is loaded, and where each figure came from

- **132,000 covered employees.** Bank of America's Human Capital Management Update, March 2024: more than 132,000 US teammates enrolled in a national medical plan. The plan's Form 5500 reports 172,475 participants across the whole welfare program, which is the wider number.
- **$2.0 billion in plan spend.** Brian Moynihan, CNBC, August 5, 2026\. The benefits trust's Form 990 puts the whole welfare program at $2.82 billion for 2024, of which roughly $2.47 billion is health-side once non-health insured premium comes out.
- **8% turnover.** Stated in Bank of America's 10-K for 2023, 2024 and 2025\. Its sector separates workers at roughly 23% a year, so this is unusually low, and it is the input most favourable to the coverage case.
- **$7,400 net per treated member.** Not a Bank of America figure. Published net price estimates for 2025 and 2026 run about $570 to $725 per member per month; this sits in the middle. Nobody asked whether the disclosed $250 million was gross or net, so switch the rebate toggle to see the other case.
- **40% eligible, 18% starting.** Eligibility is the CDC figure for BMI 30 or above. Uptake is a modelling choice, not a disclosure, chosen so the result sits in a plausible band. It is the softest number here.
- **4% trend, five years.** Bank of America's own filed cost per participant, 2019 to 2024, rather than a market benchmark. Their 2020 to 2024 figure runs 6.5%, but that measures from the pandemic trough and is mostly the rebound.

**One scope difference that matters.** This calculator models weight management only. Moynihan said "GLPs," which reads as the whole class including the type 2 diabetes indications. So the therapy cost below should land well under his $250 million, and the gap between them is roughly the diabetes half of the class. Read this as the weight-management slice, not as a reconstruction of his number.

Reset to the loaded scenario

The eligible population

Covered employees132,000

Employees enrolled on the plan, not total covered lives

Adults covered per employee1.40

Set this to 1 plus the share of enrolled employees who cover a spouse or adult partner. Cover a spouse on 40% of contracts and the multiplier is 1.40\. Children are excluded, because the weight management labels are adult indications with narrow adolescent exceptions. Your enrollment census answers this exactly, and it is the input most worth getting right

Clinically eligible40%

CDC puts BMI 30 or above at 40.3% of US adults. The label also allows 27 or above with a comorbidity, which widens the pool; plans that tighten to 35 or 40 narrow it, with BMI 40 or above at 9.7%

Of those eligible, share who start18%

Eligibility is not utilisation. Employers who added coverage in 2023 and 2024 consistently report take-up above their projections

Does the plan capture rebates?Yes

Yes, net of rebates No, list price 

Fully insured and some smaller self-funded plans never see the rebate. If yours does not, the plan is exposed to list

Net cost per member, per year$7,400

After rebates. Employer guidance commonly assumes $600 to $900 per member per month

Who is still here, and still on it

Annual turnover8%

All separations, voluntary and involuntary

Still on therapy after a year66%

Commercial claims studies put one-year persistence between roughly a third and two thirds, improving as shortages eased

The plan, and what you expect back

Total annual health plan spend$2.00B

Expected reduction in plan spend2%

Reduction in **total** plan spend, not in the treated members' own costs

Medical trend4%

Annual growth in plan spend. Mercer projected 6.7% for 2026 after planned cost-reduction measures and near 9% without them; Aon projected 9.5%. Savings track this rate, which holds the reduction constant as a percentage

Therapy cost trend0%

Net cost **per treated member**, not total GLP-1 spend. Pharmacy trend of 13% to 15% is largely utilisation, which this tool already models separately, so per-member net cost is held flat by default. Set it negative if you expect deflationary pressure on unit cost through rebates or contracting, positive if you expect the opposite. Equal trends on both sides very nearly cancel, so it is the gap between these two that moves the answer

