Author: Peng Zhao (Founder of Zhicifang, Co-Founding Partner of Yunhe Capital)
When we placed the core sketch of this white paper in front of a factory director, we deliberately offered no explanation. While waiting, I was anxious; in industrial research, the greatest fear is falling into self-indulgent "echo chambers." After sixty seconds, the director looked up and immediately named the diagram: "A Phased Planning Blueprint for AI Integration into Factories," then pointed out where his team currently stood on the map.
At that moment, my anxiety subsided.
Because this diagram addresses the biggest pain point facing industrial AI today: The industry does not lack technology; it lacks a "construction map" that allows suppliers and demanders to speak the same language.
NVIDIA CEO Jensen Huang once used the "Five-Layer Cake"—energy, chips, data centers, models, and applications—to establish coordinates for the AI industry, providing a common language. However, at the very top of this cake lies the "Application Layer," which carries the most complex process mechanisms, business workflows, and human-machine collaboration relationships found on the industrial floor. It condenses the myriad implications of true AI implementation yet has not been clearly articulated until now.
This has resulted in bottlenecks on both ends:
Factories are stuck with "Data but No Judgment": MES, ERP, and SCADA systems are all present, but the value of the data remains unrealized;
The Industry is stuck with "Practice but No Map": Some talk about platforms, others about Copilots, and still others about agents. The same term may refer to entirely different things depending on the vendor. When clients ask, "To what level has your solution progressed?", they receive answers that cannot be compared horizontally. When vendors say, "Our system is intelligent," their promises remain unverifiable.
An entrepreneur who has been deeply involved in industrial computing for over forty years put it bluntly: "Everyone is exploring the industrialization of AI, but there is a difference between articulating the underlying architecture and standards versus staying silent."
Today, I want to briefly outline the White Paper on Edge Intelligence Application Frameworks and Industrial Agent Grading, which aims to do exactly this: slice open the hardest layer of the "Application Layer" cake and mark its evolution from "Assisting Humans" to "Liberating Humans."
From "5-Hour Firefighting" to "Confident Delegation": A Trust Handover Inside a Pick-and-Place Machine
On a Tuesday morning at 10:00 AM, a pick-and-place machine on an SMT production line in an electronics factory in East China triggered a cycle time overrun alarm.
The production supervisor spent 40 minutes switching back and forth between three systems—MES, SCADA, and QMS—to piece together data. The data pointed to either "feeding abnormality" or "nozzle wear," but the only senior technician capable of making the final call was in another workshop. Even after a judgment was made, determining the specific handling method required repeatedly aligning with equipment, process, and production departments via WeChat groups...
It wasn't until 3:00 PM that the production line finally resumed. As for the root cause of this failure and the troubleshooting experience, they were scattered across chat logs and the minds of those few individuals. Two weeks later, the same anomaly occurred again, and everything had to start from scratch.
Three months later, the same line exceeded limits again. This time, the AI assistant automatically aligned data from the three systems in advance, directly listing two possible root causes, the basis for each judgment, and a recommended troubleshooting sequence. Based on this, the supervisor arranged personnel to check the nozzle first, precisely locating the issue in 20 minutes and restoring the line in one hour.
Six months after that, high-frequency anomalies like "nozzle wear," which had been repeatedly verified, saw their handling authority formally transferred to the system. Within preset rules, the system automatically generated work orders, called for spare parts, and fine-tuned scheduling seamlessly. Engineers no longer needed to rush around putting out fires; they could focus solely on exceptions outside the rules.
These scenes are not isolated cases of a single enterprise but rather a microcosm of industrial AI evolution distilled from practices in multiple frontline factories. Each stage described here is currently playing out in different workshops.
Looking down this production line, the three core questions that industrial AI cannot avoid are laid bare on the table: Where does the data come from? Where does the judgment come from? And ultimately, who makes the decision?
The path these factories have walked step by step is precisely the two core coordinate axes drawn by this white paper: The vertical axis defines the architecture, while the horizontal axis measures trust.
Vertical Axis Defines Architecture: Three-Layer Decoupling from "Data Fidelity" to "Autonomous Intelligence"
How do we deconstruct the vast and complex application layer of industrial AI?
