CII
ControlledIndustrialIntelligence
Inaugural Issuen2t.net/ark:66356/c100010
Industrial AI
From Intelligence to Action.
cii.network
Published byCII NETWORK
CIICII № 01 · The Gap · The Substrate · The Authority to Act
Founding Note
CII № 01
Inaugural Issue
Founding Note
Artificial intelligence is advancing at extraordinary speed.
Industrial AI is not keeping pace.
AI research, across academia and industry, continues to break new ground, with many results performing strongly in benchmarks and real-world applications.
Industrial practitioners seek productivity gains from AI, but must contend with complex industrial data, uncertain outcomes, and potential risks.
As Industrial AI moves from breakthroughs to real-world deployment, explainability, verifiability, and controllability become essential.
This is where
Controlled Industrial Intelligence
begins.
In this issue
CII № 01
Contents
Define the Problem
The Gap
Build the Foundation
The Substrate
Govern the Action
The Authority to Act
CII Atlas
Publisher
CII NETWORK
CII NETWORK is a Swiss association dedicated to exploring and advancing Controlled Industrial Intelligence, connecting practitioners across industry and academia worldwide to bring intelligence into action.
Network Members
- Quality and Data Analytics Lab · HKUST
- AI & Mobility Lab · TUM
- Chinese Academy of Intelligent Decisions & Control · Taipei
- ASCETEX INTERNATIONAL LIMITED · Hong Kong
Part 01 · Define the Problem
01
The Gap
Between Intelligence and Action
01 · The Gap
Why is Industrial AI
hard to deploy?
AI capabilities are advancing rapidly,
but moving from intelligence to action
brings new challenges to deployment.
AI is scaling fast
AI can now understand and generate language, images, video and code, reason across complex tasks, and increasingly use tools and software.
88%of respondents report their organization uses AI in at least one business function.
75%of surveyed enterprise AI users can now complete tasks they could not before.
Stanford AI Index 2026, on McKinsey data · OpenAI, The State of Enterprise AI 2025
Industrial AI is scaling
more slowly
AI is already used in manufacturing for inspection, predictive maintenance, production optimization and engineering design, but at limited scale, with many projects still in proof of concept or pilot.
29%are using AI/ML at the facility or network level.
23%are piloting AI/ML technologies.
The geography of Industrial AI
sits at the intersection of AI capability and industry base.
AI capability maps technology and talent; industry base captures industrial depth and potential. Their overlap is where Industrial AI matters most.
- AI capability
- Research · Talent · Models · Compute
- Industry base
- Manufacturing · Automation · Robotics · Semiconductors · Process Industries
Fig. 1The geography of Industrial AI
01 · The Gap
CII № 01 · Controlled Industrial Intelligence
The geography of Industrial AI
sits at the intersection of AI capability and industry base.
AI capability maps technology and talent; industry base captures industrial depth and potential. Their overlap is where Industrial AI matters most.
- AI capability
- Research · Talent · Models · Compute
- Industry base
- Manufacturing · Automation · Robotics · Semiconductors · Process Industries
Fig. 1The geography of Industrial AI
01 · The Gap
Between Intelligence and Action
The landscape of Industrial AI applications
Industrial AI is proving its value across many stages of production,
yet the breadth and depth of adoption still vary across use cases.
SizePotential industrial impact (relative)
ColorHigher adoptionLower adoption
ReferenceCase or survey data available
Observe
Sense the physical world
Understand
Turn data into insights
Advise
Generate and evaluate options
Optimize
Improve industrial outcomes
Act
Execute in the physical world
Visual inspectionDefect detection and classification
Metrology & dimensional inspection
In-line testing
Environmental monitoring
Predictive maintenanceEquipment health and remaining useful life
Process quality analytics
Anomaly detection
Root-cause analysis
Failure diagnosis
Asset performance management
Engineering copilotsDesign, programming and troubleshooting
Industrial knowledge assistants
Industrial code generation
Simulation & virtual validationProcess, line and factory simulation
Design optimization
Safety & risk analysis
Compliance & documentation
Production optimizationThroughput, cycle time and OEE
Energy optimization
Yield optimization
Material flow optimization
Production scheduling
Inventory optimization
Process parameter optimization
Quality prediction
Autonomous operationsSelf-directed production and operation
Closed-loop interventionDirect adjustment of physical processes
Learning-based adaptive control
Real-time optimization
Embodied task execution (VLA)
Adaptive manipulation
Autonomous intralogistics
01Proven value
Already widely deployed
in many industrial settings.
02Expanding frontier
Rapidly growing, with increasing
deployment across industries.
03Untapped potential
High potential impact,
but adoption remains limited.
