How to launch an agentic commerce pilot that earns its next phase
An agentic commerce pilot should test one customer task, one route to purchase and one commercial hypothesis. Start with a limited catalog and an explicit approval model, prove that orders and exceptions can be handled, then compare contribution after fees and support costs with an appropriate baseline. Expand only when the evidence supports the next step.
For a commerce leader, the difficult question is where to start. A broad launch can combine unfamiliar traffic, new checkout behavior, new payment arrangements and new customer expectations. If results disappoint, the team cannot tell which change caused the problem. A narrow pilot makes that diagnosis possible.
The framework below is AIR's proposed operating approach, not an industry certification or a promise of a particular result. It follows one illustrative retailer selling replacement water-filter cartridges. The product has a repeat purchase need, but compatibility still matters: a customer ordering the wrong cartridge creates a support problem even when payment succeeds.
Start with a customer task that has a clear finish
Write the task in a sentence: “Help an existing customer reorder the exact cartridge that fits their filter, within their chosen budget.” That is more useful than “launch AI shopping.” The first sentence identifies the buyer, product decision and constraint. It also makes clear what the pilot excludes: recommending a different filtration system, creating a subscription or choosing an unapproved substitute.
Select products whose identifiers, variant attributes, delivery options and return terms are dependable. High demand alone does not make a good pilot category. Products with frequent stock corrections or ambiguous compatibility can turn a small experiment into a test of unresolved catalog problems.
AIR's guide to how AI shopping agents rank products explains the discovery stage. For the pilot, move one step further: decide what evidence will prove that a particular recommendation can become the intended order.
Choose the purchase route before choosing an integration
Three starting scopes produce different evidence:
| Pilot scope | What it tests | What it cannot establish alone |
|---|---|---|
| Agent recommendation with a link to the store | Whether shoppers arrive with useful purchase intent | Whether an agent can create or complete an order |
| Agent-prepared cart, completed by the customer | Whether the selected products survive the handoff accurately | Whether delegated purchasing works without that handoff |
| Checkout within an assistant or agent experience | Whether the merchant can support that surface's transaction flow | Whether every step is autonomous or portable to another platform |
The distinction is practical. Google's UCP integration guide, updated November 17, 2026, describes separate validation paths for transfer to a merchant cart and native checkout on Google. Native checkout also requires order-status validation and a support-team review before activation. Availability and onboarding conditions must be checked for the merchant's own account. Source: Google Merchant Center UCP integration guide.
For the cartridge retailer, a customer-approved cart may be the right first experiment. It can test accurate product selection and reduced re-entry without also changing the customer's payment experience. A subsequent phase could test a broader authorization model, if the earlier results justify it.
Put named owners behind four launch gates
Dates organize work, but evidence should control launch. Assign an owner who can refuse progression at each gate.
Catalog gate: commerce operations. The agent's selected cartridge ID must map to the purchasable variant, with a current price and deliverable destination. Test similar product names and incompatible models deliberately. The acceptance record should show which inputs were tested and which version of the catalog supplied the answer.
Authority gate: product and risk. Record what the customer allowed, which actions still need confirmation and when permission ends. Recognizing an agent, recognizing the customer, proving eligibility for an offer and approving payment answer different questions. AIR's merchant Know Your Agent checklist develops those checks. A returning-customer badge should never be treated as permission to place another order.
Order gate: engineering and payments. Prove that the merchant can reconcile the order and payment after a lost response. Exercise unavailable stock, changed totals and repeated requests. A customer-facing “something went wrong” message cannot be the only record of a potentially completed purchase.
Service gate: customer operations. Give support a way to find the order, understand what was approved and route a cancellation or return. The agent channel should appear in the support workflow before the first live customer needs it.
This approach is consistent with treating evaluation as part of deployment. The US National Institute of Standards and Technology (NIST) released its voluntary AI Risk Management Framework on January 26, 2023. It addresses trustworthiness across design, development, use and evaluation. The four commerce gates above are our application of that principle, not NIST-prescribed commerce requirements. Source: NIST AI RMF.
Test failures that could change the customer's decision
The most revealing rehearsal starts with the right product and then changes the conditions. Raise the delivery charge, remove the last unit from stock, expire the customer's permission or make the confirmation response unavailable. Check whether the system pauses, recalculates or asks for approval at the appropriate point.
