# Why commerce partnerships take different forms 13 October 2026 · Internal analytical result · Held research cutoff 7 October 2026 ## Main conclusion **Partnership form reflects which parts of a purchase can travel independently, which parts still need to work together, and whose rules govern their connection.** AI gives a new actor the ability to interpret intent, compare offers and initiate action. That creates an opportunity to separate the customer interaction from the systems performing the purchase. The resulting partnership does not follow automatically from the industry's name or a company's size. It depends on the purchase scope, provider capabilities, customer state, installed routes to suppliers, permission requirements and the architecture the AI channel offers. The coding covers all **81 direct episode/role records across 58 case bundles**, with 12 explicit condition statuses per record. These are judgement-based derivations over the recovered source-linked audit, not 81 independent commercial alliances. Seven context cases and one boundary case remain outside the direct matrices. Every positive classification has a case-specific rationale and the inherited source apparatus. No website changes were made. `DISTRIBUTION-MATRICES.md` gives all cells and sensitivity views. `distribution-matrices.json` provides the contributing case and episode IDs for every cell. `coded-arrangements.json` preserves the detailed bases, unknowns, dates and explanatory leads. ## 1. A transportable purchase supports channel checkout; a coordinated bundle retains more provider execution Eight case bundles explicitly describe shipped-goods scope. All eight include channel checkout, and two also include later handoff arrangements. Eight cases explicitly describe a multi-item or grocery cart: four handoffs and four embedded provider apps. No channel-hosted checkout is coded in those basket episodes. The wider grocery/multi-item task cohort has 11 cases: the same eight, two unresolved execution routes and one test without checkout exposed. The latter three do not establish cart construction and therefore do not receive the cart-present code. The distinction survives a bounded comparison inside Google: six shipped-goods cases have channel checkout (four documented, two planned); the two explicit basket cases use Instacart handoff and a planned Walmart embedded experience. This narrows the explanation beyond a simple OpenAI-versus-Google contrast, though company composition and dates still differ. **Mechanism:** a specified item with price, payment and shipping details can cross a relatively narrow interface. Combining local inventory, an existing basket, membership and fulfilment choices may require carrying a larger part of the provider's system into the new channel, or returning the user to that system. Baldwin's crossing-point lens asks where the purchase can be separated without forcing the new channel to reconstruct the whole task network. **Judgement:** operational coupling is a plausible explanation for the contrast, supported by explicit eligibility exclusions in Shopify–Google and the wider bundles in Target/Walmart apps. Grocery alone is not evidence that a specific party handles substitutions or dispatch; many sources do not allocate those roles. Integration maturity and channel policy remain complementary explanations. **Consequence:** greater portability can shift customer interaction without removing the provider's execution role. If interfaces begin exposing complete baskets and coordination reliably, the boundary can move again. The prediction is about changing interfaces and dependencies, not permanent protection for grocery firms. **Cases:** CP1:SH, CP1:ET, CP1:WF, CP1:TG, CP1:SN, CP1:GP; CF5:WM-2, CF5:TG-1, CP1:WM; CP4:G-IC-O/A/GA/AM, CP4:G-DD/SW/SWI. ## 2. A customized order can leave the provider's interface while remaining in its operating system Five case bundles explicitly expose customized ordering or a POS connection. Three execute an externally initiated order through provider/POS systems (Swiggy Food, Square and Toast); two import AI into their own ordering systems (Papa Johns and Wendy's). **Mechanism:** a callable order interface allows the assistant to construct or transmit an order while the kitchen, menu rules and operational execution stay with the provider. The alternative is to bring conversational capability directly into the existing ordering channel. Both preserve operational coordination; they differ in who supplies customer access and who operates the conversation. **Judgement:** complexity does not require a handoff. The important distinction is between moving the interaction and moving the operating system. The Swiggy test's cash-on-delivery restriction also shows that payment readiness can limit completion even where order tools already work. Restaurant context without explicit customization (Uber Eats) is not placed in this five-case row. **Consequence:** POS platforms can become routes around existing demand intermediaries, provided restaurants can acquire demand through the new channel at acceptable terms. This is a possible change in intermediation, not evidence that restaurants have already improved margins. **Cases:** CP4:R-SWF, CP4:R-AM (pooled Square/Toast), CP4:R-PJ, CP4:R-WD. The five-case count falls to four when the rollout-ambiguous Papa Johns record is removed. ## 3. Customer continuity and interface location are separate choices Eighteen case bundles have an explicit provider account, loyalty, identity or customer-state connection somewhere in their coded tracks. Three have that condition alongside channel checkout: Target–Google, Target–Microsoft and PayPal–Perplexity. Four have it in embedded provider apps. It also occurs in handoffs and own-system AI. Five other