The Agent Is Not the Architecture: Four Competing Models for Agentic Supply Chain Planning
Supply chain planning software vendors have reached an unusual point of apparent agreement. o9 has APEX. Kinaxis has Maestro Agents. OMP has UnisonIQ and Decision-Centric Planning. Anaplan is positioning itself as decision infrastructure for the Agentic Enterprise.
Read the announcements quickly and the strategies can look similar: agents sense changes, analyze situations, run scenarios, recommend actions, automate routine decisions and leave higher-value judgment to humans. They are not the same strategy.
The important difference is not the language model, assistant interface or number of agents. It is where each vendor believes decision intelligence should live, what representation of the enterprise gives an agent its context, which planning engine determines feasibility, and where human accountability enters the loop.
This matters because Supply Chain is an unforgiving environment for agentic AI. A plausible answer is not necessarily a feasible plan. A locally optimal recommendation can damage service, capacity or cash elsewhere. An agent that understands a planner's question but not the constraints behind the plan is an intelligent interface, not yet an intelligent decision system.
Gartner's 2026 Critical Capabilities framework is revealing because AI Planning/Decision Automation sits alongside constraint-based supply planning, scenario and financial impact, workflow orchestration, unified data integration and decision-centric planning. AI does not replace the planning architecture; it has to operate through it. McKinsey frames the same problem at operating-model level: which decisions should be automated, augmented or escalated?
Four vendors, four architectural bets
o9: the agent as part of an enterprise operating model. o9's APEX model—Agile, Adaptive, Autonomous Planning & Execution—is positioned as an operating model in which the Digital Brain connects decisions across functions and planning horizons. Its Enterprise Knowledge Graph provides a machine-readable representation of relationships, rules and institutional knowledge, while its neuro-symbolic direction combines language and statistical AI with symbolic enterprise representations. The architectural bet is clear: make enterprise knowledge machine-readable, then give agents access to it. The condition for value is equally clear: the knowledge graph must reflect how the enterprise actually makes decisions. Otherwise AI can reason elegantly over an outdated operating model. The key question is who maintains the decision knowledge when policies, constraints and decision rights change.
Kinaxis: the agent inside the live planning model. Kinaxis starts from concurrency. Maestro Agents operate inside the concurrent planning environment that connects demand, inventory, capacity, production and other constraints. Its emerging hierarchy spans orchestrator, process, task and tool agents, with Agent Studio allowing customers to compose additional agents. Where o9 emphasizes enterprise knowledge representation, Kinaxis emphasizes the continuously synchronized planning state. That reduces the gap between linguistic reasoning and planning feasibility: an agent can evaluate a disruption against the same model used to calculate consequences. The trade-off is architectural dependence. Agent quality depends on how completely the relevant decision and constraints are represented inside Maestro. The key question is whether the agent acts on the complete decision model or only the portion of the enterprise currently modeled there.
OMP: the agent as an orchestrator of decisions and optimization. OMP begins with the decision. Decision-Centric Planning aims to continuously identify decision situations, prepare scenarios and trigger action rather than waiting for a periodic planning cycle. UnisonIQ connects agents and generative AI with established optimization engines. This suggests a durable division of labor: LLMs interpret and orchestrate; agents monitor and coordinate; optimization engines solve; scenario models expose trade-offs; humans validate consequential decisions. The risk is decision proliferation. If agents discover more situations and optimization engines generate hundreds of alternatives, decision velocity only improves if the architecture also knows which situations to ignore, automate or escalate. The key question is whether agentic planning reduces the number of decisions humans need to make rather than merely producing better analysis for more decisions.
Anaplan: agents on deterministic decision infrastructure. Anaplan frames its platform as decision infrastructure for the Agentic Enterprise, connecting role-based AI agents across Supply Chain, Finance, Sales and Workforce Planning. Its proposition combines deterministic enterprise calculations with probabilistic AI, connected data and workflows. This creates an important distinction for high-stakes planning: probabilistic intelligence can detect or estimate a change, while deterministic models preserve business relationships, constraints and financial identities when evaluating consequences. Anaplan's cross-functional scope is strategically relevant because inventory, capacity and S&OP decisions are simultaneously operational and financial. The key test is where connected enterprise planning stops and detailed Supply Chain constraint logic begins.