Years to model5

Adults on therapy at any one time19

That is this share of covered adults4.8%

Therapy cost, first year$0

That is this share of plan spend0%

Therapy cost over 5 years$0

Expected savings if nobody left or stopped$0

Expected savings you actually keep$0

On therapy a full year, from year two onward56%

Reduction needed to break even0%

Reduction needed if nobody left or stopped0%

Net position over 5 years$0

Three ways to close the gap

DOOR 01 · START HERE

Buy it cheaper

$0

DOOR 02

Get more from it

0%

DOOR 03

Cover fewer people

0%

Cumulative therapy cost Savings you keep Savings if nobody left or stopped 

Cumulative therapy cost against expected savings Lines over the modelled horizon comparing what the therapy costs against the savings the employer retains after turnover and discontinuation. 

Year one returns nothing, because a benefit that depends on sustained therapy takes a year to arrive. From year two the plan carries a steady mix in which only the members who have held on for a full year can have produced anything.

**The short version of the notes.** This models a programme that is launching rather than one already running. Treated headcount is held constant, and no benefit is credited until a member has been on therapy a full year, so year one returns nothing. It prices one question only: whether the therapy pays for itself out of plan spend. 
- **What it does not account for.** It models no clinical outcome. The expected reduction is your assumption, not a published result, and the tool takes no view on whether it is achievable. It also ignores stop-loss reimbursement, rebate timing, specialty carve-outs, and any cost sharing borne by the member rather than the plan.
- **Nothing you type leaves your browser.** No account, no submission, nothing stored or transmitted.
- **This is not an actuarial opinion and must not be used as one.** Anything you intend to act on should go to your actuary, carrier, or consultant first.
- The full method, the source behind every default, and the complete limitations are on the [GLP-1 coverage calculator](https://www.thefringetheory.com/glp-1-coverage-calculator/) page.

Fringe Theory is independent and unaffiliated. Views expressed are my own and do not represent those of my employer. Nothing here is legal, tax, medical, or investment advice.

At those settings the weight-management side alone puts about 13,300 adults on therapy. That costs roughly $98 million in the first year, just under 5% of the plan. Over five years the drugs cost $492 million, and the savings the plan actually keeps, after people leave the company or stop taking them, come to $107 million.

To break even, the drugs would have to cut *total* plan spending by 9.2%. Put another way, they would have to save about 1.9 times what they cost.

The figure behind door one is the one worth sitting with. For price alone to close it, the therapy would have to be sourced at about $134 per member per month. Medicare's negotiated price for semaglutide, effective 2027, is $274 for 30 days, and the manufacturer's cash channel runs about $350\. So the arithmetic is asking for roughly half of what the federal government got using the strongest bargaining position anyone in this market has.

The corollary

There is an awkward corollary buried in this arithmetic. Because the savings are counted as a share of plan spend, a plan growing slowly throws off fewer dollars to recover than one growing fast. Run this at Bank of America's own filed trend of about 4% rather than the market's 6.5% and the price the therapy would have to hit falls from $143 a month to $134\. The better a plan is run, the weaker its cost-offset case becomes. That is worth sitting with before anybody uses this kind of arithmetic as a coverage test.

Which is the useful thing this tool does, and also its honest limit. It prices exactly one question: does the therapy pay for itself out of plan spend. On these inputs the answer is no, it is not close, and no combination of price, retention or tightened eligibility gets there. That is worth knowing. But it is not the question Moynihan answered, and the calculator cannot price a cardiac event that does not happen to a 58-year-old operations manager, or being the employer that covered this when the company down the road did not. It measures the cost of a decision made on other grounds.

## If I were sitting in that seat

None of what follows is a criticism of how the bank is running this. It is what I would want on my own desk if I owned a nine-figure drug line, and four of the five apply at a tenth the size.