case bundles have an explicit guest route, prohibition on required provider sign-in or omission of loyalty at launch; all five have channel checkout. These are three Shopify channel routes, Etsy–Microsoft and the planned Gap–Google route. The five are heterogeneous restrictions, not five cases of zero data access. The strongest comparison is **Target versus Shopify within the same channel**, repeated in Google and Microsoft. Target carries its account relationship into checkout. The Shopify route cannot require merchant sign-in. The interface location is similar while the account arrangement differs. **Mechanism:** reusable customer identity, loyalty and account state can remain a provider complement even when the channel hosts the transaction. Alternatively, a standardized guest path can make providers easier to substitute and reduce the friction of onboarding many merchants. **Judgement:** the account rules are observed; the reason each firm secured them is not. Bargaining, implementation choices and channel strategy compete as explanations. Company size alone is not established as the cause. Target's detailed Google record has a post-cutoff retrieval qualification; excluding it leaves the Microsoft contrast intact. **Consequence:** customer continuity is a better candidate dimension of strategic position than checkout location alone. A channel could simplify purchase while preserving provider loyalty, or simplify purchase by weakening the need for that relationship. Repeat purchase, data access and substitution behavior would distinguish their economic effects. ## 4. Installed networks support several forms: their shared importance is connection and governance Twenty-seven case bundles supply a reusable connection across merchants, sellers, restaurants or retailers. Fourteen include channel checkout, seven handoff, three provider/POS ordering, one an embedded app, two own-system AI, four payment integration and three unresolved routes. These memberships overlap and include plans and a reusable prototype. This factor is common across forms rather than predicting a single form. Shopify, Etsy, Instacart, payment activation partners and POS systems contribute different kinds of reach. A marketplace's seller population is not the same asset as a merchant-software installation base, so the contributions and routes remain separate in the coding. **Mechanism:** connecting one installed network can lower the cost of connecting many providers. The network's existing functions then shape the bundle: catalogue and checkout for merchant platforms, cart/retailer coordination for grocery networks, order transmission for POS, credentials or merchant activation for payments. **Judgement:** integration economies explain the appeal of the connection. The stronger claim that the connector captures a bottleneck depends on eligibility, defaults, exit and alternatives. Shopify default enrollment and partial exit differ from Microsoft applications through payment activation partners. These rules matter more than merely counting connected firms. **Consequence:** storefront functions can unbundle while coordination and access rebundle around existing intermediaries. If one connector becomes hard to replace across several channels, its governing terms may become a new dependency. Multi-homing and effective exit are the strongest counterforces. The count includes upstream Salesforce and downstream SharkNinja, and common programme rollouts. It is not 27 independent tests or a measure of the number of merchants reached. ## 5. Payment arrangements solve a separate coordination problem: delegated authority Ten case bundles explicitly expose credential, mandate, user-limit, wallet, identity-protection or financing-permission conditions. Nine include payment integration and three standard collaboration; three also support channel checkout. These forms overlap. Only three of the ten have a documented, non-specification permission episode in the held record; much of this part of the architecture remains planned. **Mechanism:** allowing an AI to express purchase intent does not itself give the AI authority to spend, prove approval, select an instrument or accept an order. Tokens, mandates, checks and protections let different actors perform those functions without all of them becoming the merchant or payment processor. **Judgement:** this is a distinct layer of partnership form, not a competing shopper-interface type. The concentration in payment/standards partly follows the role definition and is not an independent causal discovery. The analytical result is the separation of permissions from execution: Stripe's token can coexist with another processor; Klarna retains affordability checks; merchant acceptance remains distinct. **Consequence:** simpler customer interaction can require a more layered industry architecture. The permission provider may gain a role even where the visible channel changes. Adoption, liability allocation and substitution across permission providers determine whether that role becomes durable power. ## 6. Importing intelligence can be a response to valuable provider systems, rather than a failure to externalize the journey Ten case bundles explicitly combine AI with named provider knowledge, rules, CRM or history; all ten use own-system AI in those episodes. Four cases separately expose post-purchase/account-service operations, also within own-system AI. The 10/10 concentration is sensitive to what sources disclose: named knowledge bases are more visible in implementation case studies than in channel launch announcements. It does not show external-channel partners lack such assets. **Mechanism:** AI becomes more useful when it can interpret a firm's menus, product knowledge, records and operating rules. A firm can import