The Costco and SCAN partnership got read across the industry this week as a distribution story. Shelf space, foot traffic, a trusted brand at the front door. That’s the smallest thing in the deal.
Read what Sachin Jain actually said about how it works. The plans will connect to Costco’s pharmacy, its over-the-counter benefit, its vision business, and its audiology business. Those are four of the largest supplemental benefit categories in Medicare Advantage, and they are about to run through a retailer that has been watching these same people shop every week for twenty years.
That’s the deal. Not the storefront.
Costco has better longitudinal data on these members than the plan ever will
A health plan sees a member a handful of times a year, and only after something has already happened. A claim gets filed. An encounter gets submitted. A prescription gets picked up. The picture arrives late and full of holes, and every plan in the country has built an entire operation around working with a signal that thin.
Costco sees the same person forty or fifty times a year, for a decade before they ever age into Medicare, at the item level, with a payment method attached to a verified identity.
I’ve spent twenty years building the systems that move data between payers and providers, and I want to be precise about why this is different. It isn’t more data. Plans already drown in data. It’s earlier data, on a person the plan has not met yet, generated by behavior instead of by billing.
This is not theoretical, and the research is better than most people realize
Researchers at Imperial College London ran a study on six years of loyalty card purchases from two retailers, comparing 153 women later diagnosed with ovarian cancer against 120 who weren’t.
Women who went on to be diagnosed were buying more pain medication and more indigestion medication up to eight months before diagnosis. On average those same women didn’t recognize their own symptoms until four and a half months before diagnosis, and didn’t see a doctor until three and a half months before.
The receipts knew first. Published in JMIR Public Health and Surveillance, and it’s one of a growing pile of studies in the same direction.
Now put that next to a Medicare Advantage book. A plan’s hardest population is the member who is deteriorating and hasn’t shown up in a claim yet, because there is nothing to code, nothing to close, and no way to find her until she lands in an emergency room. That member is buying antacids at Costco right now.
Here’s my actual prediction
The next round of retail health partnerships will be negotiated over data, not shelf space.
Not because anyone plans it that way. The first conversation will be about distribution, because distribution is easy to price and easy to explain to a board. But somebody on the payer side is going to run the numbers on what a retailer’s purchase history is worth against a population they can’t otherwise see, and the terms of these deals will change fast.
I’d expect the shift inside eighteen months. Every plan in the country is starved for early signal on members, the supplemental benefit spend is already flowing through retailers, and the co-branding gives both sides a reason to sit at the same table with lawyers who understand what they’re building.
Whoever writes the first serious version of that agreement is going to set the template for the rest of the decade.
And here’s the part that will get plans in trouble
A purchase is not a diagnosis.
You cannot code from a shopping cart. You cannot submit a shopping cart. No auditor has ever accepted a shopping cart, and none ever will. A member buying antacids might have ovarian cancer, or acid reflux, or a son visiting for the holidays who likes spicy food.
This matters more than it sounds like it does, because our industry has a specific failure pattern with new data sources. We acquire the feed, we get excited about the correlation, and then we discover there is no path from the signal to anything we can defend. The data sits in a warehouse generating slide decks. Meanwhile the plan has taken on the privacy exposure, the vendor cost, and the expectation that something will come of it.
Signal you can’t turn into evidence isn’t an asset. It’s a liability with a subscription fee.
The retailers solved distribution decades ago. The plans that get anything out of these partnerships will be the ones who already knew what to do with a signal before the signal showed up.
What has to exist between the receipt and the record
There’s a specific chain that has to work, and most plans can’t run it today.
A non-clinical signal says look at this member. Something has to decide whether that’s worth acting on, or whether it’s noise, because a retail feed will generate far more suspicion than any plan can chase. Then a real clinical touch has to happen, an assessment or a visit, where a licensed person examines the member and documents what’s actually there. Then the coding has to be supported by evidence in the record, not by the thing that pointed you at her. Then the submission has to be accepted, linked, and hold up when someone comes back and asks you to prove it two years later.
Break any link and the whole thing is theater.