**Run a market check on net cost, not a renewal.** At $250 million, one point of net price is $2.5 million a year, which is real money for a procurement exercise that costs a few hundred thousand to run properly. A full request for proposal, with the definitions locked before pricing is compared rather than after. And I would price the alternatives next to the current arrangement rather than assuming the current arrangement is the market. Full pass-through, meaning every rebate dollar comes back to the plan against a definition that is written down and auditable. Cost-plus, meaning you pay what the drug actually costs plus a stated fee. And carving this drug class out on its own terms rather than leaving it inside a contract priced as a whole. One more worth pricing, and it is slightly uncomfortable. The manufacturers now sell direct to consumers at somewhere between roughly $350 and $500 a month. In some plan designs that is at or below what the plan is already paying, after rebates, for the same drug. Any of these needs a compliance and plan-document read before it goes anywhere near members, and some will not survive it. Price them anyway, so the decision is informed rather than inherited.

**Read the pharmacy agreement the way somebody else's lawyer would.** The definitions are the contract. What counts as a rebate, what counts as specialty, what counts as a generic, and who decides when a drug moves between those buckets, will move more money than the discount headline. Then the audit rights: scope, frequency, who pays, what happens when the audit finds something, and whether the clause survives a formulary change. Since 2024 employees have sued their own employers over prescription drug pricing under ERISA, which is what moved this from a procurement question to a governance one. Two are worth knowing about. Lewandowski v. Johnson and Johnson was filed in New Jersey federal court in February 2024, and Navarro v. Wells Fargo followed in Minnesota. Both alleged that the employer failed to make sure the prices its plan paid were reasonable. Read the outcomes before drawing conclusions: the courts have set a high bar on standing, the J&J claims were dismissed in November 2025 and are now on appeal, and Wells Fargo was dismissed as well. So this is not a settled area of law and nobody has been found liable. I am reporting that the cases exist and what they allege, not that any of it applies here or that it would succeed. Whether a given arrangement raises a prohibited transaction question under ERISA is a matter for benefits counsel, and it is exactly the sort of thing worth asking counsel before somebody else asks it for you.

**Re-run the model against 2027, not against last year.** Three price events land inside about a year of each other, and the projection everybody is working from predates all of them. I would model the rebate assumption in both directions, since even the gross-or-net question on the public figure is unresolved. And I would run it per participant rather than gross, for the reason the second chart shows. On a gross basis a shrinking population can hide a rising cost for a decade. A growing one can manufacture a crisis that is not there.

**Ask what everybody in the chain is paid, in writing.** The plan's Schedule A names a broker of record and reports commissions, and on most lines of a program this size those commissions read as nothing, which usually means the compensation is arranged somewhere other than the commission field. That is ordinary and it is not a scandal. It is also not an answer. A 2021 federal law, the Consolidated Appropriations Act, gave plan fiduciaries the right to be told what their service providers are paid, directly and indirectly. The useful version of that right is a written statement from every party in the chain. Everything they receive from any source connected to this plan, including from carriers and the pharmacy benefit manager. And whether any part of it moves up when drug spending does. Most sponsors have the disclosure in a file somewhere and have never read it against the contract.

**Decide now what would change your mind.** Write down, before the evidence arrives, what result would make you narrow this and what result would make you widen it. Cost per treated member, persistence at two and three years, cardiac events, disability days, turnover among the treated group against everybody else. Sponsors almost never do this, which is why coverage decisions in this category tend to get made twice, once on enthusiasm and once on a renewal shock, and neither time on evidence.

## Eight percent

Which brings me to the number I actually cannot stop thinking about.

### Eight percent: what Bank of America reports losing each year

Annual turnover as stated in the company's own 10-K, against separations across finance and insurance. Both lines are public. Neither one explains the other.

Bank of America, stated in the 10-K Finance and insurance, all separations 

**Bank of America's figures are stated, not derived.** Each appears verbatim in the Human Capital section of a Form 10-K, for example "Our turnover among employees was stable at 8 percent in both 2025 and 2024." They cover total turnover, voluntary and involuntary together, company-wide. The sector line is the Bureau of Labor Statistics JOLTS annual average separations rate for finance and insurance, multiplied by twelve to put a monthly rate on the same footing as an annual one, which is our arithmetic and not the BLS's. **The two measures are cousins, not twins.** JOLTS counts every employer in the industry including small firms, where turnover runs structurally higher, so some of this gap is size rather than retention. The gap is wide enough to survive that caveat: BLS puts total separations across all industries near 40% a year over the same period. For a third reference point, a survey of financial institutions put front-line turnover at 23.4% in 2022 and 19.8% in 2023\. Independent and unaffiliated; not advice.