model capability into those complements rather than export the full customer relationship. Amazon's explicit multi-model routing makes this division particularly clear: the lab supplies an input to orchestration controlled elsewhere. **Judgement:** the contributing assets are visible; their distinctiveness and substitutability are not measured. Teece's complement lens identifies what must be evaluated before claiming protected margins. A generic assertion that every provider's data is a moat would be unsupported. **Consequence:** an incumbent can become more capable while retaining the customer surface. If the model input becomes easier to switch while the provider's coordination knowledge remains difficult to reproduce, dependence may favor the incumbent. That outcome is conditional; a single multi-model example cannot establish market-wide switching ease. ## 7. Channel policy explains reversals that static industry categories cannot The same Shopify and Etsy relationships contain historical channel checkout followed by handoff; Walmart contains historical channel checkout followed by an embedded provider experience. Asset and industry categories alone cannot account for those changes. The held records identify OpenAI's checkout withdrawal. Across partners, IHG's OpenAI app sends guests to direct booking, while the held Google hotel cohort includes hosted booking and provider processing. Treat IHG's inclusion in that cohort as a matched qualitative comparison, not a newly added independent case. **Mechanism:** labs choose which interfaces, payment arrangements and provider integrations their channels offer. These choices set the available menu within which industry constraints and provider objectives operate. **Judgement:** policy/product architecture is directly implicated in these changes. Why the policy changed - adoption, conversion, liability, economics or coordination problems - is not isolated by the matrix. It is a complementary explanation to operational coupling, not a reason to discard coupling. **Consequence:** a firm can lose an interface position without losing its partnership. Monitoring changed terms and product capabilities can matter as much as monitoring new announcements. ## A framework that explains the distribution The emerging explanation is a sequence of connected choices: 1. **What can AI now do?** Interpret intent, compare, build a cart, initiate an order, reason over service records or support delegation. 2. **What does completing the task still require?** Inventory/capacity, coordinated execution, customer state, expertise and permission. 3. **Which actor supplies that bundle most effectively?** A provider, merchant platform, POS network, payment institution or the AI channel itself. 4. **What does the channel permit, and what do the parties seek?** A hosted transaction, embedded provider experience, handoff, or model input inside the incumbent system; defaults and account rules allocate the connection. This generates four forms of rebundling that can coexist: around the AI channel, around the provider experience, inside the incumbent's operating system, and around connector/permission infrastructure. The matrix informs this theory; it does not reduce the theory to a regression over publicity records. Commercial objectives remain a live explanatory factor. Virgin Voyages explicitly emphasizes relationship-building across a long purchase journey; IHG emphasizes new reach with direct booking; Klarna describes combining its knowledge and UX to challenge shopping incumbents. Most episode summaries do not contain comparable objective statements. The coding records this gap rather than inferring a company's objective from the form it chose. ## Checks, sensitivity and limits - All 81 direct records and all 58 direct case bundles covered once at their respective levels; 12 condition statuses per record. Counts reproduce from the manual coding table. - Same-episode joins prevent Walmart's later account arrangement being attached to historical checkout, or Home Depot's own-channel knowledge being attached to an unresolved external route. - A condition absent from the recovered record is `not_stated`, not a negative finding. Positive co-occurrence cannot establish how often the condition exists in the market or its absence elsewhere. - Removing plans/specifications/prototypes leaves 44 case bundles. Shipped-goods checkout remains in six cases including historical routes; basket handoffs remain four and embedded apps three. Removing explicitly historical records leaves four shipped-goods checkout cases. There is no claim that all other documented records remain live today. - Case bundles vary in granularity and share programmes. No significance tests, causal coefficients, market-share estimates or volume-weighted distributions are warranted. - The review corrected overextensions before release: recipe-card availability alone no longer establishes a cart in Instacart–Copilot, nor do item suggestions establish a cart in the Swiggy grocery test; general SharkNinja engagement/service no longer establishes a named knowledge asset or post-purchase operations. Swiggy's specific grocery test is attributed to ChatGPT rather than all assistants. - Same-analyst coding and review; no independent coder. This extends the held audit rather than refreshing the market census. **Method state:** IIR existing-project adoption; interpretation layer. Bounded corpus integration recovered and extension locally checked; SI-1 theory work applied, global theory registration still adopting; all scoped records represented in the analytical universe, formal global migration incomplete; public expression not evaluated and site redesign paused. See `DESIGN-AND-CODEBOOK.md` for exact scope and the method-check limitation.