Half-caf, no whip: Starbucks drops GLP-1 coverage for weight loss

I was on vacation last week, so you can imagine how pleased I was to see another headline when I got back home: Starbucks is dropping GLP-1 coverage.

Why this matters, to me at least. This is more than I could have asked for. First I built a tool to let an employer price GLP-1 coverage against its own census and find the break-even. A week later a CEO went on television and became the worked example for keeping it. Now an employer has decided to drop it.

Two employers, opposite decisions, the same drug, the same month, both on the public record. A calculator is only as good as the cases you can test it against, and these are the two poles.

I dove in, but not to critique. This is an autopsy of a decision, and it will bring a perspective none of the other headlines will offer.

The research

Starbucks funds its plans differently than Bank of America does, and that is the punchline of the whole piece.

Bank of America pays claims. Starbucks buys insurance.

Before going further, let me be clear that I cannot testify to the state of mind of the witness. I have no relationship with Starbucks. They may have well-rooted reasons for procuring benefits the way they do. Both companies file enough for us to say how different they are rather than assume it.

Two honest notes on that table. The pay medians are both global figures and they are not measured the same way: Bank of America's includes employer-paid health benefits and Starbucks' explicitly excludes them, so the disclosed gap understates the cash difference and overstates the total-rewards difference at the same time. The two turnover figures are not measured alike either, which I will come back to.

Because Starbucks uses insured products for its health plan, we get access to granular data through its Form 5500 filings.

The two stories got filed under the same heading. They are not the same story, and the difference is not scale or philosophy. It is procurement, and perhaps an awakening.

Bank of America's Brian Moynihan was direct about it on CNBC on August 5. "We spend about $250 million or more on GLPs," he said, "and that's up from zero." Asked about the cost, he answered with a return that was not financial: what the company sees is an impact on employees, including lower near-term incidence of heart problems among people taking the drugs, "even if they don't have all the attributes." That is the language of a business talking about a long-horizon investment, not a line item.

Every dollar of that $250 million is a claim Bank of America pays out of its own money, so the decision to keep paying is a decision about its own claims. Starbucks does not, on the evidence of its own filings, run its medical coverage that way. Every medical contract in its welfare plan filing is an insurance contract: thirty-five contracts in all for the year ended September 30, 2025, twenty-two of them medical across twelve carriers, carrying $774.9 million of premium in total and $726.5 million on the medical lines. The plan reports no trust. Twenty of those twenty-two medical contracts are reported as not experience-rated.

As we discussed in the GLP-1 calculator, turnover is one of the factors that decides whether this coverage can pay for itself. The chart below is why that matters here.

Bank of America discloses turnover in its 10-K every year, and has reported 8 percent in each of the last two years. Starbucks' 10-K says only that the company has "low turnover," with no number; the 45 percent figure appears in one unfootnoted row of the Fiscal 2025 Global Impact Report data tables, outside that report's assurance scope. Run each rate forward and the horizons separate fast. Of 100 people who start therapy today, roughly 66 are still on the Bank of America plan at year five. At Starbucks' rate, roughly five. Add the eligibility gate underneath it, an average of 20 hours a week tested against 520 hours in each six-month measurement period, and a therapy started in month three has to clear that test twice before month twelve.

That is not an argument that Starbucks should or shouldn't cover the drug. It is an argument that they would either need to a) buy it for less, b) get more from it, or c) treat it as a perk or an investment in their workforce. It is the reason the same spreadsheet returns a different answer in the two buildings.

Where a plan like this files its numbers, and where it doesn't

An important nuance to discuss before going further, because it drives everything after it. The Form 5500 settles the funding question in one line. Plan 506, the Starbucks Corporation Welfare Benefits Plan, reports its funding arrangement as insurance and general assets of the sponsor. Trust is not checked. The 10-K is consistent with that, showing an insurance reserve of $282.3 million inside accrued liabilities at the end of FY2025 rather than assets held apart, though the filing does not say what risks that reserve covers and Starbucks self-insures other exposures.

That single line removes almost everything a reader would want. An unfunded welfare plan takes a limited exemption, so no Schedule H, no financial statement, no total. There are no plan assets, so there is no Schedule C, the schedule that reports what service providers were paid. The public record contains no claims figure, no spend figure, and no vendor compensation schedule.

