Past two weeks have been packed with signals that the agentic AI moment for supply chains is no longer approaching - it is here, and the economics of it are starting to bite. Let me reflect on a few that provoked my thinking.
0️⃣ CNBC: Tokens or Humans? The new corporate trade-off.
AI is turning out to be far more expensive than anyone expected, and CFOs at major companies are now facing a brutal new choice. Enterprise AI has moved through three distinct phases: boards demanding their CEOs do something about AI, then tokenmaxxing - using AI by any means regardless of cost - and now a third phase, where leadership teams are asking whether they actually need premium frontier models for every task. Roughly 95% of enterprise AI still runs on the most expensive models, even for tasks that do not require that level of intelligence.
Implications for supply chains. This is the frame for everything else in this edition. The CFO question - do we need frontier-model intelligence for this task? - is exactly what supply chain leaders should be asking about every agent workflow in their portfolio right now. A replenishment calculation, a standard PO approval, a track-and-trace query - none of these require the same model as a complex disruption scenario.
Organisations that build deliberate model tiering into their supply chain agent architecture now carry a structural cost advantage into 2027. Those that do not will fund the gap from headcount - which is where tokens or humans stops being a metaphor.
Blue Yonder announced a Model Training Factory built with NVIDIA - a repeatable system for fine-tuning specialised supply chain agents on NVIDIA's open Nemotron models rather than calling expensive frontier LLM APIs. The first models target warehouse management decisions: allocation shorts, inventory exception handling, due-time urgency, and yard and receiving-trailer inventory. CEO Duncan Angove's framing was deliberate: owned intelligence, not rented intelligence.
Implications for supply chains. This is the most directly relevant development in this edition, and the clearest practical answer to item 0. The rented intelligence model - calling a frontier API for every warehouse decision - is expensive, slow and exposes operational data to third-party infrastructure. Blue Yonder's approach inverts this: train compact domain-specific models on your own operational data, run them locally, and own the resulting capability. For a global manufacturer or 3PL running millions of daily warehouse decisions, the economics are transformative.
The model factory concept also carries a governance advantage - a fine-tuned domain model is interpretable in ways a black-box frontier API is not, and its decision boundaries can be audited with precision.
2️⃣ Coupa acquires Tonkean - the agentic procurement stack is now complete.
Coupa announced the acquisition of Tonkean, an AI-native intake and orchestration platform, giving Coupa fully integrated Agentic-as-a-Service capabilities across its network of more than 3,500 buyers and 10 million suppliers. Tonkean adds a natural language interface, a multi-agent orchestration framework, Agent-to-Agent coordination, and more than 250 native integrations. This is Coupa's fourth acquisition in rapid succession - following Cirtuo, Scoutbee and Rossum - all in service of a single stated ambition: the world's foremost agentic trade network.
Implications for supply chains. In the space of two weeks Coupa moved from a procurement platform with AI features to a vertically integrated agentic trade network. Rossum handles unstructured documents, Scoutbee handles supplier intelligence, Cirtuo handles category strategy - and Tonkean now routes work across all of them via natural language and 250 native connectors, without requiring a rip-and-replace of existing systems. A fully agent-orchestrated procurement cycle from intake request through sourcing, document processing, approval and payment is now available as a single-vendor deployment on a 10 trillion dollar spend dataset. The no-code process builder is particularly relevant for supply chain teams without dedicated engineering resources.
Figure AI announced a commercial agreement to deploy its humanoid robots at Catalyst Brands' distribution logistics centre in Reno, Nevada. Catalyst operates more than 1,800 stores across JCPenney, Aeropostale, Brooks Brothers, Lucky Brand and Nautica. The initial deployment focuses on assisting warehouse associates in the facility's sorting-system sequencing operation. Figure CEO Brett Adcock framed the thesis: humanoids provide a standardised labour solution deployable across diverse industries without facility redesign.
Implications for supply chains. The significance here is not the scale - robot counts and throughput figures have not been disclosed. It is the category of customer and operational context. A traditional brick-and-mortar retail holding company deploying humanoids into an existing, conventionally-designed DC, alongside human associates, in a high-variability fashion retail environment - that is a meaningful proof-of-concept for the drop-in humanoid model. For supply chain leaders modelling the medium-term labour picture, this is the moment to begin scenario planning rather than waiting for throughput data.
The question is not whether humanoid logistics automation will be viable - the Figure and Catalyst deal confirms it is past proof-of-concept. It is how quickly the cost-per-unit-handled will fall to the point where it changes the DC labour planning conversation.
Gartner warned that applying uniform governance across AI agents - treating all agents the same regardless of their autonomy level and decision scope - will lead to enterprise AI agent failure. The firm predicts 40% of enterprises will demote or decommission autonomous AI agents by 2027 due to governance gaps identified only after production incidents. The recommendation is proportional governance: classify agents by their autonomy level and assign controls that match - not a binary choice between locked-down and fully trusted.
Implications for supply chains. This is the governance warning every supply chain leader deploying agents in 2026 needs to read before going live, not after. The failure modes are predictable: a procurement agent with write access that books an unintended order, a replenishment agent that over-corrects on a demand signal and creates a stock spike, a freight-booking agent that selects a carrier it should not have because it was not constrained at the data layer. The practical implication is a tiered governance model built before deployment. Classify every supply chain agent by what it can do - read only, recommend, execute within bounds, execute freely - and assign matching controls: scoped API permissions, human-in-the-loop checkpoints, shadow-mode validation, named ownership, and audit logging. The governance architecture should exist before the first production transaction, not after the first incident.