Google's native-checkout documentation distinguishes incomplete, ready-for-payment, ready-for-complete and completed sessions. Its documented flow also includes a buyer-facing checkout interface; an agent-related purchase is not necessarily a purchase made without human input. These details show why the pilot must record the exact flow being evaluated. Source: Google native checkout overview, updated November 3, 2026.
For the retailer, an incompatible cartridge suggestion should stop the journey even if it is cheaper. A total above the customer's approved limit should trigger a new decision. A delayed confirmation should trigger order reconciliation. Capture the expected response before running each scenario, so a convincing chat answer does not become the acceptance criterion.
Make the expansion decision with contribution, not order count alone
Consider an illustrative pilot with 100 completed orders at $60 each. Revenue is $6,000. Assume $3,600 product cost, $600 fulfillment subsidy, $180 channel and payment fees, $240 incremental support cost and a $180 allowance for expected returns. Contribution after those stated variable costs is $1,200, or $12 per order. These are invented planning numbers, not AIR or market results.
Suppose a comparable existing channel contributes $14 per order. The pilot has not yet demonstrated better unit economics. It may still be worthwhile if it reaches incremental buyers or produces repeat purchases, but those are hypotheses to measure. Reassigning customers who would have bought anyway does not create the same value as attracting additional demand.
Keep setup costs separate, then include them in the investment decision. At $12 contribution per genuinely incremental order, $12,000 of pilot setup cost would require 1,000 such orders to recover, assuming the same economics and excluding ongoing fixed costs. That arithmetic makes the scale question concrete.
Use comparable cohorts where practical, record channel overlap and let the return window mature. A short pilot can establish that the operating flow works; it may be too small to establish sustained commercial lift. The appropriate next decision can be a longer measurement period rather than immediate expansion.
Introduce customer evidence when it changes the offer
Once the basic flow works, a useful next experiment might test a member benefit. The retailer should specify the accepted membership issuer, required status, validity and permitted use before asking the customer to share evidence.
AIR Identity supports credential issuance and verification with user-controlled sharing. That makes customer eligibility a relevant area to assess with AIR. It does not establish that this hypothetical retailer has integrated AIR or that any particular checkout protocol is supported. Keep the eligibility test separate from the customer's permission to spend.
The expansion proposal should name the one new capability being tested, the owner, the expected commercial effect and the condition for stopping. That keeps each phase understandable enough to fund or decline on evidence.
Frequently asked questions
What is an agentic commerce pilot?
It is a limited test of an AI agent's role in a specific shopping task. Define the products, customer group, purchase route, approval model and success measures before allowing live orders.
Do merchants need autonomous payments for a first pilot?
No. A pilot can start with product recommendations or an agent-prepared cart that the customer reviews and pays for. The chosen scope determines what the results can demonstrate.
How long should an agentic commerce pilot run?
Long enough to observe the intended customer task, operational exceptions and relevant return or repeat-purchase window. A calendar deadline alone does not establish that the sample is sufficient to make an expansion decision.
What should stop a pilot from launching?
Unresolved product mismatches, unclear customer authority, orders that cannot be reconciled with payments, or a missing service route should block the affected flow. Set the acceptance criteria and responsible owners before testing.
Where can AIR fit in the pilot?
AIR Identity is relevant when the merchant needs customer credential evidence, such as eligibility for a member offer. Assess the exact integration separately from checkout execution, spending permission and payment approval.
Partner with AIR
Planning a merchant pilot that needs to recognize eligible customers across participating services? Talk to AIR about the credential evidence your commercial flow requires.
Moca Network, the identity network of Animoca Brands.
Partner with AIR
AIR is one integration suite for identity, money and loyalty. AIR Identity lets trusted issuers issue verifiable credentials and businesses verify them, so acquisition starts with users who already qualify, with minimal data custody. AIR Money embeds compliant global financial infrastructure, opening new markets without building new rails. AIR Loyalty gives brands a programmable, stable-value points engine that keeps users coming back.
Partner with us to grow your business.
AIR is built by Moca Network, the identity network of Animoca Brands.




.png)