Enterprise software can report success one service at a time while the business outcome still fails. A catalog record is created. A pricing call returns. A promotion is configured. Yet the product may remain unavailable because a regional tax rule, an incomplete price, or a failed eligibility check sits between those individually successful steps.
That gap between technical completion and commercial readiness was the focus of Amit Kumar Padhy’s presentation during the virtual program of the 10th Toronto Machine Learning Summit. Padhy, who works as a Senior Computer Scientist II and Lead Architect at Adobe, presented “Beyond ‘System Complete’: Shipping Globally with Agentic Commerce Orchestration” on June 16, 2026. The session took place from 3:35 p.m. to 4:05 p.m. in the summit’s Technical / Engineering Talks track, under Enterprise Adoption / Team Design.
The talk places agentic AI inside one of enterprise technology’s least forgiving environments: distributed commerce systems responsible for coordinating pricing, catalog, promotions, billing, taxation and compliance across markets. In such systems, intelligence is useful only when it operates within clear technical and regulatory boundaries.
When “system complete” is not business complete
Padhy’s session begins with a problem familiar to engineers responsible for large commerce platforms: a workflow can finish without producing the intended business result. A product may exist in a catalog while its pricing remains incomplete. A promotion may be configured but fail to qualify. A tax dependency or compliance rule may prevent an offer from being made available in several regions.
From the perspective of individual services, each expected task may appear complete. From the perspective of the business, however, the product still cannot be sold.
This distinction forms the central argument of Padhy’s session. Distributed commerce platforms, he proposed, should not be optimized merely for workflow completion. They should be designed for business-outcome convergence: reaching a verified state in which a product is genuinely ready to be sold in its intended markets.
The change in perspective is significant. Traditional workflow engines generally execute a predefined sequence and report whether each step succeeded. Global commerce operations are less predictable. Pricing, billing, catalog, promotions, taxation and compliance services can produce conflicting or incomplete states, while market conditions may vary by geography, customer segment or commercial program.
Padhy’s presentation explored whether a coordinated system of specialized agents could respond to those conditions without turning a general-purpose language model into an unrestricted controller of the commerce platform.
A coordinated model, not a single autonomous agent
The architecture presented by Padhy, divides responsibility among multiple agent categories, each with a bounded role. Planner agents use ReAct-style reasoning to translate an onboarding objective into a dynamic execution graph. Domain agents invoke live enterprise capabilities, including pricing runtime, billing-preview and tax services. Validator agents enforce pricing and regulatory integrity, while a coordinator maintains shared state through a blackboard-pattern memory layer connected to Kafka-backed events.
The important design choice is separation. Planning, execution, validation and coordination are treated as distinct responsibilities. The language model helps interpret goals and adapt plans, but established domain services continue to perform the underlying business operations.
The presentation and depth of the topic reflect the professional background of Padhy, his work in cloud-native platforms, distributed systems and AI-enabled digital commerce. It also highlights his expertise in architecture and modernization of mission-critical, event-driven microservices, with attention to reliability, performance, cost optimization and platform governance.
The proposed orchestration model is also positioned as a layer over existing systems rather than a demand to replace them. For organizations operating mature infrastructure, this may be one of the presentation’s most practical ideas. Enterprise platforms cannot ordinarily be rewritten simply because a new AI architecture becomes available. New capabilities must coexist with existing APIs, contracts, controls, ownership models and operational responsibilities.
Keeping language models within clear boundaries
A notable aspect of Padhy’s session is its attention to where large language models should not be used. He argues invoking an LLM at every execution step introduced unnecessary latency and cost. The design therefore narrows model use primarily to planning and exception handling, while routine service execution remains deterministic.
The same principle applies more strongly at regulatory boundaries. Tax treatment, pricing integrity and compliance decisions cannot depend exclusively on probabilistic reasoning. Padhy’s model places deterministic, rule-based validators around these areas, creating safeguards and fallback mechanisms where repeatability and auditability are essential.
The talk also identified centralized orchestration as a bottleneck and described a move toward loosely coupled, domain-specific agents. That shift matters because a single intelligent controller can become both a performance constraint and an operational point of concentration. Distributed ownership allows agents to remain closer to the domains they understand while the coordinator maintains the shared business objective.
Together, these choices present a measured vision of enterprise agentic AI. Autonomy is introduced where interpretation and adaptation are useful, but constrained where correctness, evidence and regulatory consistency are non-negotiable.
A practitioner-oriented contribution to enterprise AI
The Toronto Machine Learning Summit is a practitioner-focused community centered on applied AI, AI infrastructure and enterprise adoption. It brings together more than 500 attendees, including machine-learning and AI professionals, technical practitioners and leaders working directly with production AI systems.
Within this setting, Amit Kumar Padhy’s presentation offered a practical perspective on autonomous agents. Rather than proposing that agents replace enterprise architecture, deterministic controls or domain ownership, it described a bounded coordination model in which AI supports planning, adaptation and exception management while established services continue to execute core business operations.
In this model, validators protect sensitive regulatory and pricing boundaries, compensation mechanisms enable recovery from distributed failures, and auditable decision traces provide greater accountability and operational visibility.
The presentation’s broader message is that the success of enterprise agentic AI will not be determined solely by how intelligently an individual agent can reason. It will also depend on how reliably multiple agents collaborate with existing systems, how carefully their authority is constrained and whether their decisions can be understood and reconstructed when something goes wrong.
For global commerce platforms, being “system complete” is not enough. The meaningful measure is whether the entire system has produced a correct, compliant, explainable and commercially usable outcome.