Note what that exemption also means: because nothing financial has to be filed, the absence of a self-funded line is not proof there isn't one. What the filing establishes is that there is no trust, and that every medical contract it does report is insured. That is enough to work with, and it is where the stopping point was on this journey.

It is worth being precise about one more thing the filing cannot tell you. A Schedule A is filed only for insurance contracts, and the instructions say plainly not to file one for an administrative-services-only arrangement. So those thirty-five contracts are thirty-five genuine insurance policies rather than an administrative shell. But a carved-out clinical vendor is a service arrangement, not insurance, so it would generate no Schedule A at all, and with no plan assets there is no Schedule C to catch it either. The plan's own filing proves the point: Starbucks has run a named mental health vendor since April 2020 and a second wellbeing app alongside it, and neither appears anywhere in these filings. This document does not detect point solutions.

Which raises a fair question I cannot answer: how were they covering it in the first place? No public source names a pharmacy benefit manager, a weight-management vendor, or a manufacturer channel. So I will say only what I know. Something was covered. It stops in October. The filings do not show me the channel, and anyone telling you they have verified it from public records has not.

What the filing contains instead is thirty-five Schedule A filings, one per insurance contract, each carrying the premium and the producer compensation. That turns out to be the better document.

Six years of filed premium

Gross medical premium grew 36 percent across that window. That number needs no denominator and nobody can argue with it.

The denominator is where it gets interesting, and where it gets hard. Starbucks' US head count peaked at 258,000 in fiscal 2022 and fell to 211,000 by fiscal 2024, an 18 percent contraction. Put the premium over the people and the burden per US employee ran down 10 percent across the first three years, while the company was hiring, then rose 18.4 percent and 24.3 percent in the two years that followed. Up 41 percent across the whole window, about 7 percent a year.

Two things to say plainly about that, and the first one matters more than anything else on this page. That is per person on the US payroll, not per person enrolled. Most Starbucks partners are part-time, coverage starts at an average of 20 hours a week, and a large share of that payroll never elects anything. Implied enrollment runs somewhere between 40 and 50 percent, which means the real cost per enrolled employee is roughly double the figure in that column. So do not set $3,443 against a published per-employee benchmark and conclude Starbucks buys cheap coverage. Those benchmarks are quoted per enrolled employee and this column is not that. Nobody publishes the enrolled count, so this is a workforce-cost lens, and its only real virtue is that Starbucks states the denominator plainly every year in the same words.

The second is that I got this wrong on the first pass, and the way I got it wrong is worth more than the correction. The plan's own Schedule A head count says cost per covered person rose 73 percent. It didn't. One option's reported count halved in a single year while its premium rose, which is a change in what is being counted rather than what is being paid, and that one change is larger than the whole book's apparent drop in covered people that year. Hold it on its prior basis and the answer is about 7.2 percent a year. Divide by the plan's own participant count from the Form 5500 instead and it is 7.15. Per US employee, 7.09. Four denominators, one answer, and the outlier was the one that looked most dramatic.

This is an imperfect science. Each plan year is set against the head count reported at the fiscal year end that falls within days of the plan year's start, so it is the count at the opening of the window rather than the close. Run the whole series off the closing count instead and the five-year total lands in the same place, about 39 percent against 41, but the single largest jump moves from the last year to 2022. The 2019 base year also runs through the first pandemic autumn, and none of these denominators is the one you would actually want. The direction is reliable. The decimal places are not.

That 2024 filing was received by the Department of Labor on April 23, 2026. Business Insider reported the GLP-1 decision about fifteen weeks later.

It is worth a pause to review these numbers. The writing was likely on the wall long before headlines broke.

It is also worth noting that Starbucks has been through a hard stretch, and the filed benefits record sits inside it. In fiscal 2025 margins came down and diluted earnings per share halved on revenue that rose. In September 2025 it filed an 8-K disclosing a restructuring it estimated at roughly $1 billion, closed 627 stores in the fourth quarter of that year and eliminated about 900 non-retail roles, with 90 percent of the cost in North America.