Context is becoming the real battleground
All four vendors converge on one principle: a useful Supply Chain agent needs context. But context means something different in each architecture. For o9 it is heavily expressed through the Enterprise Knowledge Graph and Digital Brain. For Kinaxis it is the current state of the concurrent end-to-end planning model. For OMP it is the decision situation, scenario space and specialized optimization machinery. For Anaplan it is the connected deterministic enterprise model enriched by probabilistic AI and workflows.
That difference may become one of the most important architecture choices in next-generation APS. Agents will become easier to build. Foundation models will improve. Natural-language interfaces will commoditize. What remains difficult is encoding the physics of the decision: dependencies, constraints, objectives, trade-offs, policies, assumptions, financial consequences and organizational authority.
This is where mature planning vendors have an advantage over generic enterprise AI platforms. They already possess years—sometimes decades—of Supply Chain decision logic. The race is therefore not simply to put AI into APS. It is to determine whether existing APS decision logic becomes the substrate for agents, one tool used by agents, or is reorganized around a new agentic operating model.
The autonomy trap
Enterprise enthusiasm is running ahead of enterprise readiness. McKinsey's Supply Chain research found that while many companies were planning, blueprinting or piloting AI use cases, only a minority were deploying AI tools at scale. Demand forecasting, inventory optimization and supply planning remain natural high-value targets. The gap between experimentation and scale should make buyers skeptical of autonomy claims.
The hardest step is not moving from chatbot to agent. It is moving from recommendation to delegated decision rights. An agent that summarizes exceptions creates limited organizational risk. An agent that reallocates supply creates more. An agent that commits inventory, changes production priorities or triggers procurement crosses into a different category.
At that point architecture must answer operational questions: What objective is being optimized? Which constraints are hard? Whose policy wins when service, working capital and margin conflict? What confidence permits autonomous action? When must the agent escalate? Who owns the outcome? Can the decision be reconstructed months later? These are not generic AI-governance questions. They are planning-system design questions.
What Supply Chain leaders should actually compare
1. What gives the agent its world model? A knowledge graph, concurrent planning model, optimization model and multidimensional enterprise model are not interchangeable.
2. Can the agent distinguish plausible from feasible? Native access to constraints and optimization becomes critical once recommendations affect real supply, capacity or inventory.
3. Does the agent interrogate the plan or change the decision workflow? The larger opportunity is eliminating unnecessary analysis and coordination, not making dashboards conversational.
4. What learns? Buyers should distinguish improvement of an ML model from learning across forecast error, overrides, scenario choices and realized outcomes.
5. Where is authority encoded? Human-in-the-loop cannot mean placing an approval button after every recommendation. Mature autonomy requires explicit rules for what agents may decide, execute and escalate.
6. Can decisions cross functional boundaries? High-value Supply Chain decisions are often simultaneously commercial, operational and financial.
Dataleo perspective: AI will commoditize intelligence before it commoditizes decision architecture
The important development is not that APS vendors now have agents. It is that agents are forcing vendors to expose what their platforms really believe a decision is.
o9 emphasizes enterprise knowledge and a learning operating model. Kinaxis emphasizes concurrent awareness of end-to-end consequences. OMP emphasizes decision situations, scenarios and orchestration of specialized intelligence. Anaplan emphasizes trusted deterministic enterprise models augmented by probabilistic intelligence and role-based agents.
None is intrinsically the universal winner. They solve different parts of the same problem. But they expose a useful rule for buyers: do not evaluate the intelligence of the agent before evaluating the decision architecture underneath it.
The generative layer will improve rapidly and increasingly become interchangeable. The hard problem remains the layer that knows whether a recommendation is feasible, valuable, authorized and consistent with the rest of the enterprise.
The future of Supply Chain AI may therefore look less like replacing APS with agents and more like turning the APS into an executable decision model that agents can safely operate.
The decisive competitive question may ultimately be simple: when the agent is wrong, what part of the architecture knows it?
Sources and further reading
Primary vendor research: o9 APEX; o9 AI architecture; Kinaxis Maestro Agents; Kinaxis Agent Studio coverage; OMP Decision-Centric Planning; OMP agentic AI analysis; Anaplan Agentic Enterprise. Independent context: Gartner Critical Capabilities research; McKinsey Supply Chain risk survey; McKinsey on agentic operating models. Related Radar coverage: o9 Solutions, Kinaxis, OMP, Anaplan, agentic AI.