Data table 

| Year | Bank of America (stated) | Finance & insurance (derived) |
| ---- | ------------------------ | ----------------------------- |
| 2018 | 12%                      | –                             |
| 2019 | 11%                      | –                             |
| 2020 | 7%                       | –                             |
| 2021 | 12%                      | \~26%                         |
| 2022 | 13%                      | \~28%                         |
| 2023 | 8%                       | \~23%                         |
| 2024 | 8%                       | \~23%                         |
| 2025 | 8%                       | \~25%                         |

Bank of America reports 8% annual turnover. It has reported 8% for three years running, and it puts the figure in the Human Capital section of every 10-K, which plenty of companies do not bother to do. Its sector separates workers at something north of 20% a year on the government's figures, and the whole economy runs near 40%. Even allowing that the measures are cousins rather than twins, and they are, that is a genuinely impressive place to be, and it is checkable by anybody with twenty minutes and EDGAR.

Now the tempting part. GLP-1 coverage started around 2021 or 2022\. Turnover ran 12% and 13% in those years and has been 8% every year since. It would be easy to draw a line between those two facts and call it a retention story.

I am not going to, because the sector did roughly the same thing over the same window. Finance and insurance separations peaked in 2022 and fell in 2023, which is when the labor market cooled for everyone. The timing fits the macro at least as well as it fits the benefit, and one company's series cannot separate them.

But look at what is true. Something is working at this employer. The benefits investment is real, sustained, and unusually well documented. Premiums held flat for the lowest-paid teammates for more than a decade, a $25 minimum wage as of October 2025, and now a nine-figure drug line the CEO will defend on camera. The retention outcome is excellent. The two sit side by side in the public record and nobody outside the company can tell you whether one causes the other. That is not a criticism of the bank. It is a description of the limits of public data, and it is the whole reason the next section exists.

## The case study nobody has run

The first question I would ask, if anyone at the bank is reading: have you already done this internally? A company four or five years into covering these drugs, with a stable population and a CEO willing to say "good investment" on television, has almost certainly looked at its own claims. Whether it looked at cost, at retention, at cardiac events, at all three, and what it found, is not something any filing will ever tell you.

The second question is whether it becomes a real study, because the field is openly asking for one. In May 2025 the National Institute of Diabetes and Digestive and Kidney Diseases convened a workshop, co-sponsored by other NIH institutes along with the CDC and FDA, on leveraging real-world evidence to assess the benefits and risks of GLP-1 therapies. Its stated purpose was identifying gaps in what we know and how real-world evidence could close them, and among the questions it named was the effect of these drugs on healthcare costs. Two synopses have since run in the Annals of Internal Medicine. The data sources it points to are electronic health records and claims.

Employer plans hold exactly that claims data, and almost none of it is ever studied. What gets published tends to be vendor analyses of pooled books, useful but not independent, and none of it follows a defined population long enough to answer the question everyone is arguing about.

Bank of America is an unusually good candidate, and the reasons are all in the filings. Roughly 132,000 employees enrolled in a national medical plan, plus dependents. A single pharmacy administrator across every carrier from 2026, so the drug data is not fragmented. Four or five years of coverage already behind them. And that 8% turnover, which sounds like a footnote and is actually the entire thing: at 8% you can follow a treated cohort for five years and still have two thirds of it. Almost no American employer can offer a researcher that. The low attrition that makes this a good place to work is the same fact that makes it a good place to study.