The scale of division should not follow the inertia of traditional IT tech stack classifications but must return to the three progressively deeper core values of the industrial floor: Making the physical world computable, making industrial capabilities composable, and making decision-making authority transferable.
These three value propositions clearly cut the vertical axis into a three-layer architecture, each corresponding to distinctly different commercial logics.
Layer 1: Entity Intelligent Computing Foundation
In the past, we habitually called this the "Connection Layer," which severely underestimated its value. Its true core output is "Data Fidelity": ensuring that every objective fact entering the intelligent system is complete, authentic, and traceable.
Many industrial agents do not die from algorithmic inference failures but implicitly from "data contamination." Once, a factory's equipment early-warning model issued false alarms for three consecutive weeks and was nearly taken offline. Investigation revealed the root cause was merely sensor baseline drift. The model was not wrong; the underlying "facts" were.
Layer 2: Domain Intelligence Layer
This is the truly "expert" layer, currently undergoing a profound "chimney-breaking movement." Past industrial software was locked as monolithic packages, like a custom banquet where dishes cannot be changed. Now, it has evolved into "Framework + Skill," where professional capabilities are decomposed into replicable, shareable, and tradable standard blocks.
The core asset of this layer is industry Know-how, which cannot be bought or copied but must grow bit by bit from the shop floor. Its ultimate form is returning the "assembly rights" to the frontline: in some pioneering factories, workshop engineers are already using these blocks to independently build dozens of AI scenarios that solve actual pain points.
Layer 3: Autonomous Intelligence Layer
AI assistants and agents are not two isolated products but different positions of the same system on the "Trust Spectrum." "Autonomy" is absolutely not uncontrolled full automation; it is bounded and governed autonomy.
Therefore, the most critical component of this layer is not the large model itself, but the "Governance Cabin": authorization chains, permission boundaries, process auditing, and exception circuit breakers form a safety foundation equally important as the core loop of "Perception—Observation—Reasoning—Action." An industrial agent without a governance cabin dies faster the more capable it becomes.
In actual implementation, to determine which layer a technical component belongs to, simply look at whom its final output is delivered to: Trusted facts belong to the Foundation, callable capabilities belong to the Domain Layer, and authorized decisions and actions belong to the Autonomous Layer. This echoes the golden rule of complex decision-making systems: "Facts, Understanding, and Judgment" must be strictly decoupled for the system to endure.
Furthermore, at the bottom of this architectural diagram, the industrial floor is the "Horizon" supporting everything, not one of the three layers.
This three-layer architecture is absolutely not intended to overthrow or replace existing MES, ERP, or SCADA systems in factories, but to enable these legacy systems to "grow" AI capabilities. Previous digital investments by enterprises are the most solid foundation, not sunk burdens.
Horizontal Axis Measures Trust: Five-Level Evolution from "Human-in-the-Loop" to "Human-on-the-Loop"
The core proposition of industrial AI implementation has never been "how smart the model is," but "who holds the decision-making power, and how to confidently hand it over."
Therefore, when constructing the agent grading system, the white paper deliberately avoided considering model parameters or computing power size. Large models iterate monthly; grading by technical metrics will become obsolete in half a year. However, the transfer of decision-making authority follows the underlying laws of organizational management and human nature, remaining unchanged for decades. Just as car owners buying autonomous driving vehicles don't care whether the underlying solution is pure vision or LiDAR, what they truly care about is: Under what road conditions can I safely let go of the steering wheel?
How is this trust of "daring to let go" built bit by bit? Consider the story of a factory HVAC unit.
Two years ago, every Friday before leaving work, facility engineers would spend a long time pondering weekend schedules and weather forecasts, agonizing over whether to turn off the air conditioning and when. Later, the factory deployed an AI system. Every Friday, the system automatically pushed shutdown recommendations, faithfully attaching a long list of judgment bases. Initially, engineers were skeptical, checking each basis item by item before daring to click "Confirm." By the third month, they found the system calculated more accurately than themselves and never missed, so they began looking only at the conclusion and approving directly.