Fig. 2The landscape of Industrial AI applications
01 · The Gap
CII № 01 · Controlled Industrial Intelligence
The landscape of Industrial AI applications
Industrial AI is proving its value across many stages of production,
yet the breadth and depth of adoption still vary across use cases.
SizePotential industrial impact (relative)
ColorHigher adoptionLower adoption
ReferenceCase or survey data available
Observe
Sense the physical world
Understand
Turn data into insights
Advise
Generate and evaluate options
Optimize
Improve industrial outcomes
Act
Execute in the physical world
Visual inspectionDefect detection and classification
Metrology & dimensional inspection
In-line testing
Environmental monitoring
Predictive maintenanceEquipment health and remaining useful life
Process quality analytics
Anomaly detection
Root-cause analysis
Failure diagnosis
Asset performance management
Engineering copilotsDesign, programming and troubleshooting
Industrial knowledge assistants
Industrial code generation
Simulation & virtual validationProcess, line and factory simulation
Design optimization
Safety & risk analysis
Compliance & documentation
Production optimizationThroughput, cycle time and OEE
Energy optimization
Yield optimization
Material flow optimization
Production scheduling
Inventory optimization
Process parameter optimization
Quality prediction
Autonomous operationsSelf-directed production and operation
Closed-loop interventionDirect adjustment of physical processes
Learning-based adaptive control
Real-time optimization
Embodied task execution (VLA)
Adaptive manipulation
Autonomous intralogistics
01Proven value
Already widely deployed
in many industrial settings.
02Expanding frontier
Rapidly growing, with increasing
deployment across industries.
03Untapped potential
High potential impact,
but adoption remains limited.
Fig. 2The landscape of Industrial AI applications
01 · The Gap
Between Intelligence and Action
The closer Industrial AI gets to action,the harder it becomes to deploy
Let’s push the boundary TOGETHER.
— CII NETWORK
01 · The Gap
CII № 01 · Controlled Industrial Intelligence
The closer Industrial AI gets to action,the harder it becomes to deploy
Let’s push the boundary TOGETHER.
— CII NETWORK
01 · The Gap
Between Intelligence and Action
Industrial control remains fragmented
across controllers, protocols and system layers.
Part 02 · Build the Foundation
02
The Substrate
Software-Defined Control Systems
02 · The Substrate
What carries intelligence into action?
Control systems connect computing to physical execution — and software-defined control offers a stronger foundation.
AI acts onExamplesActs throughControl layer
Digital systems
Information & transactions
Finance, procurement, planning
APIs & workflows
—
Physical operations
Physical processes & operations
Energy, logistics, buildings, process industries
OT & supervisory systems
Relevant
Machine control
Machines & motion
Industrial equipment, robotics, automation systems
Controllers & real‑time control
Critical
Our focus Where intelligence meets the physical world.
Table 1Industrial AI across digital and physical systems
How intelligence
reaches machines
AI analyzes data and proposes actions.Control systems observe machine states and execute commands in real time.
The rise of software‑defined
control systems
Decoupling from dedicated hardware enables broader data integration and more advanced computation.
Fig. 3The control substrate
Fig. 4From hardware-defined to software-defined control
Hard, soft and virtual PLCs will likely coexist in the future.
Automation.com, 2025
01Compute
General‑purpose computing is now powerful enough for real‑time, deterministic control.
02Integration
Functions once split across controllers can now share one platform.
03Flexibility
Control capabilities can evolve through software updates, not hardware replacement.
02 · The Substrate
CII № 01 · Controlled Industrial Intelligence
How intelligence
reaches machines
AI analyzes data and proposes actions.Control systems observe machine states and execute commands in real time.
The rise of software‑defined
control systems
Decoupling from dedicated hardware enables broader data integration and more advanced computation.
Fig. 3The control substrate
Fig. 4From hardware-defined to software-defined control
Hard, soft and virtual PLCs will likely coexist in the future.
Automation.com, 2025
01Compute
General‑purpose computing is now powerful enough for real‑time, deterministic control.
02Integration
Functions once split across controllers can now share one platform.
03Flexibility
Control capabilities can evolve through software updates, not hardware replacement.
02 · The Substrate
Software-Defined Control Systems
AI and control run on different clocks,
but they must work together
01DeloitteTech Trends 2026
Language and vision models work in “human time”, Ayanna Howard tells Deloitte. For a walking robot, “a one- or two-second delay means it drops something, crashes, or potentially hurts someone.”
02NVIDIAIGX ThorJetson ThorHoloscan
A real-time Linux kernel and isolated GPU partitions give real-time work “hard performance guarantees” beside AI. Holoscan runs from “AI reasoning” to “deterministic real-world actions”.
03BainFrom control to intelligence
“Industrial automation is no longer about controlling machines—it’s about orchestrating intelligence.” Bain sees “AI-native workflows” moving “from analytics layers into the operational core”.
