The Partnership Yield Problem | PastBehavior Research

The opening contrast

State Farm Bank

1999 to 2020

The access

Roughly 19,000 agents and approximately 84 million policies and accounts. Every agent held a standing relationship with customers who kept money somewhere. FACT

The outcome

Twenty-one years after founding, the bank held $16.4 billion in assets, produced a 0.35% return on assets in 2019, and shrank by $500 million that year. State Farm announced its exit from banking in March 2020, handing deposits and card accounts to U.S. Bank. FACT

The cause cannot be isolated. Bank capital rules inside an insurer, a decade of low rates, and an industry-wide retreat from insurer-owned banks all point the same direction as the channel mechanics. UNRESOLVED

Financial Engines

1996 to 2018

The access

Managed accounts held roughly 4% of defined contribution plan assets, and more than 70% of plan participants could not correctly identify what a managed account is. A product most eligible people do not understand, cannot name, and do not select. FACT

The outcome

The company reached $169 billion in assets under management, served more than 750 of the largest U.S. employers and the largest recordkeepers, and sold to Hellman & Friedman for $3.02 billion at $45.00 per share, announced April 30, 2018. FACT

Apple Card

2019 to 2026

The access

Embedded in the Wallet app on a device base in the billions, with no fees and a signup flow measured in taps. As close to ideal distribution architecture as the record contains. FACT

The outcome

The portfolio reached roughly $20 billion in balances. Goldman Sachs lost billions on it, and on January 7, 2026 JPMorgan announced it would take over the portfolio, with Goldman accepting a discount reported at more than $1 billion. FACT

Reach tells you how many customers might be available. It tells you the size of the opportunity, not the value of the channel.

01 / Access is a ceiling

Announced reach is not actual distribution.

A partnership announcement gives you one number, and it is the flattering one. Announced reach appears in nearly every press release. Activation, the number of people who actually encountered the product, appears in maybe one in five. That asymmetry is not an accident of disclosure rules. Neither party benefits from publishing the second number unless it is remarkable.

10 million customers≠10 million exposures≠10 million users≠a good business

The Hinge Health funnel

Access20.1M

Trailing-twelve-month average eligible lives at the end of 2025, across more than 2,250 client organizations. The number a press release would print. FACT

↓

Adoption783K

Members at the end of 2025. A yield of roughly 3.9% of the eligible base. FACT

↓

Economics$588M

Revenue for 2025, up 51%, with net dollar retention above 110% and client retention of 98%. About $29 of annual revenue per eligible life. FACT

Hinge Health prospectus (filed March 2025) and 2025 earnings disclosures. The company went public at $32 a share.

The point is not that 3.9% enrollment is bad. The point is the opposite. A relatively small percentage of an enormous eligible base can create an excellent business when each converted customer is economically valuable. Anyone reasoning from the first number alone would have been wrong about the business in both directions: overestimating how many people the product reaches, and underestimating what a small share of them is worth.

Installed-base size sets a ceiling. It does not set the outcome. The question is never how far below the ceiling you land. It is what each landed customer is worth, and for how long.

02 / Once vs. every time

The distinction that predicts outcomes is not human versus machine.

Every partnership requires behavior between the agreement and the customer. The structural question is how many times that behavior has to happen. Map the two extremes side by side.

Compounding behavior

Affirm through Shopify

Affirm signs one exclusive commercial agreement with Shopify, July 2020. FACT

A merchant enables Shop Pay Installments in the admin dashboard. Affirm describes it as taking a few clicks. FACT

Every subsequent checkout surfaces the option and underwrites the shopper in real time. The merchant does nothing further, ever. FACT

One integration by the platform. One toggle per merchant. Zero behaviors per transaction.

Active merchants went from 6,500 in Q1 2020 to 102,000 by September 2021 to 168,000 by December 2021, with Affirm attributing the substantial majority of the increase to Shop Pay Installments. Affirm handled $10.8 billion of gross merchandise volume in the quarter ended September 2025. FACT

Repeated behavior

A digital advice platform through a wirehouse

The bank signs one central agreement. Advisors are trained.

The advisor recognizes that a client fits the offering.

The advisor remembers the offering exists.

The advisor judges it appropriate, explains it, and hands the client off, knowing the client may never come back.

One integration. One training event per advisor. Then a full chain of judgments per client, per conversation, forever, performed by someone whose compensation is not improved by performing it.

