What Commitment Governance Looks Like in a Realistic Agentic Transaction

Part 3 of a three-part series.

Parts one and two of this series did the heavy lifting. We named the governance gap that agentic transactions create, then laid out Lu Zhang's Commitment Governance Framework: progressive binding, decision outcomes, and the evidence object that ties them together.

Now we make it concrete.

What follows is an illustrative walkthrough, not a claim of production deployment. The scenario is realistic. The transaction details are the kind product teams deal with every day. The decision points map directly to systems that card networks and processors already maintain. But the framework itself is a proposal, and this walkthrough is designed to show what it would look like in practice.

One transaction. A consumer's AI agent ordering weekly groceries. Six moments where the framework would make a call. By the end, you will see exactly how commitment governance would handle routine fulfillment, catch a scope violation in real time, and resolve a dispute in minutes that would otherwise take weeks.

Every decision point in this scenario maps to a system that product teams at card networks and processors already maintain. The framework plugs into existing infrastructure. It does not replace it.

Decision Point One: The Agent Initiates

Sarah configured her AI shopping agent last month. Weekly grocery run, $150 limit, three approved merchants, grocery category only. The agent selects 22 items from FreshCart, her preferred store, and assembles a $127.43 basket.

Here is where commitment governance would diverge from current systems. The framework would evaluate the delegation scope before anything moves downstream. Amount within bounds. Merchant on the approved list. Category is grocery. All three constraints satisfied.

Decision outcome: allow to proceed. A pending order is created. But the framework would not bind. No payment instruction fires. No merchant system receives the order. No downstream party has relied on anything yet.

That distinction matters. In current systems, the agent's order would already be in a merchant queue. Under commitment governance, it would remain a proposal. Nothing more.

Decision Point Two: Payment Authorizes

The agent submits payment. Visa's network authorizes the $127.43 charge against Sarah's card. Funds are reserved.

Most systems would treat this as the green light. Authorization means go. But the framework would ask a different set of questions. Is the merchant ready to accept? Is inventory confirmed for these specific 22 items? Is the delivery address validated and within the merchant's service area?

Authorization proves the consumer can pay. It does not prove the merchant can deliver.

Decision outcome: conditional binding. Payment is reserved, and FreshCart may begin preparation. But the framework would hold back fulfillment release. The transaction is partially bound. The merchant knows funds exist. The consumer's card reflects the hold. Neither party is committed to completion.

Decision Point Three: The Merchant Accepts

FreshCart's system responds. All 22 items confirmed in stock. Delivery slot available for 2:00 to 4:00 PM. Pricing matches the agent's basket. Store is operational and within delivery range.

Decision outcome: allow to bind for merchant-side preparation. FreshCart could now allocate inventory and schedule the driver. The framework would create an evidence object capturing everything checked so far: delegation scope (amount, merchant, category), payment state (authorized, reserved), merchant readiness (inventory, slot, pricing, operational status), and every constraint evaluated at each decision point.

This evidence object would not be a log file buried in a database. It would be a first-class artifact designed for one purpose: proving, at any future point, exactly what was known and confirmed at the moment binding occurred.

Decision Point Four: Something Goes Wrong

FreshCart's system flags two items as out of stock. Organic whole milk and a specific brand of pasta. The AI agent, doing what agents do, finds substitutes. Different brand of milk from FreshCart. Fine. But the pasta? The agent locates it at GreenBasket, a different merchant entirely. It queues both substitutions.

The framework would catch this instantly. The milk substitution is routine. Same merchant, same category, similar price. But the pasta substitution crosses a boundary. GreenBasket is not on Sarah's approved merchant list. This is not a continuation of the original order. It is a new commitment event.

Decision outcome: the milk swap proceeds. The pasta substitution would require step-up confirmation. Sarah receives a notification on her phone: "Your agent wants to order Barilla penne from GreenBasket ($3.49) because DeLallo penne is out of stock at FreshCart. Approve?"

The agent could not proceed on the pasta until Sarah confirms. The rest of the order would be unaffected. No blocking. No cancellation. Just a pause on the one item that exceeded the delegation scope.

Current systems have no clean mechanism for this. The agent either substitutes freely, creating disputes later, or it blocks the entire order and waits for human input on everything. Commitment governance would thread the needle.

Decision Point Five: Fulfillment Releases

Sarah taps approve. The framework would now re-evaluate the complete order with updated constraints. Delegation scope: $150 limit (new total $128.91, within bounds), approved merchants (FreshCart plus GreenBasket, now authorized via step-up), grocery category (confirmed). Payment: captured for the full amount across both merchants. Delivery: both merchants confirmed, addresses validated.

Decision outcome: allow to bind. Fulfillment would release. Both merchants begin final preparation and dispatch.

The transaction would now be fully binding. The evidence object would record the complete chain: original scope, authorization, merchant confirmation, the out-of-stock trigger, the substitution attempt, the scope violation detection, the step-up request (timestamped), Sarah's approval (timestamped, method recorded), and the fulfillment release with all constraints satisfied.

Every decision point. Every check. Every outcome. One object.

Decision Point Six: The Dispute

Two weeks pass. Sarah reviews her statement and disputes the GreenBasket charge. She does not remember authorizing it.

Her bank pulls the evidence object. It shows the original delegation ($150, approved merchants, grocery), the FreshCart out-of-stock event, the agent's substitution attempt, the step-up confirmation request sent at 10:47 AM on March 24, Sarah's approval received at 10:52 AM via push notification tap, and the fulfillment release at 10:53 AM with all constraints marked satisfied.

The dispute would be resolved in minutes. Sarah authorized the purchase. The evidence is unambiguous.

Without an evidence object, this kind of dispute takes weeks. The merchant absorbs the cost because nobody can reconstruct what happened. With it, the entire chain of intent, confirmation, and binding would be available on demand.

The dispute crisis we have been covering is not a technology problem. It is an evidence problem. Commitment governance is designed to solve it at the source.

What This Walkthrough Shows

That single grocery order illustrated five things current systems cannot do.

The framework would distinguish between payment-ready and fulfillment-ready. Authorization and binding become separate events with separate criteria. Current payment rails collapse them into one.

It would catch a scope expansion before it created binding consequence. The agent tried to add an unapproved merchant. The framework would flag it before any downstream party relied on the transaction.

It would require human confirmation for a novel action without blocking the rest of the order. Sarah confirmed the pasta substitution. The other 21 items moved forward without interruption.

It would produce a complete evidence trail purpose-built for dispute resolution. Not scattered logs across three systems. One object with every decision, every constraint check, every timestamp.

And it would resolve the dispute from that evidence object alone. No manual reconstruction. No merchant eating the cost because the proof did not exist.

The Counter-Scenario

Run the same transaction without commitment governance and consider what would likely happen.

Sarah's agent orders groceries. Two items are out of stock. The agent substitutes from a different merchant. Payment authorizes. Both merchants fulfill. Two weeks later, Sarah disputes the charge she does not recognize.

In most current systems, nobody could prove the agent was authorized to shop at GreenBasket. The authorization was a card network event. The substitution was an agent decision. The connection between them would live nowhere. The merchant could not produce evidence of consumer intent because no system captured it.

GreenBasket would absorb the $3.49 chargeback. The processing cost runs $25 to $50. Multiply that by millions of agent transactions per day and you have the dispute crisis we have been warning about: not a wave of fraud, but a structural inability to prove that legitimate transactions were properly governed.

The question is no longer whether agentic transactions need governance. How fast can the payments industry adopt it?

Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.