04ARCSoftware-defined automation
One theme stood out to ARC: industrial AI “will scale only when intelligence can be connected to deterministic, real-time action on heterogeneous equipment.”
Asynchronous
Variable-latency
Real-time
Deterministic
A software-defined
control architecture
Unified compute, memory and control.
Runtime actions ↓Real-time data ↑
Fig. 5The clocks of AI and control
Fig. 6AI and control on one industrial PC
01Bring compute, control and memory closer
The same trend can be seen across computing architectures. In-memory computing (IMC) brings compute closer to memory; co-packaged optics (CPO) brings optical I/O closer to compute. For Industrial AI, compute, control and memory (CCM) can follow the same principle — bringing AI, machine data and real-time control closer together on one machine.
02Different loops, direct connection
AI runs asynchronously on the GPU/NPU, while control runs independently on real-time CPU cores with deterministic timing. They keep their own loops, but exchange machine data and AI commands directly through shared memory.
asynchronous · variable-latency~10–200 Hz~1 kHz – multi-kHz
02 · The Substrate
CII № 01 · Controlled Industrial Intelligence
AI and control run on different clocks,
but they must work together
01DeloitteTech Trends 2026
Language and vision models work in “human time”, Ayanna Howard tells Deloitte. For a walking robot, “a one- or two-second delay means it drops something, crashes, or potentially hurts someone.”
02NVIDIAIGX ThorJetson ThorHoloscan
A real-time Linux kernel and isolated GPU partitions give real-time work “hard performance guarantees” beside AI. Holoscan runs from “AI reasoning” to “deterministic real-world actions”.
03BainFrom control to intelligence
“Industrial automation is no longer about controlling machines—it’s about orchestrating intelligence.” Bain sees “AI-native workflows” moving “from analytics layers into the operational core”.
04ARCSoftware-defined automation
One theme stood out to ARC: industrial AI “will scale only when intelligence can be connected to deterministic, real-time action on heterogeneous equipment.”
Asynchronous
Variable-latency
Real-time
Deterministic
A software-defined
control architecture
Unified compute, memory and control.
Runtime actions ↓Real-time data ↑
Fig. 5The clocks of AI and control
Fig. 6AI and control on one industrial PC
01Bring compute, control and memory closer
The same trend can be seen across computing architectures. In-memory computing (IMC) brings compute closer to memory; co-packaged optics (CPO) brings optical I/O closer to compute. For Industrial AI, compute, control and memory (CCM) can follow the same principle — bringing AI, machine data and real-time control closer together on one machine.
02Different loops, direct connection
AI runs asynchronously on the GPU/NPU, while control runs independently on real-time CPU cores with deterministic timing. They keep their own loops, but exchange machine data and AI commands directly through shared memory.
asynchronous · variable-latency~10–200 Hz~1 kHz – multi-kHz
02 · The Substrate
Software-Defined Control Systems
The Software-Defined
Control Ecosystem
From hardware and OS to software-defined control and applications.
Applications
Software-defined control
Operating Systems
Hardware
Machines, systems and industrial applications
Motion and logic as reusable software
Real-time execution platforms
x86 and Arm compute platforms
For further insights and practical applications, contact us.
contact@cii.network
02 · The Substrate
CII № 01 · Controlled Industrial Intelligence
The Software-Defined
Control Ecosystem
From hardware and OS to software-defined control and applications.
Applications
Software-defined control
Operating Systems
Hardware
Machines, systems and industrial applications
Motion and logic as reusable software
Real-time execution platforms
x86 and Arm compute platforms
For further insights and practical applications, contact us.
contact@cii.network
02 · The Substrate
Software-Defined Control Systems
Part 03 · Govern the Action
03
The Authority to Act
Evidence, Risk, and Human Oversight
03 · The Authority to Act
How can intelligence be authorized to act?
Capability does not imply authority. A well-defined authorization mechanism enables AI to act autonomously within clear conditions and boundaries.
03 · The Authority to Act
CII № 01 · Controlled Industrial Intelligence
Autonomous AI is already at work
Yokogawa × Aramco2025
10%–15%
“reduction in its amine and steam usage” at Aramco’s Fadhili Gas Plant, where multiple autonomous control AI agents now run acid gas removal.
Fadhili Gas Plant · Aramco, via Yokogawa
Caterpillar2026
Nearly700
“Cat autonomous trucks are in operation around the world and together they have safely hauled over 11 billion tonnes of material.”
Autonomous Cat truck · RobSimmons223311, CC BY-SA 4.0
Figure × BMW2025
90,000+
“parts loaded” by Figure 02 on an active assembly line at BMW Group Plant Spartanburg: “1,250+ hours of runtime” in an 11-month deployment.
