Where AI Value Lands

Vertical AI Orchestration

6 min read

Vertical AI Orchestration

One-line

In asset-heavy verticals (energy, mining, manufacturing, utilities, supply chain), the durable AI moat is not the underlying model but the multi-agent orchestration platform that converts model outputs into trusted operational actions — bound by safety, regulatory, and incident-history constraints that do not transfer across organizations.

The structural argument

Per deloitte-ai-dossier-eri: in ER&I, AI use cases compose into multi-agent workflows where:

  • One agent diagnoses a fault (predictive maintenance).
  • A second agent reschedules technicians around the fault.
  • A third agent ensures regulatory compliance for the intervention.
  • A fourth agent reroutes the affected supply chain.
  • A fifth agent dispatches a drone for visual confirmation.
  • A human is in the loop only at consequential decision points (autonomy-slider — “Iron Man suit”).

The economic unit of value is the orchestrator that holds this composition together, not any individual agent and certainly not the underlying foundation model.

Why this is hard to commoditize

Three properties:

  1. Operational context is sticky. The orchestrator’s escalation logic, recovery patterns, and trust thresholds are calibrated against years of incident reports specific to a refinery, a grid, a fleet. None of this calibration transfers to a competitor’s installation.
  2. Regulatory regimes are jurisdiction-specific. ER&I orchestrators must encode local compliance rules (NERC, OSHA, EU ETS, EPA permits). Updates to these rules are continuous and require human-in-the-loop change management.
  3. Liability allocation is contractual. Who is responsible when the orchestrator’s autonomous reroute strands a customer or trips a grid? The answer is encoded in service agreements, insurance riders, and SLAs that constrain the orchestrator’s behavior. Switching orchestrators means renegotiating these contracts.

These properties are absent from horizontal SaaS AI tools and from generic chatbot wrappers. They are precisely the properties that make industrial automation defensible.

Relationship to middle-layer-defensibility

This is the same thesis as middle-layer-defensibility applied outside software. The “workflow-embedded platform” pattern (Cursor, Windsurf, LangGraph) survives in software because of proprietary context, behavioral data, and distribution. Vertical AI orchestration survives in industrials for the analogous reasons — proprietary operational context, incident history, and contractual position.

The distinction worth preserving: software workflow-embedded platforms compete on developer productivity. Vertical AI orchestrators compete on avoided catastrophic loss and regulatory permission to operate. The pricing power is different (productivity gains are negotiable; permission to operate is not).

Relationship to H2_u-curve-of-value

A clean vertical-domain instantiation of the U-curve:

  • Top of the U in ER&I = the multi-agent orchestrator (Aker BP’s Yggdrasil-style platform, supply-chain digital twins, autonomous drone fleets with orchestration logic).
  • Bottom of the U in ER&I = the proprietary operational data + the physical assets themselves (oil wells, transmission infrastructure, fleet sensors, captive geological data). Without these, the orchestrator has no surface to act on.
  • Squeezed middle = generic foundation models and generic agent-orchestration frameworks that lack operational specialization. They become OEM inputs to the vertical orchestrator, not value-capturing layers themselves.

The taxonomy mismatch middle-layer-defensibility surfaces in software applies here too: by position, the orchestrator sits “in the middle” between physical assets below and foundation models above. By function, it captures the entire downstream value of the integrated operation.

Compared to horizontal AI orchestration

LangGraph, AutoGen, CrewAI, OpenAI Agents SDK are horizontal orchestration substrates. They do not own operational context for any specific vertical. They sell shovels in the gold rush.

Vertical AI orchestrators (Aker BP’s Yggdrasil program, Siemens MindSphere extensions, BMW’s multi-agent supply-chain genAI, SBB’s pantograph vision agent, hypothetical players in mining and grid management) own one operational domain deeply. They cannot be lifted-and-shifted across verticals.

The economic prediction: a small number of dominant vertical orchestrators per industry per geography, with switching costs dominated by regulatory and contractual position, not technical lock-in.

Tensions and risks

  • Paradigm-reset risk (cf. world-models-jepa). Vertical orchestrators built on current-generation LLMs face the same architectural obsolescence risk as horizontal ones. A JEPA-class shift would force re-engineering of the orchestration substrate.
  • Adversarial cyber surface. Per deloitte-ai-dossier-eri, the orchestrator is the new attack surface for nation-state and criminal actors targeting physical infrastructure. The moat is also the target.
  • OOD generalization failure. Per the scaling-wall vertical-application manifestation, orchestrators fail unpredictably on “novel ores,” unfamiliar geology, or scenarios outside training distribution. The trust accumulated in normal operation can be lost in a single anomalous incident.
  • Capture by the asset owner. If incumbent ER&I players (Shell, Siemens, ArcelorMittal, the major utilities) build orchestrators in-house, the orchestration vendor market never forms as an independent layer. Value accrues to the asset owner instead of the AI vendor. The empirical question is whether ER&I incumbents have the digital-engineering talent depth to do this — most evidence so far suggests they partner.

Data

Graph View

Backlinks