© Supply Chain Research 2026
Field Guide · AI
Almost everything written about AI in supply chain is written by someone selling it. This page is the other thing: an independent map of where AI is shipping in production, where it is early, and where the story is ahead of the software — across the vendors SCR covers.
Read it with the matrix legend in mind: ● means customers run it in production today, ◐ means early or partial, ○ means not a current focus. Every mark is an editorial judgement, not a vendor submission, and the right way to use any of it is to structure your own scripted demos.
Demand forecasting and planning ML is the most proven category: machine-learned models beating statistical baselines is a decade-old result, and platforms like Blue Yonder, Kinaxis, RELEX, ToolsGroup, and o9 run it in production at scale. The demo test is forecast value added on your history, item by item — not an aggregate accuracy chart.
Optimization at scale — routing, slotting, wave and task assignment, network design — predates the AI label and quietly delivers most of the money. Manhattan, Oracle OTM, and the robotics fleet orchestrators (Locus, Symbotic) are optimization businesses wearing an execution interface.
GenAI copilots earn their keep on investigation and configuration: “why is this order stuck,” “build me this report,” “draft the carrier dispute.” SAP’s Joule, Kinaxis Maestro, and Manhattan’s assistants are furthest along. Copilots fail the value test when they only summarize screens you could already read.
Agentic execution — software that takes actions, not just suggests them — is 2026’s loudest claim and thinnest evidence base. Real deployments exist (touchless exception resolution in freight audit and order management) but are narrow. Buy agents the way you would hire a new planner: limited authority, monitored decisions, expanding scope with track record.
Physical and vision AI is the warehouse robotics story: perception for picking and unloading, and inventory intelligence (Dexory, Simbe). Here the AI is inseparable from the hardware economics — evaluate units-per-hour and exception rates, not model names.
20 vendors, five capability areas. Legend: ● shipping in production · ◐ early or partial · ○ not a current focus.
| Vendor | ML forecasting & planning | Optimization at scale | GenAI copilot | Agentic execution | Physical & vision AI | SCR note |
|---|---|---|---|---|---|---|
| Manhattan Associates | ◐ | ● | ● | ◐ | ○ | Active-platform assistants and optimization are real; agentic execution is arriving feature by feature. |
| Blue Yonder | ● | ● | ◐ | ◐ | ○ | Deep ML planning heritage; the copilot story is newer than the forecasting one. |
| SAP (IBP / EWM + Joule) | ● | ● | ● | ◐ | ○ | Joule spreads a uniform copilot across the estate; planning ML is mature. |
| Oracle | ◐ | ● | ◐ | ◐ | ○ | OTM optimization is the standout; AI features track the broader Fusion cadence. |
| Kinaxis | ● | ● | ● | ◐ | ○ | Concurrency is itself the optimization story; Maestro’s AI layers onto it credibly. |
| o9 Solutions | ● | ● | ● | ◐ | ○ | The “digital brain” pitch is broad; insist on your data, not the demo graph. |
| RELEX Solutions | ● | ● | ◐ | ◐ | ○ | Retail/grocery ML forecasting with real production depth. |
| ToolsGroup | ● | ● | ◐ | ○ | ○ | Probabilistic forecasting pioneer; narrower surface, deep where it plays. |
| e2open | ◐ | ● | ◐ | ◐ | ○ | Network data is the asset; AI value scales with how much of your chain runs through it. |
| Pando | ◐ | ◐ | ● | ● | ○ | The most aggressive agentic-freight pitch in the field — test it on your exceptions, not theirs. |
| IBM Sterling | ◐ | ◐ | ◐ | ◐ | ○ | watsonx integration is advancing against a proven but older core. |
| Salesforce OMS | ◐ | ○ | ● | ● | ○ | Agentforce makes it the strongest CRM-side agent story; fulfillment depth is the check. |
| Infor | ◐ | ◐ | ◐ | ○ | ○ | Steady platform-level AI; strongest inside CloudSuite verticals. |
| Infios (Körber) | ◐ | ◐ | ◐ | ○ | ○ | Post-rebrand roadmap clarity is the thing to verify. |
| Siemens Opcenter | ◐ | ◐ | ◐ | ○ | ◐ | Industrial AI lives closer to the automation layer than the MES seat. |
| Rockwell (Plex) | ◐ | ◐ | ◐ | ○ | ◐ | Edge/plant-floor analytics are the differentiated part. |
| Tulip | ◐ | ○ | ● | ◐ | ◐ | Frontline copilots on composable apps — the most buyer-accessible GenAI in manufacturing. |
| Locus Robotics | ○ | ● | ◐ | ◐ | ● | Task optimization across fleets is production AI, even if nobody calls it that. |
| Symbotic | ○ | ● | ○ | ◐ | ● | Perception and dense-case orchestration at industrial scale. |
| Dexory | ◐ | ○ | ◐ | ○ | ● | Vision-first inventory intelligence; the AI is the product. |
Four places, in descending order of proof: demand forecasting (ML beating statistical baselines is well established), large-scale optimization (routing, slotting, task assignment), document and audit automation (freight audit, order capture), and physical AI in warehouse robotics (perception and fleet orchestration). GenAI copilots are genuinely useful for investigation and configuration work; fully agentic execution is where claims most outrun production evidence in 2026.
Three tests: run it on your data, not the vendor’s (a copilot that only answers on curated demo data is a prototype); ask for the production reference running the specific AI feature at your scale; and ask what happens when the model is wrong — real deployments have override workflows, monitoring, and rollback, marketing decks do not.
It should be weighted, not decisive. Platform fundamentals — data model, integration, scalability — determine whether any future agent can act reliably. A strong platform with an emerging agent story usually beats a weak platform with a loud one; SCR’s selection scorecard lets you weight this explicitly.
No. Supply Chain Research accepts no vendor compensation for coverage, ratings, or placement. The matrix is SCR’s independent editorial assessment, last reviewed August 2026, and is deliberately directional — it exists to structure demos, not replace them.
The matrix is SCR's independent editorial assessment of publicly evidenced capability, last reviewed August 2026, and will be revised as production evidence accumulates. Vendor ratings are directional, based on public analyst research, reference calls, and market experience (last reviewed August 2026). They are a shortlisting aid — always validate capabilities against your requirements in scripted demos and references. Supply Chain Research accepts no vendor compensation for coverage, ratings, or placement.
