Physical AI's Missing Middle: The Decision Layer Between Sensors and Robots

On a Siemens blueprint factory floor in Erlangen this week, a wheeled humanoid — the HMND 01, running NVIDIA's Jetson Thor edge module — executed an autonomous logistics run that would have been a research demo twelve months ago. Down the hall at Hannover Messe 2026, which closes today, SAP detailed a Q2–Q3 general availability schedule for agents that orchestrate production master data, material reservations, field service dispatch, and asset health. On the same day, Infor launched its Agentic Orchestrator into limited availability. The symmetry is hard to miss: after a decade of AI confined to information systems, the agent has stepped onto the shop floor.

The Week Physical AI Became Operational

Hannover Messe 2026 was not a product launch. It was a convergence. NVIDIA and partners including Siemens, SAP, ABB, Dassault Systèmes, Microsoft, and Hexagon staged AI-driven manufacturing at factory scale — Hexagon's AEON system, for example, is slated for assembly operations at BMW's Leipzig plant. Humanoid's simulation-first pipeline compressed typical hardware development from roughly two years to seven months. Valeo, running Gemini across 100,000 employees, reports that 35% of its code is now AI-generated or optimized. In parallel, SAP's Hannover announcements push ERP execution toward the edge of operations rather than the center of IT.

The supply chain numbers are catching up. AI-driven demand forecasting cuts forecast error by 20–50% relative to traditional methods, according to McKinsey, with lost-sales reductions reaching 65%. U.S. distribution operators report warehouse throughput gains of 30–50% when AI and robotics are integrated. Autonomous mobile robots grew from 14% warehouse adoption in 2022 to 32% in 2026. Yet McKinsey's own State of AI finds that while 79% of organizations are experimenting with generative AI, fewer than 10% have scaled agents — and roughly 40% report EBIT impact, most of it under 5%.

A Three-Layer Lens on Industrial AI

Physical AI stacks cleanly into three layers. The sensing layer captures reality: IoT telemetry, vision systems, MES and SCADA streams, order books, and condition monitoring. It is widely deployed and, for most industrial operators, already a commodity. The acting layer executes in the world: robotic picking, humanoid logistics, AGV fleets, field-service dispatch, automated machine-tool routines. It is scaling rapidly, pushed by the NVIDIA–Siemens–SAP triangle. The decision layer sits between them — the operational intelligence that translates noisy signal into coherent, auditable intent and binds it to downstream execution.

Most industrial enterprises have invested heavily in sensing and acting. The middle is thin. Without a mature decision layer, sensors produce alerts that nobody prioritizes, and robots execute plans that ignore upstream disruption. It is the industrial equivalent of measuring blood pressure and performing surgery while skipping the diagnosis.

What the Decision Layer Actually Does

Consider four concrete decision loops now in production or near-production. A production planning agent reconciles live MRP signals with bill-of-materials changes and capacity constraints, recommending order-release adjustments within minutes rather than shift cycles. A field service dispatch agent re-sequences technician routes in response to real-time asset health telemetry, prioritizing interventions by revenue impact rather than ticket age. An asset health triage agent aggregates vibration, thermal, and cycle-count data to predict failure windows, routing the output to maintenance planners and spare-parts inventory. A warehouse demand-forecasting agent closes the loop between slow-moving SKU signals and pick-face replenishment, compressing forecast error by double digits.

None of these are novel in isolation. What is new is that vendors are shipping them as coordinated agents rather than point tools — and expecting enterprises to operate them. That expectation exposes the weakness.

From Telemetry to Coherent Action

Implementation is where POC theater dies. A working decision layer requires a semantic data layer that unifies product, plant, order, and asset ontologies across MES, ERP, WMS, and IoT sources. It requires model routing — small models at the edge for millisecond telemetry, frontier models for multi-constraint planning — rather than a single foundation-model call for every decision. It requires a decision ledger: which data, which model, which policy, which action, which outcome. And it requires human-in-the-loop gates on decisions that carry safety, financial, or reputational risk. Interoperability or it doesn't scale: without MCP-style tool access and A2A-style agent coordination, each use case becomes an integration project that never amortizes.

Where Deployments Break: Governance and Accountability

The EU AI Act's high-risk obligations under Article 6 take full effect on August 2, 2026. Many industrial AI use cases — safety-critical dispatching, predictive maintenance on regulated assets, worker-adjacent robotics — will fall inside scope. Conformity assessments typically run six to twelve months; organizations with fewer than five months of runway need to move now. Beyond the regulatory surface, three governance questions tend to go unanswered in physical-AI pilots: who owns the decision when an agent recommends an action, who is liable when an autonomous sequence misfires, and how is the decision reconstructed months later for audit. If an enterprise cannot answer all three in writing, it is not ready to move from advisory mode to autonomous mode.

KPIs Before APIs

Operational AI is measured at the loop, not at the model. Five KPIs tend to separate production deployments from showcase installations. Forecast error percentage against a frozen baseline, with a 25–40% reduction as a defensible target. On-time-in-full (OTIF) delivery rate, isolated to the decision loops the agent touches. Mean time to repair (MTTR) and first-time-fix rate, for asset health and field service use cases. Automation rate of qualifying operational decisions — the share of decisions made without escalation. And the decision auditability ratio: the share of decisions traceable to a specific data input, model version, and governing policy. A human override rate that climbs over time is an early indicator of drift or misaligned policy, not a failure of adoption.

From Pilot to Policy: A Twelve-Month Sequence

In the first ninety days, map the decision-sensor-actor triangles already present in operations, identify two high-frequency loops with clear measurement, and stand up a semantic layer that unifies at minimum three source systems. From day 90 to 180, deploy decision agents in advisory mode with mandatory human approval, instrument the baseline KPIs, and formalize the decision ledger. From 180 to 360, transition the highest-confidence loops to autonomous-with-oversight under Article-6-compatible governance, expand to adjacent loops, and move the decision layer from a project to a product owned by operations — not by IT. The sequence is deliberately linear. Parallel scaling is where POC theater reappears under a different name.