AI Can Execute Transactions -- Lu Zhang Is Focused on Whether Actually Should Be

As commerce moves from AI-assisted tools toward AI-executed decisions, Zhang is focused on the rules that determine when an automated commercial action should proceed.

A payment can clear even when fulfillment is impossible. A market can be technically reachable while the economics make entering it a bad decision. Lu Zhang has spent much of her career working inside those gaps, where commerce systems can execute individual actions without necessarily knowing whether the larger transaction makes sense. Now a product leader developing AI automation and governance systems for global commerce, Zhang believes that problem will become more consequential as software moves from helping people make decisions to acting on their behalf.

“Commerce is getting much better at execution,” Zhang said. “The harder question is whether the system has enough information to know that execution should happen at all.” Her work centers on transaction integrity, merchant profitability, and the operational conditions required to complete a commercial promise. Those concerns have become increasingly connected as AI begins entering purchasing, payments, merchant operations, and other workflows that previously required more direct human involvement.

Zhang sees a growing mismatch between the speed of automation and the quality of the decisions underneath it. A payment system may know that a charge can be processed, but that tells it little about whether inventory is available, whether fulfillment is viable, or whether the merchant can complete the order profitably. In cross-border commerce, additional conditions such as serviceability, customs, logistics, and compliance can turn a seemingly valid transaction into an expensive failure after execution has already begun. Zhang supports greater automation, but she questions systems that treat technical permission as sufficient evidence that a commercial action should proceed.

“Execution speed is increasing faster than decision quality,” Zhang said. “You can automate a bad decision just as efficiently as a good one.” That distinction has shaped Zhang’s work across several layers of commerce infrastructure. At Easyship, she served as Product Lead for the international shipping partnership with eBay, owning the roadmap, integration, global rollout, and live optimization of cross-border fulfillment infrastructure supporting more than 9 million annual international shipments across more than 200 countries. Later, at TikTok Shop, she worked on core transaction architecture connecting pricing, payment, merchant conditions, fulfillment, policy, and compliance while supporting new commercial models.

Those environments exposed the same underlying problem in different forms: a transaction is rarely one decision. It is the result of several conditions becoming true at the same time, and Zhang argues that automated commerce needs a way to evaluate those conditions before committing resources or creating obligations downstream. Her focus is on the point where technical capability, business logic, and operational readiness have to converge.

Zhang describes her preferred model as bounded automation. AI should have greater freedom when the conditions are clear and the consequences are limited, while more consequential actions require stronger controls, including explicit rules, confidence thresholds, human escalation, and a record of why the system was allowed to proceed. “I do not think the goal should be maximum autonomy,” Zhang said. “The goal should be appropriate autonomy. A system should know when it has enough confidence to act and when the ambiguity is telling it to stop.”

That approach treats governance as part of the product architecture rather than a policy document sitting outside the transaction itself. Zhang argues that a model may be useful for reasoning through incomplete information, while some decisions still require deterministic boundaries. Money, policy constraints, compliance requirements, and high-impact operating decisions may need rules that cannot simply be overridden by a probabilistic recommendation.

That is why Zhang advises product teams to begin with the decision rather than the technology. Before asking what AI can automate, she believes they should define what must be true before the system acts, which conditions cannot be overridden, who has authority when the information is incomplete, and whether the final decision can later be reconstructed. The goal is to make the logic behind execution explicit enough that both people and systems can understand why an action was allowed to move forward.

Zhang also cautions against optimizing individual parts of the commerce chain without examining the complete result. Conversion can improve while margins deteriorate, and payment success can rise even as fulfillment problems create refunds or operational recovery work. Faster shipping decisions are of little value if warehouse capacity or customs constraints make the commitment impossible to honor. “Commerce systems are often measured one function at a time,” Zhang said. “The customer and merchant experience the result of all of those functions together.”

Her systems-level approach connects transaction success to merchant economics and fulfillment feasibility rather than treating each metric as an independent measure of performance. The cost of a poor decision can also grow after execution begins. Zhang points to rework, returns, detention, disputes, and manual recovery as downstream consequences of weak upstream decisions. At platform scale, dependencies that were manageable in smaller systems become harder to ignore because pricing, risk, payment, logistics, merchant configuration, and fulfillment can no longer be treated as isolated domains.

Zhang has also developed these ideas through the professional supply-chain community. She holds the Certified Professional in Supply Management, or CPSM, credential and has published three practitioner articles in Inside Supply Management®, the professional publication of the Institute for Supply Management. Her articles examine executable governance in cross-border supply chains, operational workarounds as signs of missing system controls, and the redesign required when autonomous AI agents begin initiating commercial transactions.

Founded in 1915, ISM is the world’s first professional association for supply chain and serves a community of more than 200,000 professionals across more than 100 countries. Its Manufacturing and Services PMI reports are closely followed by business leaders, policymakers, and financial markets as indicators of U.S. economic activity. Publication through ISM places Zhang’s work within an established professional discussion about supply-chain operations, governance, and the future of automated commerce.

Zhang has formalized the recurring decision logic behind this work in the Commitment Decision Framework for Transaction Systems. The framework addresses what must be true before a commercial action proceeds, where automated authority should stop, and what decision record should remain when the action creates financial or operational consequences. Its application to agentic commerce has also been discussed in a three-part editorial series published by Major Matters.

For Zhang, the central challenge facing AI-driven commerce is deciding when automated systems should move forward without creating avoidable financial, operational, or compliance consequences. More systems will gain the technical ability to act without waiting for a person to complete every step, but speed alone does not make the underlying decision sound. “Greater automation should give a business more control over good decisions, not less control over consequential ones,” Zhang said. “The system should make the conditions for action clearer.” Her work is focused on building that decision logic into commerce before execution begins.