UBS held equity in SigFig from 2016 and private-labeled its technology as Advice Advantage from 2018. Its advisors decided client by client. UBS disclosed in a Form ADV Part 2 update dated March 31, 2025 that it would close the service. Nine years, a central agreement, and a decentralized surface the vendor never touched. FACT

A behavior that happens once and keeps producing distribution becomes an asset. A behavior that must happen again for every customer remains an operating expense.

Aflac

The counter-case that disciplines the finding

Aflac looks like the most human-mediated business imaginable: independent agents selling worksite policies one employee at a time. But look at what happens after the sale. Premiums come out through payroll deduction. Agents are paid commissions on first-year and renewal premiums. Premium persistency ran 78.6% in 2023. The human behavior happens once per policy, and the revenue arrives every pay period for years without anyone acting again. FACT

Aflac's actively producing U.S. agent count fell from an average of 17,300 in 2005 to roughly 5,300 in 2025, while new annualized U.S. premium sales reached $1,558 million in 2023. Fewer humans, more sales, because the channel shifted toward brokers and digital enrollment platforms, and because the installed book keeps paying regardless. FACT

The finding is not that human channels are weak. It is that the question is whether the distribution behavior has to be recreated for every incremental dollar of revenue.

03 / What the partner gets

Does the partner benefit from signing, or from the partnership working?

The question that changes how several cases read: why does the party controlling distribution want this product used? The record shows a structural distinction between partners who benefit from signing and partners who benefit every time the product is used, and they behave completely differently.

$8.2B

Delta's remuneration from American Express in 2025, up 11%. The company's 10-K notes it expects the figure to reach $10 billion. FACT

~14%

Approximate share of Delta's operating revenue represented by that payment stream. FACT

1M+

Card acquisitions reported by Delta for the fourth consecutive year. FACT

American Express pays Delta for SkyMiles currency, and the airline's own P&L moves with the partnership. Delta's cabin announcements, its app, its check-in flow, and its gate agents all promote the card. Ed Bastian discusses the partnership on earnings calls. Nobody at Delta has to be trained to remember the card exists, because the partner's revenue depends on it. FACT

Benefit at signing

The partner gets a checkbox, a differentiation talking point, a small placement fee, or a venture position.

↓

The incentive is satisfied at signature. Everything after it is a cost center inside the partner's organization.

↓

Activation remains primarily the vendor's problem.

Benefit at usage

The partner gets value every time the customer uses the product.

↓

The partner's own economics improve when the partnership performs.

↓

The partner has its own reason to activate distribution, and does work the vendor could never buy.

Walmart provided shelf space for Green Dot's cards and received a commission. The channel worked: the Walmart MoneyCard exceeded a million account holders, and Walmart's share of Green Dot's operating revenue was reported at 63% in a 2010 quarter. FACT

Then the share declined steadily: 46% in 2015, 40% in 2017, 34% in 2019, 27% in 2020. Walmart eventually pursued its own consumer financial venture. A retailer earning a commission on shelf space has a bounded interest in the outcome. FACT Whether the decline reflects channel mechanics or the retailer's own ambitions cannot be cleanly separated. UNRESOLVED

Shopify took an equity stake in Affirm, reported as worth roughly $2 billion at Affirm's IPO, and made Affirm the exclusive provider of Shop Pay Installments. Shopify's interest is not a placement fee. It is higher merchant conversion and higher average order value across its own gross merchandise volume, the thing Shopify is measured on. That alignment is why merchant-facing enablement is one toggle rather than a sales process. FACT

Test the partner's incentive before the architecture. When the partner's economics improve with usage, the vendor is no longer solely responsible for creating activation.

04 / Activation is not value

Two cases break the intuitive assumption from opposite directions.

The intuitive assumption is that higher activation means a better partnership. Low activation can create an exceptional channel when customer value and persistence are high. Near-perfect activation can destroy billions when the underlying customer economics are bad.