Figure 02 at BMW Spartanburg · Figure AI
The broader agentic AI landscape
WEF × Capgemini2026
ACAP
Introducing its Agent Capability and Authorization Profile, the WEF notes that organizations “face new challenges in defining the conditions under which these agents are authorized to act and in ensuring this authority is enforced as systems evolve.”
Gartner2025
Over40%
“of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls.”
Dynatrace2026
Only13%
“of organizations use fully autonomous agents,” and “69% of agentic AI–powered decisions are still verified by humans.”
03 · The Authority to Act
Evidence, Risk, and Human Oversight
But authority is still an open question
Autonomy is already here, but it is uneven
Aramco says the Fadhili system “has now been permanently adopted.” It went in step by step: trained and checked on a simulator, then connected to the plant’s existing control system. Haulage, process control and general AI agents are different problems. Autonomy succeeds first where the task is clear, the setting controlled and the boundary verifiable.
2%of organizations have deployed AI agents at scale, 12% at partial scale. Most are still piloting (23%) or exploring (61%), in a survey of 1,500 executives.Capgemini, Jul 2025
Authority is becoming everyone’s question
The WEF’s ACAP spells out an agent’s permissions, checkpoints, owners and evidence; NIST calls for “appropriate identification and authorization controls.” From traditional automation to today’s AI agents, governing autonomous behavior has long been a central challenge.
Human approval is not the final answer
Claude Code users approve 93% of permission prompts, later 97%: approval fatigue, “where people stop paying close attention to what they’re approving.” Developers are not plant operators, but if every low-risk step needs a click, the approvals that matter get lost. Authority should also let work that is fit to run alone stop waiting.
13.6%of dangerous commands caught by human review in a controlled experiment; auto mode caught 89%.Anthropic, 2026
03 · The Authority to Act
CII № 01 · Controlled Industrial Intelligence
Five approaches
Exploring exploration
From automation to AI, industry and academia have long explored how to guide and constrain machine behavior.
01 · Control engineering
Rules
Interlocks and permissive checks enforce the conditions under which a machine may operate.
02 · Process control
Optimization
A controller uses a process model to meet an objective within defined operating constraints.
03 · Human factors
Supervision
People and automated systems share the work of gathering information, choosing actions and executing them.
04 · Machine learning
Behavioral Learning
Data, rewards and human feedback shape behaviour, guided by preferences and explicit principles.
05 · Robotics · Aerospace
Uncertainty & Runtime Assurance
Uncertainty estimates and runtime monitors help determine when to seek assistance or switch to a backup.
03 · The Authority to Act
Evidence, Risk, and Human Oversight
Five open questions
Challenges That Remain
As AI moves from reasoning to action, fundamental questions remain.
- How accurate is accurate enough?
- If not accuracy, what should we measure?
- Where are the boundaries, and who sets them?
- What should humans do: approve, supervise, or step in?
- Who is accountable when AI acts?
The question behind them
Why is moving from intelligence to action so difficult?
Part of the answer may lie in
authority.
03 · The Authority to Act
Evidence, Risk, and Human Oversight
Real-world case
Authority-Gated Fault Management
Communication faults in semiconductor manufacturing
01 The challenge
In semiconductor manufacturing, AI can quickly diagnose communication faults and recommend corrective actions. But with expensive equipment and costly downtime, engineers are cautious about allowing AI to act autonomously.
02 The approach
Working together, engineers and data scientists defined a 3×3 authority matrix based on evidence strength and action risk. The matrix determines how far AI can proceed—from monitoring and proposing fixes to autonomous execution or human approval.
With sufficient evidence and the required checks, AI can handle low- and medium-risk recovery actions autonomously. High-risk actions still require human approval.
| Evidence strength | Action risk | ||
|---|---|---|---|
| Low | Medium | High | |
| Strong | Execute+ post-check | after verificationExecuteafter verification | Humanapproval |
| Moderate | Propose | Propose+ verify | Humanapproval |
| Weak | Monitor | Monitor | Humanattention |
- AI executes
- AI proposes
- AI monitors
- Human decides
For more cases and practical insights, contact us.
Atlas, Rockefeller Center, New York (Lee Lawrie, 1937). Shouldering the celestial sphere, the figure echoes the center’s art theme: New Frontiers and the March of Civilization.
n2t.net/ark:66356/c100062The exchange
CII Atlas
Connecting Needs with Solutions
CII Atlas
Connecting
- NeedsReal industrial challenges
- MatchingConnecting needs and expertise
- SolutionsResearch and technical solutions
- ValidationTested in real environments
- DeploymentIntelligence into action
Test · Refine · Repeat
Visitors worldwide