It cuts the other way too, and it is only fair to say so: fiscal 2026 has been a recovery, with comparable sales up 7.9 percent globally in the most recent quarter. The contraction was real, and it happens to be exactly the window the filed benefits record covers.

Philosophies in managing cost

Here is where the two stories come apart.

Starbucks, procuring benefits the way it does, can see the removal of coverage as a saving: premium it won't have to spend, for a risk it no longer has to insure. Bank of America described the same category as an investment, with early signs it is mitigating future spend. Neither of those is a wrong answer. They are answers to different questions, and the procurement method is what decides which question gets asked.

Starbucks does not buy one health plan. Its benefits site tells partners today that "there are multiple coverage levels and insurance carriers." The company announced the design in July 2016 as "a unique private exchange model" built with Aon, offering "as many as six national and regional carriers and five coverage levels."

A decade on I cannot tell you from public records what that arrangement is called today, and it turns out not to matter, because the filing shows what it does. Twenty-two medical contracts. Twelve carriers. Blocks as small as fifty-two covered people on a regional HMO, and several more under two hundred. No employer negotiates a standalone insured contract for fifty-two employees. Contracts that size exist because somebody is aggregating employers, which is what an exchange does whatever the brand on it happens to say. Aon entities are the producer of record on nearly every contract in the filing, and half the carriers run Aon-branded member sites for the 2026 plan year.

That is how one workforce ends up split across twelve carriers and dozens of contracts, and my read is that most of those blocks are too small and too self-selected to experience-rate credibly. The architecture that gives employees choice is the same architecture that moves the employer's claims to the back seat.

That leaves Starbucks with the harder question: if you drop the coverage, does the premium come down by that amount, or does it come off a renewal you were never going to see?

There is one more way to size it. Set the plan's premium against what the North America business earned that year. For four straight years it ran between 11 and 13 percent of North America operating income. In the year ended September 2025 it was 23 percent. Strip out the restructuring charge recognized in the year, about $892 million and roughly 90 percent of it in North America, and it is still 18.3 percent. The benefit line did not double. The profit it comes out of halved.

What the record will not settle

None of this will be settled from outside. These are decisions made in a boardroom, and I am only here to lay out the data on how decisions like this may have been made.

Start with what is actually known, which is less than the coverage suggests. Starbucks did not announce this. A reporter found it and a spokesperson confirmed a previously unreported change without elaborating. There is no company statement on the record. The change is reported to apply to the weight-loss indication, and the same reporting says the drugs may still be covered for other medical uses. May. Nobody has confirmed what happens to the other half. Which also means the $250 million it is being measured against is not a like-for-like number, since that figure was never broken out by indication either. That line between the weight-loss indication and the other medical uses is where the next drug lands, and I picked it up in load bearing.

In the simplest terms I can offer: for a plan built like this one, reducing coverage reduces premium. It was a renewal decision made on a renewal calendar. This is the ecosystem where it is easiest to stare at the angry hog's snout and ignore the tail.

If it saves money, the saving is a renewal that came in lower than the one that would otherwise have been quoted. There is no filing for that. The 2026 premiums will appear on Schedule A eventually, net of enrollment changes, network changes, a store footprint that has been moving, and whatever else was negotiated in the same conversation. A 41 percent rise in premium per US employee across five years on the filed record is checkable now. The effect of the decision itself never will be, and the anatomy of it will not be known unless somebody discloses it.

Which leaves the odd position everyone is in. The employer that self-funds and keeps paying gets to state a number on television and be argued with. The employer that buys insurance and stops buying one piece of it is legible only as an absence, and an absence is not evidence of much. It may be a company that looked hard at a cost line rising about 40 percent per employee across five years and made a defensible call on a benefit its carriers price into a pooled rate it cannot audit. It may be the item that happened to move in a renewal conversation. From out here, those look identical.

One last thing the filings show, and I cannot tell you what it means. The Forms 5500 for plan years 2021, 2022 and 2023 were each re-filed on a single day in January 2026, and all three are marked as amended returns. Three consecutive years, restated at once, fifteen months before anyone was asking about this plan. What changed is not in the public record, because the superseded originals are no longer in it either. By the time that record catches up with a decision like this one, the decision is two renewals old and the people arguing about it have moved on.

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

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.