6 April 2026

RELEX "State of Supply Chain 2026": AI Moves Into Core Decisions as Tariffs Bite

A survey of 514 retail, manufacturing, and supply chain leaders (January 2026) finds 67% report increased confidence in AI for supply chain decisions year-on-year. Only 10% trust fully autonomous AI — 54% prefer AI-recommends, human-decides. 86% report trade policy/tariff impacts on operations. 47% are using or planning AI-driven inventory optimisation; 41% logistics routing. 71% plan to invest in generative and agentic AI over the next three to five years.

Why it matters: The "AI recommends, human decides" majority tells practitioners exactly what architecture to build: explainable recommendations, human-in-the-loop approval workflows, and full audit trails. With tariff volatility driving demand for rapid scenario replanning, AI-enabled replanning velocity is becoming a structural competitive differentiator.

Palantir AIP Analyst: Conversational Decision Intelligence Grounded in the Ontology

AIP Analyst reaches general availability the week of April 13, 2026 for all AIP-enabled users. It is a conversational AI interface that lets both technical and non-technical users interrogate their operational ontology in natural language. Responses are backed by SQL queries, Vega chart visualisations, and Foundry function calls. Every answer exposes an interactive dependency graph tracing the chain from question to answer — providing full auditability of each analytical step.

Why it matters: Grounding conversational AI in a semantically modelled operational ontology prevents hallucinated decisions. The dependency-graph transparency layer is a practical answer to enterprise explainability requirements and directly enables the kind of trusted, auditable AI that the RELEX survey majority demands before ceding decision authority.

Capgemini × Microsoft Unveil CargoPilot Agent at Hannover Messe 2026

Capgemini is launching an end-to-end agentic supply chain offering built on Microsoft's Work IQ, Foundry IQ, and Fabric IQ intelligence stack at Hannover Messe 2026. The CargoPilot Agent continuously analyses transport modes, routes, cost structures, carbon impact, and cycle times in real time, generating optimised shipment recommendations that balance speed, sustainability, and cost. The offering spans demand planning, inventory, and customer service across the full enterprise context.

Why it matters: Multi-objective transport optimisation — balancing cost, carbon, and speed simultaneously — is one of the hardest real-world OR problems. Embedding it inside an agentic continuous-monitoring workflow turns a periodic planning exercise into an always-on decision engine, with carbon now treated as an operational constraint rather than a reporting afterthought.

Gurobi 13.0: PDHG GPU-Accelerated LP, 16% Faster MIP, and Expanded MINLP

Gurobi 13.0 (November 2025, updated to 13.0.1 January 2026) introduces a Primal-Dual Hybrid Gradient (PDHG) LP solver with CPU and GPU support — enabling rapid solution of very large-scale LPs that exceed practical limits for simplex methods. Core MIP performance improves approximately 16% on difficult instances. The MINLP solver achieves more than 2× speedup on non-trivial nonlinear problems. The matrix-friendly gurobipy API receives significant enhancements for Python-first workflows.

Why it matters: PDHG with GPU acceleration brings Gurobi into direct competition with cuOpt for large-scale LP workloads, extending GPU-native solving beyond the VRP domain. For practitioners, a single commercial solver now covers GPU-accelerated LP, cutting-edge MIP, and improved MINLP — eliminating the need to stitch together separate solver stacks for different problem types in the same pipeline.

Execution-Verified Reinforcement Learning for Optimization Modeling (EVOM)

Anonymous authors — arXiv preprint, April 2026

EVOM treats a mathematical programming solver as a deterministic, interactive verifier for RL training of LLMs. Given a natural-language optimisation problem and target solver, the model generates solver-specific code; the code is executed in a sandboxed harness; and execution outcomes (feasibility, optimality, error traces) are converted into scalar rewards optimised via GRPO and DAPO in a closed-loop generate–execute–feedback–update cycle. Cross-solver generalisation is achieved simply by switching the verification environment.

What problem it solves: Existing LLM-for-OR approaches either rely on costly closed-source agentic pipelines with high inference latency, or overfit fine-tuned models to a single solver API. EVOM's outcome-only reward formulation removes the need for process-level supervision and enables generalisation across solvers.

Why it matters: Using solver execution as a verifiable oracle is an elegant extension of the RLVR (RL with Verifiable Rewards) paradigm proven in mathematical reasoning. It creates a self-improving loop — better model → better code → solver confirms correctness → stronger signal — and is the training methodology most likely to underpin the next generation of open LLM+OR models.

Scenario Theory for Multi-Criteria Data-Driven Decision Making

Simone Garatti, Lucrezia Manieri, Alessandro Falsone, Algo Carè, Marco C. Campi, Maria Prandini — Politecnico di Milano & Università degli Studi di Brescia

The scenario approach provides a powerful data-driven framework for designing solutions under uncertainty with rigorous probabilistic robustness guarantees. This paper extends it to multi-criteria settings where multiple datasets and simultaneous appropriateness criteria must be assessed jointly. The central innovation is a collective treatment of cross-criteria risks that yields substantially tighter robustness certificates than per-criterion analysis, covering multi-agent decision problems with heterogeneous data sources.

What problem it solves: Real-world decisions must satisfy multiple conflicting criteria simultaneously. Existing robust optimisation frameworks either handle one criterion at a time or produce overly conservative certificates; this work delivers tight guarantees for the full multi-criteria regime.

Why it matters: As enterprise AI moves toward multi-objective decision support (cost + carbon + speed, as in CargoPilot), the theoretical foundations for certifying solution quality under data uncertainty become essential. This paper provides those foundations for the multi-criteria regime that practitioners are actually building for.

Stein Variational Uncertainty-Adaptive Model Predictive Control

Efstathios Iliakis, Wallace Gian Yion Tan, Liang Wu, Jan Drgona, Richard D. Braatz — submitted to CDC 2026

This paper proposes an MPC framework using Stein Variational Gradient Descent (SVGD) to maintain a particle-based representation of system uncertainty, adapting control decisions in real time as uncertainty estimates evolve. Constraint tightening and objective weighting automatically respond to the current epistemic state of the system model, enabling more informed optimisation under non-Gaussian, multi-modal uncertainty distributions that standard stochastic MPC cannot represent.

What problem it solves: Standard stochastic MPC assumes Gaussian uncertainty or fixed scenario sets; neither captures the complex, asymmetric uncertainty profiles common in real supply chains, logistics, and energy systems. SVGD-based adaptation gives the controller a richer, dynamically updated model of what it does not know.

Why it matters: As agentic decision loops operate in increasingly volatile environments — the tariff-disrupted supply chains described by RELEX — uncertainty-adaptive control frameworks provide principled machinery for decisions that are robust to what is genuinely unknown at decision time, not merely to what was anticipated at design time.

For practitioners: Build verifiability into your decision AI stack now — audit trails, dependency graphs, solver-validated outputs. The 54% majority that demands human-in-the-loop is your near-term market; the transparent, explainable systems you build for them will be the trusted infrastructure when autonomy thresholds rise.