Nobody outside can do this alone. The claims are theirs, the privacy obligations are real, and it needs an academic partner and an IRB rather than a benefits consultant with a spreadsheet. But the New England Journal published SELECT, and JAMA, Health Affairs and Annals publish work like this constantly. What none of them has is a large, stable, single-PBM employer population followed for five years with a sponsor willing to be named. Bank of America could be the first, and it would be the most useful thing anybody has done in this category since the trials.

In the meantime the filed record runs about eighteen months behind. The trust's FY2025 return lands around the end of this year, and the year Moynihan was describing will not be filed until 2027.

Run your own numbers: the [GLP-1 coverage calculator](https://www.thefringetheory.com/glp-1-coverage-calculator/).

### Sources

1. CNBC, Squawk Box interview with Brian Moynihan and published excerpts, August 5, 2026.
2. Fortune, August 7, 2026; Business Insider, August 5, 2026; Quartz, August 5, 2026.
3. Bank of America Group Benefits Program Trust, Forms 990, EIN 95-7055654, fiscal years 2011 through 2024, via the IRS and ProPublica Nonprofit Explorer.
4. Bank of America Retiree Group Benefits Program, Form 990, EIN 30-6599079, fiscal year 2024.
5. Bank of America Group Benefits Program, Forms 5500, EIN 56-0906609, plan 501, plan years 2012 through 2024 (participant counts; Schedule A insured premiums); and Forms 5500 for plan 001, The Bank of America Pension Plan, and plan 003, The Bank of America 401(k) Plan, plan year 2024 (trust assets). All via U.S. Department of Labor EFAST2.
6. Bank of America Corporation, Forms 10-K, fiscal years 2018 through 2025, SEC EDGAR (headcount and turnover, Human Capital sections).
7. 26 CFR 1.501(c)(9)-3, on the benefits a voluntary employees' beneficiary association may provide, via the electronic Code of Federal Regulations.
8. Lewandowski v. Johnson and Johnson, No. 3:24-cv-00671 (D.N.J., filed February 5, 2024; claims dismissed November 26, 2025; notice of appeal January 16, 2026); Navarro et al. v. Wells Fargo & Company (D. Minn.), dismissed on standing grounds. Procedural history via the Georgetown Health Care Litigation Tracker.
9. U.S. Bureau of Labor Statistics, JOLTS annual tables 20 and 22 (separations and quits by industry); "Employee Tenure in 2024," September 26, 2024\. Crowe LLP financial institutions compensation survey, 2022 and 2023 editions.
10. Bank of America, Human Capital Management Update, March 2024 (medical plan enrollment); newsroom materials on the $25 minimum wage, October 2025.
11. Bank of America, 2026 U.S. Benefits Enrollment guide and prior-year guides (carriers and pharmacy administrator).
12. Lincoff AM et al., "Semaglutide and Cardiovascular Outcomes in Obesity without Diabetes" (SELECT), New England Journal of Medicine, 2023.
13. National Institute of Diabetes and Digestive and Kidney Diseases, "Leveraging Real-World Evidence to Assess Benefits and Risks of GLP-1-Based Therapies," workshop held May 7 and 8, 2025; workshop synopses, Annals of Internal Medicine, 2025 and 2026.
14. Peterson Health Technology Institute, "Employer Approaches to GLP-1 Coverage," December 2025; EBRI Issue Brief No. 644, October 9, 2025; Willis Towers Watson, March 2026 (net price estimates).
15. Mercer, National Survey of Employer-Sponsored Health Plans: 2024 results published November 2024 ($16,501 average total health benefit cost per employee) and 2025 results published November 2025 (trend rates and the 2026 projection).
16. CMS, negotiated maximum fair price for semaglutide effective 2027, announced November 2025; Novo Nordisk, U.S. list price reduction announcement, February 24, 2026.
17. CDC, National Center for Health Statistics, NHANES August 2021 to August 2023 (obesity prevalence).

Fringe Theory is independent and unaffiliated. Views expressed are my own and do not represent those of my employer. Nothing here is legal, tax, medical, or investment advice. Figures described as modeled or derived are my own arithmetic from the public sources listed, not disclosed data.