Low penetration, exceptional outcome

Financial Engines

Penetration: Roughly 4% of total defined contribution plan assets in managed accounts (Cerulli survey data). FACT
Comprehension: More than 70% of plan participants could not correctly identify what a managed account is (Cerulli's 2024 research with Edelman Financial Engines). FACT
Scale: $169 billion in assets under management; more than 750 of the largest U.S. employers; more than 10 million employees with access. FACT
Outcome: Sold to Hellman & Friedman for $3.02 billion, announced April 30, 2018, twenty-two years after founding. FACT
The meter: A percentage of assets in the accounts it managed. A participant who enrolled once kept paying, and the amount grew with markets and contributions, without anyone acting again. FACT

Near-perfect activation, catastrophic economics

Apple Card

Architecture: Launched 2019 with Goldman Sachs as issuer. Embedded in the Wallet app on a device base in the billions. No fees, up to 3% Daily Cash, a signup flow measured in taps. FACT
Scale: Activation as good as consumer credit activation gets. The portfolio reached roughly $20 billion in balances. FACT
Outcome: Goldman lost billions. JPMorgan announced on January 7, 2026 that it would take over the portfolio, with Goldman accepting a discount reported at more than $1 billion and JPMorgan setting aside an additional $2.2 billion provision for future credit losses. FACT
Aftermath: Goldman's second quarter 2026 8-K shows Platform Solutions net revenues of $221 million, down 64% year over year, primarily reflecting markdowns on the Apple Card portfolio moved to held for sale. FACT
The meter: Consumer credit economics on terms the partner set, including no fees. The distribution design worked perfectly and delivered negative unit economics faster. FACT

Distribution architecture is a multiplier on unit economics. It cannot rescue bad economics. Applied to a negative number, it makes things worse faster.

The difference is the economic meter. Different partnerships get paid through different meters, and identical activation produces radically different value depending on which meter is running.

SubscriptionTransaction feesReferral feesPercentage of assetsCommissionsRevenue sharePer-user pricingUsage

Teladoc and Hinge Health sell to the same buyers, employers and health plans, and both make a population eligible and then need individuals to act. The meters differ, and so does what activation is worth.

Teladoc

Per eligible head per month, whether or not anyone uses anything.

Utilization sat under 12% for most of its first decade: 2.1% in 2013, 8% to 11.1% by quarter in 2019, 17.7% in the fourth quarter of 2020. Quarterly PEPM guidance in late 2018 was around $1.10, roughly $13 per member per year. By 2022 the company stated that one in four Americans had access through an employer or health plan. FACT

When the meter prices eligibility, activation is a cost.

Hinge Health

Contracts that increasingly price engagement.

About 50% of eligible lives had moved to an engagement-based pricing model by the end of 2025, and the company spends its own money enrolling: its prospectus states that for the term of each contract it is able to enroll, engage, and re-engage the client's eligible lives. FACT

When the meter prices engagement, activation is an investment.

The arithmetic, in the abstract

Channel A activates 40% of an eligible base at $12 of annual revenue per activated person.
Channel B activates 4% of an eligible base at $400 of annual revenue per activated person, compounding for fifteen years.
Channel A yields $4.80 per eligible person per year and stops when the contract ends.
Channel B yields $16 per eligible person in year one, and materially more thereafter.

Channel A looks better on every activation dashboard in both companies. Activation rates are the variable that is easiest to measure and easiest to move. They tell you how much distribution you got, not what it was worth.

Activation determines how much distribution you get. The economic meter determines how much that distribution is worth.

05 / The partnership yield model

Distribution multiplies. So does value. They do not weigh the same.

Distribution

Installed base×Eligibility×Trigger×Activation×Conversion

Varies within roughly one order of magnitude across the measurable cases. The most heavily managed terms.

Economic value

Value per customer×Persistence×Retained share

Varies across four or five orders of magnitude. The least visible terms before signing.

Partnership yield = distribution × economic value

Not a formula to compute. A discipline about which terms to negotiate, instrument, and protect.

The terms on the left are the ones companies manage, because they are the ones that are visible: installed base, advisors, merchants, employees, members. Across the measurable cases in this study, the entire reach chain varies within roughly one order of magnitude, with realized yields between about 4% and about 40% of the eligible population. FACT

The terms on the right vary across four or five orders of magnitude: from about a dollar a month per eligible head to a compounding percentage of a growing asset balance held for a working lifetime. FACT

No model that treats these as co-equal factors will serve its users, because the terms that vary most are precisely the ones no one sees before signing. INFERENCE

What is observable before signing

What is actually observable before signing, drawn from the cases in this study:

Installed baseAlways. It is in the announcement.
EligibilityUsually.
Activation and conversionOnly where the vendor controls the enrollment surface and chooses to publish. In this study: Hinge Health, Teladoc, Trupanion, Affirm.
Value per customerFrom filings, if the vendor is public and the channel is most of the business.
PersistenceFrom retention and persistency disclosures, where they exist.
Retained shareEssentially never in advance. Only in retrospect, through renewal disclosures buried in risk factors.