This is why we built Invent Health the way we did. I spent two decades connecting systems, and I came out of it convinced that connection was the last generation’s problem and we’d largely solved it. What nobody solved is the step where a signal becomes something a plan can prove.
Our platform runs suspecting, assessment targeting, coding, quality, and encounter submission against one view of the member, because the handoffs between those functions are exactly where the chain breaks. The Coder Workbench uses a two-pass model with a certified coder in the loop. Our AI surfaces what matters. It never invents a diagnosis, and it never will, because the entire value of the output is that it survives an audit.
We run this for plans covering more than five million members. The plans that will do well with a retail data feed are the ones who already have that chain working. The feed makes a good operation faster. It makes a broken one louder.
Why this decides itself soon
Annual enrollment opens October 15. The Costco plans aren’t in this one, and the regulatory approvals are still pending. That’s not the clock that matters.
The clock that matters is that every strategy team in Medicare Advantage read the same headline this week, and some of them are already drafting the retailer they want to call. Those conversations will start with distribution and drift toward data, and the plans that show up without an answer for what they’d actually do with a purchase history will sign a deal they can’t use.
Distribution is the easy half. It always was.
Talk with our team
If your plan is thinking about what a non-clinical data source would really require to be worth anything, that’s a conversation we’d enjoy.
Frequently asked questions
What does the Costco and SCAN Health Plan partnership mean for health plans?
It signals that retailers have found a role in healthcare that fits what they’re good at. The near-term effect is a new distribution channel. The longer-term effect is that retail purchase data becomes something payers negotiate for, which no health plan has had access to at this scale before.
Can retail purchase data predict health conditions?
Research says yes for some conditions. An Imperial College London study published in JMIR Public Health and Surveillance found women later diagnosed with ovarian cancer bought more pain and indigestion medication up to eight months before diagnosis, earlier than they recognized symptoms themselves.
Can a health plan use purchase data for risk adjustment?
Not directly. A purchase is not a diagnosis and cannot support a submitted code. Non-clinical signals can only direct where a plan looks. The documented condition still has to come from a clinical encounter with evidence in the medical record that holds up under audit.
What data does Costco have on its members?
Item-level purchase history tied to a verified membership identity, accumulated over years. Costco also operates pharmacy, optical, and hearing aid businesses, which puts it closer to health-relevant purchasing than a general retailer.
Will other retailers sell Medicare Advantage plans?
Almost certainly, if this performs. The economics work for both sides, and retail distribution sidesteps rising broker acquisition costs. Expect competitors to look for their own health plan partner rather than build an insurance business.
Why did previous retail health ventures fail?
The failed ventures tried to deliver care. Walmart closed all 51 of its health centers in 2024. Clinical delivery requires licensing, staffing, and long payback periods that do not match retail economics. Distribution and merchandising are a much better fit for what retailers already do well.
What is the risk of using non-clinical data in a payer setting?
The main risk is acquiring a signal with no path to a defensible action. Plans take on privacy exposure and cost, generate more suspicion than they can chase, and end up with correlations that cannot be documented or submitted. The data becomes a reporting exercise.
How do health plans turn a signal into a documented condition?
A signal directs attention, a clinical encounter produces the evidence, a certified coder supports the code from the record, and the submission has to be accepted and linked. Every step must survive an audit. Breaking any link means the plan does the work without getting credit for it.
Do you need a Costco membership to buy the Medicare Advantage plan?
No. Federal rules prohibit requiring a separate paid membership as a condition of Medicare Advantage enrollment. The companies confirmed this when announcing the partnership.
When will the Costco Medicare plans be available?
No launch date has been announced. Co-branded Medicare Advantage plans are planned for two states and a Medicare supplement plan for a third, all pending regulatory approval in each state.
What should a health plan do before signing a retail data partnership?
Confirm the chain from signal to accepted submission works internally first. If suspecting, assessment targeting, coding, and encounter acceptance are running on disconnected systems, an external data feed adds volume to a process that already leaks.
Does AI make purchase data usable for risk adjustment?
AI can rank and prioritize a non-clinical signal so limited clinical resources go to the right members. It cannot substitute for the clinical documentation itself. Any model that generates a diagnosis rather than surfacing evidence creates audit exposure rather than value.
