I came across an article on AI-Native Supply Chain Planning by Alex Pradhan on LinkedIn. I used to get regular briefings from Alex when she was the global product strategy leader at John Galt Solutions, and I was an industry analyst. I always respected her knowledge. Like me, she used to be an industry analyst covering the supply chain space. Now she has spun off and founded her own firm called Alchemy Advisory.

It is a good article, well worth reading. While I agree with most of it, I did not agree with everything. So, I wanted to interview her, push back in a few areas, and see whether I came around to her way of seeing things. I’ve always found this is a good way to learn.

“Over the past year I've talked with many supply chain planning technology vendors that are newer to the market,” Alex wrote, “and a growing number of them share something in common: they'd be classified as AI-native, carrying architectural characteristics that come from being born AI-native, rather than retrofitted onto it. Longstanding, traditional legacy vendors have been modernizing their stacks too, some quite aggressively, but there's a notable difference between architecture designed AI-first from day one and architecture that's evolved to accommodate AI after the fact.”

That was the first paragraph. I pushed back immediately. I pointed out that three or four years ago no one was talking about AI, not the incumbents, and not newer entrants. But some vendors, Manhattan Associates and o9 Solutions , came to mind; they were doing certain things that got them ready for AI. Manhattan, I thought, led the game in breaking their solution into components and creating microservices. A componentized architecture is necessary for agentic AI .

o9 built their solution with a knowledge graph, a key piece of technology for providing context around decision-making. Alex, in turn, pointed out that a knowledge graph might support supplier risk or a demand-planning hierarchy, yet still be too narrow to fully support supply chain planning. Just having a knowledge graph is not enough to check the box.

Technology Building Blocks to Be AI Native

In her article, Pradhan listed 6 main building blocks:

A unified data and context layer. A single, living model connecting data, business logic, decisions, and real-world relationships. Some vendors formalize this as an ontology —think of it as a blueprint defining objects, attributes, and relationships working alongside a knowledge graph that reflects live data. The goal is semantic consistency across the different AI models, and it's what helps you get real context and knowledge of your supply chain data.” This is where the knowledge graph matters, but the hard work defining objects also matters.

Composable and scalable architecture. A modular, flexible technology stack that can extend and adapt.” This is a componentized, microservices architecture.

But also to be AI-native, the solutions need:

Governed, autonomous agents. Multi-agent systems that reason, coordinate, act, and learn; with permissions, audit trails, human override, and awareness of process and physical constraints. Autonomy here is a progression, part of a broader continuum of decision automation.

Shared decision authority. Humans and AI agents share decision authority based on context, criticality, urgency, and other criteria. This helps to map where a decision should be AI-supported, where it needs a human-AI mix, and where full autonomy is appropriate.

Continuous learning and adaptation . Actions and outcomes feed back into the system so it improves and recalibrates continuously, without human intervention.

AI embedded in operations and workflow . Not an intelligence layer sitting on top, but structurally part of how decisions, execution, and workflows operate. “While still emerging, I am seeing use cases being targeted to integrate and orchestrate decision-making across functions, applications, platforms, and multi-parties.”

So, who is AI-native? Alex listed: Flowlity , Horizon Solutions , Pull Logic , Lyric , and Azirella .

Alex also said that, in her mind, the ability to continuously learn and adapt is particularly important to being AI-native. Demand planning systems have had machine learning capabilities for 25 years. But she means more than this.

“You're getting bombarded with all these alerts. There are different types of decisions that require different types of automation.” On the backend, a thorough typology for deciding which tasks can be automated needs to be in place.

The Importance of Context

Can an SCP solution ever fully incorporate the full context surrounding a decision? Take, for example, an inventory planning engine in which the context is a major new tariff announcement. Companies will immediately do a big forward buy. But the planning solution will look at the purchase and say, “Cancel it; you have more than enough inventory.” If agents are fully autonomous, they may proceed with this without a human in the loop to stop it.

Alex replied, “I don't think that everything is ever going to be fully autonomous. It doesn't make sense. Alex later wrote back with deeper thoughts on a decision criteria framework for agents.

Impact Criteria: As digital triggers/events hit the plan, they are evaluated and scored based on impact criteria such as urgency, radius, investment, relevance, and value. This score quickly triages the event: is it routine enough to be automated, does it need augmentation, or does it need to be escalated to a different stakeholder?

Assign Owner, Priority, and Responsibility: Based on the degree of impact, the system identifies key stakeholders, assesses potential options, and prioritizes.

Impact and complexity are two different questions, Pradhorn pointed out, and they don't always move together. A high-impact event isn't automatically a complex one, and a low-impact event can still be genuinely ambiguous.

What Method and Process to Use: This layer maps the method and degree of automation to the nature of the problem. This ensures we don’t over-build for simple decisions or under-build for ambiguous ones.

Pradhan points to the Cynefin Framework a useful lens for mapping the type of decision to the degree of automation based on how well cause and effect is understood. This framework categorizes decisions as simple, complicated, complex, and chaotic.

“This is also where the automated/manual/mixed agent decision-making is actually made — not purely by impact score, but by impact combined with how well understood the problem is,” she explained.

Decision Quality: Once a decision lands with a person, agent, or a defined workflow, that's where classic decision quality kicks in — is the problem framed right? Are there real alternative options? Is the information reliable? Are the trade-offs clear? Is the reasoning sound? Is there real commitment to follow through?

This is supported by a decision library that works in both directions. Before a decision is made, the library surfaces how similar events were previously framed, routed, and resolved — the known options, the data typically needed, and who historically owned the call — so that the prospective decision can route faster to the appropriate type of automation.

After a decision is made, the library captures the outcome and distinguishes between whether the decision itself was sound and whether the outcome was simply lucky or unlucky. This feedback loop improves the quality of judgment over time. Those learnings fold back into how future events are scored, routed, and reasoned through. I view this framework as useful and thought-provoking. But I would still err on the side of autonomous decisions mainly for low-impact events. The example I brought up on agentic inventory planning when new tariffs are introduced comes to mind. What seems like a decision that can be easily automated because it is “clear” can shift to “chaotic” in a heartbeat. And that transition event can be hard to capture because it was created outside the system.

In Alex’s framework, that would be scored as a poor-quality decision, and this decision might shift from autonomous to human-supervised. But over time, wouldn't more and more decisions move from automated to supervised?

Alex agrees that, in contrast to genAI, which is now table stakes for supply chain planning solutions, agentic AI is in far more of a pilot mode. “We're nowhere near multi-agent workflows where everything is connecting across the entire supply chain.”

I am keen to interview these AI-native SCP suppliers and their customers and learn more about what can really be achieved.