The two terms that vary most are the two hardest to observe before signing. The one that varies least is the one printed in the press release. INFERENCE

06 / The decision surface

The contract is signed centrally. Distribution happens locally.

For every case in this study, one question: what is the last thing that determines whether an end customer encounters the product? Not the contract. The last thing. The answers cluster into five.

Vendor

The partner grants access once. The vendor enrolls continuously, on its own schedule, with its own money, measured by its own team.

Hinge Health. The prospectus is explicit: for the term of each contract, the company is able to enroll, engage, and re-engage the client's eligible lives. A partner-granted list and a vendor-controlled surface. FACT

Partner, centrally

One decision at headquarters produces exposure across the whole base.

Apple decided the Apple Card would live in the Wallet app, and every iPhone user encountered it. Shopify decided Shop Pay Installments would appear in checkout. The architecture everyone wants; two of the three examples in the record produced disasters for the vendor, for reasons that have nothing to do with the architecture. FACT

Partner, locally

A central agreement, then thousands of local decisions by people who were not party to the negotiation, do not know the terms, and have their own priorities.

UBS advisors decided client by client whether SigFig's Advice Advantage was the right destination. State Farm signed one decision; 19,000 agents executed it, or did not, one conversation at a time, for twenty-one years. FACT

Customer

The customer must decide they need something, search for it, evaluate it, and install it. Nobody recommends it to them.

The Atlassian Marketplace: no intermediary, no prompt, and more than $6 billion in lifetime sales since May 2012. Customer-triggered distribution is weak when the customer has no reason to look, and strong when the customer arrives already looking for the thing. FACT

System

Eligibility, underwriting, and the offer fire automatically at the moment of need. No human decides anything after the initial enablement.

Affirm inside Shop Pay. Once a merchant enables it, every checkout is underwritten in real time. FACT

A centralized partnership agreement routinely produces decentralized distribution, and the gap between the two is where most partnership value evaporates. A head office signs. The customer is reached by a branch, an agent, an advisor, a broker, or a store manager who was not party to the negotiation.

Control of the decision surface predicts exposure. It does not predict economics. Apple controlled the decision surface completely, and Goldman still lost billions.

07 / Growing on someone else's surface

Even when distribution works, the economics can change.

A successful partnership creates a second problem: the partner now knows exactly how valuable its distribution is. The party that owns the customer surface has leverage at renewal, and the record shows it using that leverage whether the vendor is a six-billion-dollar partner ecosystem or a single anchor customer.

The anchor customer re-prices

Marqeta at Block

Marqeta's revenue from Block ran at approximately 68% in 2023, 47% in 2024, 45% in 2025, and 41% in the second quarter of 2026. FACT

The August 2023 Block Amendment renewed the Cash App agreement on different terms, which reduced reported net revenue and made Block responsible for defining and managing the program with respect to the primary card network. Marqeta's CEO described a pricing tier that steps down at twice the size of other tiers to incentivize volume growth, which Block reached in December 2025. FACT

The platform raises its take

Atlassian's marketplace partners

The marketplace passed $1 billion in lifetime sales, $2 billion in December 2021, $4 billion in late January 2024, and more than $6 billion by November 2025, across more than 1,800 partners. FACT

Atlassian then announced that the revenue share it retains on Connect apps would rise from 15% to 20% on January 1, 2026, and to 25% on July 1, 2026. FACT

Growth on someone else's surface is priced at renewal. If the channel works spectacularly, how much of the economics will you still own five years from now?

08 / Five questions

Before you count the customers.

Before getting excited about the size of a potential partner's installed base, five questions. Each one traces to a case in this study.

01
What has to happen for one customer to encounter us?
Count the behaviors between the agreement and the encounter, and name who performs each one. Several of the clearest breakdowns occurred where distribution required repeated per-customer behavior from someone whose compensation or primary job did not depend on performing it. INFERENCE
02
Does that behavior happen once or every time?
A behavior performed once is an asset. A behavior repeated per customer is an operating expense that decays the moment training stops, the champion leaves, or a competing priority arrives. INFERENCE
03
What does the partner gain when the customer actually uses us?
Value at usage produces partners who do activation work you could never buy. Value at signing produces partners who have already collected everything the partnership was going to give them. INFERENCE
04
What is one converted customer worth, and does that value persist?
Design for the low conversion number. Nothing in this study achieved better than roughly 40% of an eligible base, most achieved under 20%, and the best economic outcome in the set operated at around 4%. If the business only works at 30% penetration, it does not work. FACT
05
If this channel works, how much of the economics will we still own?
The surface owner re-prices at renewal, whether the vendor is a $6 billion partner ecosystem or a single anchor customer. Negotiate the renewal terms at signing, not after the channel has proved out. INFERENCE

Only after answering those questions does the size of the partner's installed base matter.

