Kinaxis Runs Its Supply Chain as One Model, Not Six Silos — and Scenario Modeling Is Up 120%

Kinaxis, the Ottawa-founded supply chain software company, has moved its Maestro platform out of pilot and into daily use for customers who can no longer plan demand, inventory, logistics, and production as separate problems. Built on Microsoft Azure and Azure OpenAI, Maestro combines predictive AI, scenario modeling, and agentic AI to treat the entire network as what CEO Razat Gaurav calls “one single model” — machine learning tuned per customer generates demand forecasts, autonomous agents flag emerging issues and propose fixes, and humans keep final decision authority. Microsoft reports customers have ramped scenario-modeling activity more than 120% amid recent tariff and geopolitical instability, a usage jump that tracks real-world volatility rather than a product launch curve.

That kind of compounding adoption once a system proves itself operationally shows up elsewhere, too. Inside OpenAI’s own research organization, coding-agent usage went from a novelty to the dominant mode of work in a matter of months: by mid-August the median researcher was running agents at more than $600 a day in inference cost, and the org logged 3.1 “agent-workdays” of effort for every human workday — up from a ratio where human labor still exceeded agent runtime as recently as June.

The harder question is what keeps that compounding safe. Mistral’s own case study on migrating 40,000 lines of untested Fortran-77 to C++ found the honest failure mode on both ends: full agent autonomy produced code that was merely “Fortran retyped in C++,” while pure manual migration stalled on the hard bugs. The fix wasn’t more autonomy or less — it was structured human review gates at defined checkpoints, the same design choice Kinaxis makes by keeping a human as the final call on every agent-flagged supply chain decision. For operations leaders scaling agentic AI into functions that can’t absorb a bad autonomous call, the pattern holds across a supply-chain platform, a frontier lab’s internal R&D, and a legacy-code migration alike: the checkpoint is the feature, not the bottleneck.