5️⃣ Meta SAM 3 published at ICLR 2026: a model that sees what you describe - in real time.
SAM 3 is Meta's latest Segment Anything model, published at ICLR this week. Unlike previous versions that required clicks or bounding boxes, SAM 3 accepts plain-language concept prompts - "box", "damaged carton", "trailer without seal" - and returns segmentation masks for all matching instances in images and video, without fine-tuning. SAM 3.1, released in March, added multi-object tracking and runs on edge hardware, making near-real-time video processing viable without cloud infrastructure. Given that I am building my own home AI server and experimenting with local vision pipelines, I find this one particularly interesting.
Implications for supply chains. The supply chain relevance is in three properties that together remove the main barrier to computer vision adoption in physical operations: no fine-tuning required on your specific environment, plain-language prompts, and edge-deployable inference. A warehouse camera feed, a yard management camera, a cross-dock inbound dock - you describe what to detect and get continuous segmentation without a machine learning team maintaining a custom model per facility. The process analysis application is also direct. SAM 3 can be applied to screen recordings of ERP or WMS workflows - segmenting UI elements and interaction sequences - to automatically generate structured process maps for workflows that have never been formally documented. For supply chain teams running pre-automation audits before deploying agents, this compresses weeks of manual process mining into hours. Open-source, runs locally, no data leaves your environment.
6️⃣ 11,000 tech jobs eliminated in a single day - profitable companies trading payroll for compute.
On May 20, Meta began notifying roughly 8,000 employees that their positions were eliminated - about 10% of its workforce. The same morning, Intuit announced 3,000 additional cuts, 17% of its global staff. Combined: 11,000 tech jobs in a single Wednesday. Meta simultaneously moved 7,000 workers into four new AI-focused organisations and cancelled 6,000 open roles - reshaping close to 21,000 positions in a single week. What distinguishes this wave is not the numbers. It is that the companies executing the deepest cuts are simultaneously reporting their strongest-ever quarterly results and raising AI infrastructure budgets to figures that make human payroll look like a rounding error.
Implications for supply chains. One-third of surveyed organisations expect AI to reduce their workforce in the coming year, with anticipated cuts highest in service operations, supply chain, and software engineering. For supply chain leaders, the implication arrives in two forms. External - technology vendors, logistics service providers and 3PLs are undergoing the same restructuring, and service quality may deteriorate as headcount falls faster than agent capability matures. Internal - supply chain functions are named as high-displacement targets, and the real design challenge is not whether to automate but how to redesign operating models so the humans retained are doing work agents genuinely cannot do. Complex supplier negotiation, risk judgment under ambiguity, exception escalation, cross-functional coordination where context and relationships matter.
Organisations that treat this as a workforce redesign challenge will retain institutional knowledge. Those that treat it purely as a cost line will find out what they lost when the next major disruption hits.
7️⃣ Google Antigravity 2.0: orchestrate teams of autonomous agents with a single API call.
At Google I/O 2026 on May 19, Google repositioned Antigravity from a coding environment into a platform for developing and managing teams of autonomous AI agents - with a new desktop application, a CLI, an SDK, and managed execution in the Gemini API. The headline demo had a team of parallel agents build a working operating system core from scratch in twelve hours for under one thousand dollars in token cost. Powered by the simultaneously released Gemini 3.5 Flash, claimed to be four times faster than rival frontier models. Free public preview as of May 2026.
Implications for supply chains. Antigravity 2.0 matters for supply chain practitioners for a reason beyond coding: it is the clearest demonstration yet that orchestrating teams of specialised agents - each with a defined scope, running in parallel, coordinated toward a shared objective - is now a low-code capability accessible to operational teams. The pattern maps directly to supply chain exception management: one agent monitors inbound shipment data, a second runs compliance checks, a third drafts the supplier corrective action notice, a fourth updates the ERP record.
My top GitHub repos worth exploring for Supply Chain professionals in this issue:
microsoft/ai-agents-for-beginners - 12 structured lessons to get started building AI agents, with working code samples using Microsoft Agent Framework and Azure AI Foundry. If you have been meaning to move from reading about agents to actually building them, this is the most accessible on-ramp I have come across. Python-based, well documented, and directly relevant to the orchestration patterns we discuss in this newsletter.
facebookresearch/sam3 - Meta's open-source Segment Anything Model 3 from item 5 above. Model weights, fine-tuning code, and the SA-Co evaluation dataset. If you want to experiment with concept-based visual detection on your own warehouse or yard imagery without building a custom training pipeline, this is the starting point. Runs locally.
n8n-io/n8n - workflow automation platform with 400+ integrations and strong AI agent nodes, including native MCP support. The most practical no-code tool I have seen for connecting supply chain data sources, ERP APIs, and AI agents without a full engineering project. There is a pre-built supply chain risk monitoring workflow template that scores disruption risk by linking supplier and route data with live news signals - a useful starting point for anyone building an exception management layer.
See you in two weeks...