After several consecutive months of "zero errors," the factory made a decision: formally hand over the "Weekend Shutdown" button to the system, while drawing two red lines: Must automatically restore cooling two hours before the Monday morning shift; if workshop temperature exceeds limits during the period, immediately restore cooling and trigger an alarm.
The first summer after delegation faced a heatwave test. On a Sunday dawn, the workshop temperature rose rapidly. The system keenly captured the trend of exceeding limits, restored cooling in advance according to rules, and sent a notification to the engineer. When the engineer entered the workshop for the Monday morning shift, the site was already cool. Because "exception fallback" was written into the rules, this sudden situation did not shatter trust but instead gave the factory complete peace of mind. By the second year, this system even began autonomously coordinating staggered operation across units.
The journey of this HVAC unit corresponds exactly to the first three steps of industrial agent grading. Higher levels L4 and L5 are also unfolding in this direction:
L1 Data Dashboard: Humans view data; humans make decisions.
L2 Intelligent Recommendation: The system provides suggestions with bases; humans decide and execute.
L3 Execution Coordination: The system executes automatically within preset rules and reports to humans upon completion.
L4 Process Autonomy: The system understands goals and plans steps autonomously; humans handle only exceptions.
L5 Intent-Driven: Humans only issue intents (e.g., "Reduce annual energy consumption by 5%"); the system fully takes over implementation.
Please pay special attention to the watershed between L3 and L4.
From L1 to L3, it is "Human-in-the-Loop," where key decisions or rule boundaries remain firmly controlled by humans; From L4 to L5, it is "Human-on-the-Loop," where humans retreat to setting goals and handling exceptions. Crossing this line is not just a technological upgrade but a reconstruction of responsibility attribution and organizational structure. Therefore, technical feasibility does not equal organizational willingness to delegate.
To ensure this grading standard stands firm and runs smoothly on the industrial floor, the white paper clarifies three basic principles:
1. Grade based on "Scenario Authorization," not "Vendor Marketing": Which level an agent belongs to depends on how much authorization it receives in a specific scenario. Even the same system can be delegated to L3 in public utility energy scenarios but must stay at L2 in safety interlock scenarios.
2. Trust is "Dynamic Expansion/Contraction," capable of rising and falling: The transfer of decision-making authority is not a one-time deal but an evolutionary closed loop of "Verification—Authorization—Operation—Circuit Breaker (or Upgrade)—Re-verification." Like a ratchet mechanism, once extreme anomalies trigger a circuit breaker, the system temporarily reverts to a lower level for manual takeover. However, previously accumulated data and experience are not wiped clean and can be advanced again anytime after repair.
3. High or Low Levels Only Represent "Authority Boundaries," Not "Technical Superiority": Higher levels do not necessarily replace lower ones. Conference rooms and factory HVAC can confidently reach L3 or even L4, but safety interlocks in hazardous chemical workshops perhaps should always stay at L2. L2 does not mean technological backwardness but respect for industrial safety.
Final Thoughts
One vertical and one horizontal axis draw the coordinate system for industrial AI implementation.
The most direct value of this map is changing the dialogue between buyers and sellers. When clients stop asking "How large are your model parameters?" and start asking "What L-level has this scenario reached? What is the measured reliability? To which level can it be handed over in six months?", the competition in industrial AI completely shifts from "competing on who sounds smarter" to "competing on who earns genuine customer trust."
The coordinates are established; the practical interrogation has just begun: For a specific workshop scenario, where exactly should it stop at L-level? How are risk and reliability calculated? Which industries will lead the way?...
To answer these real frontline questions, we will subsequently launch three episodes of in-depth video interviews, dialoguing with three practitioners who have personally immersed themselves in the field:
From a technical perspective, seeing how "trust" is quantified and safeguarded;
From an industry perspective, seeing who starts running first and where the money comes from;
From a customer perspective, seeing how factories gradually hand over the steering wheel.
No concepts, only practice. Let us take this map into the industry and dismantle the real problems of industrial AI implementation one by one.
Note: This white paper is jointly launched by Zhicifang·IoT Think Tank and Advantech (the initiating research partner). The full text will be officially released at the 2026 China AIoT Industry Annual Conference. Factory scenarios in the text are organized based on corporate interviews and public cases, presented anonymously, with figures for illustrative purposes only.