The conclusion

Access is not distribution, and distribution is not value.

Partnerships are announced with the biggest number available. Twenty thousand advisors. Ten million customers. Five hundred thousand merchants. Those numbers are easy to understand. They are also the beginning of the analysis rather than the answer.

The distinction the industry needs is between the words it uses interchangeably.

Access: A list of people you are now permitted to reach. It is what gets announced.
Activation: The share of that list who actually encounter the product, set by how many behaviors stand between the agreement and the encounter, and by who has to perform them.
Adoption: The share who act, which in practice lands one to two orders of magnitude below the announced number and should be planned for there.
Economics: What each adopter is worth, how long they stay worth it, and how much of that you still hold after the partner re-prices the relationship at renewal.

Access is the only one anyone publishes. Economics is where the value of the channel is set.

The biggest partner is not necessarily the best partner. The best partner is the one whose distribution mechanics, incentives, and economics turn access into durable yield.

Methodology and limitations

Case selection: Twenty partnerships across financial services, benefits, insurance, healthcare, payments, and software, drawn from three decades of announcements. Twelve carry activation, adoption, or economic evidence strong enough to reason from. The research plan called for twenty-five to thirty; several plausible cases were dropped rather than described without measurement, including arrangements that publish nothing past the announcement.
Evidence labels: FACT: documented in filings, prospectuses, earnings disclosures, or on-record statements. INFERENCE: a conclusion drawn across cases, labeled as such. ESTIMATE: analyst estimate rather than disclosure, flagged at point of use. UNRESOLVED: the public record cannot isolate the cause.
Source hierarchy: SEC filings, prospectuses, 10-K and 8-K disclosures, Form ADV updates, and earnings calls first. Analyst estimates are flagged at the point of use. Secondary characterizations (a trade-press description of one bank's strategic intent, a founder's retrospective account) are used only where they are the sole articulation of a mechanism, and are labeled accordingly.

Known weaknesses

Activation data is missing systematically, not randomly. It is published by pure-play vendors whose entire business is one channel, and withheld by everyone else. Every conclusion here is weighted toward companies where the channel was the whole business, and away from partnerships inside diversified firms, where the quiet successes and quiet failures both live.
Failure bias. Endings are documented and continuations are not. A partnership that renews on similar terms for a decade generates zero coverage, so any pattern drawn from visible endings overstates how often these arrangements end badly.
Survivorship bias in the other direction. Financial Engines, Aflac, Trupanion, and Atlassian are legible because they succeeded. There is no way to count the vendors who signed similar arrangements, activated similarly, and never reached disclosure.
Causality is unestablished in three cases: State Farm Bank, the SigFig and FutureAdvisor pair, and Green Dot's decline. In each, a channel explanation and at least one unrelated explanation fit the same facts.
Two figures in the underlying study are analyst estimates (a William Blair note, November 2025, covering attach rates and share of wallet at two e-commerce partners). They are flagged where used, and nothing structural rests on them.
The study is U.S.-weighted and financial-services-weighted, because that is where disclosure obligations produce the numbers this question needs.

Evaluating a distribution partnership?

PastBehavior studies what has already happened in adjacent markets, separates durable patterns from attractive stories, and helps companies turn that evidence into partner strategy and execution.

Discuss a distribution problem

State Farm Bank

Financial Engines

Apple Card

Announced reach is not actual distribution.

The distinction that predicts outcomes is not human versus machine.

Affirm through Shopify

A digital advice platform through a wirehouse

Aflac

Does the partner benefit from signing, or from the partnership working?

Two cases break the intuitive assumption from opposite directions.

Financial Engines

Apple Card

Distribution multiplies. So does value. They do not weigh the same.

The contract is signed centrally. Distribution happens locally.

Even when distribution works, the economics can change.

Marqeta at Block

Atlassian's marketplace partners

Before you count the customers.

What has to happen for one customer to encounter us?

Does that behavior happen once or every time?

What does the partner gain when the customer actually uses us?

What is one converted customer worth, and does that value persist?

If this channel works, how much of the economics will we still own?

Access is not distribution, and distribution is not value.

Methodology and limitations

Evaluating a distribution partnership?