Why AI, ML, LLM, RAG, and Agents end up in a single sentence of a factory RFP

Why Factory RFPs Cram AI, ML, LLM, RAG, and Agents into a Single Sentence

And the Layer That One-Pager Is Actually Pointing To

Source: Each layer isn't a replacement for the previous one—it sits directly on top of it, elevating capabilities and responsibilities at every step.   

Terminology gets jumbled first in boardrooms and procurement documents. Statements like “We’ll catch defects with Generative AI,” “We’ll introduce LLMs,” or “We’ll run operations agentically” shove completely different layers into the exact same sentence. The attached diagram folds that confusion down into a single page.

AI forms the foundation, with ML built above it, DL above ML, followed by GenAI, LLM, and RAG stacked higher, and Agentic AI sitting at the very top. Each layer is not a substitute for the one beneath it; it assumes the layer directly below it as a prerequisite to elevate capabilities and responsibilities by one level.

The factory's real question isn't about model names. It is: Which layer is actually required for this task? And what are the inputs and outputs of that layer on top of our existing MES, SCADA, FDC, PLC, and CCTV systems?

Source: Towards AI. The layers are cumulative, not a pyramid. If you try to buy only the top Agent layer, it will end up running with the layers beneath it completely empty.

1. The Big Picture — Each layer enables the next

The core of this diagram lies on the left: THE BIG PICTURE. You can read it from the outside in, or from the bottom up.

· AI — The broadest umbrella that enables machines to think, reason, and act. This includes rule-based expert systems.

· ML — Instead of humans hand-writing every rule, models learn patterns directly from data.

· DL — Pushes that learning further using multi-layer neural networks. This unlocks complex inputs like images, audio, and time-series data where manual feature engineering is too difficult.

· GenAI — Goes beyond classification and prediction to create entirely new content, such as text, images, and audio.

· LLMs — Large-scale models within GenAI specialized in understanding and generating language.

· RAG — Connects external knowledge (documents, manuals, logs, SOPs) to the language model, grounding its answers in up-to-date facts and internal guidelines.

· Agentic AI — Moves beyond understanding and generation to decompose goals, utilize tools, plan, and execute tasks autonomously.

The right side, THE EVOLUTION, retells the same story chronologically: Think → Learn patterns → Learn complex patterns → Create content → Understand language → Knowledge + generation → Act autonomously.

The most critical statement is already written at the top of the diagram: Each layer builds on the previous one, adding more capability and intelligence. If a factory overlooks this concept, they end up purchasing an LLM chatbot while expecting it to tune recipe parameters.

Source: LinkedIn / Dr. Archana Ashok Generation and action are not in the same circle. GenAI creates text, while Agentic AI attaches tools and goals.

2. The Layer Shown by Real-World Examples — A chatbot and a chamber are not the same "AI."

Translating the REAL-WORLD EXAMPLES at the bottom of the diagram into factory language clearly separates the layers.

Layer	Examples in the Diagram	Actual Application in Factories	Delusion of Stopping at This Layer
AI	Expert systems, Robotics, Game AI	Interlocks, recipe branching, alarm logic	Assuming "we already have AI" and skipping learning and grounding altogether
ML	Spam detection, Recommendations, Fraud detection, Predictive analytics	Yield prediction, equipment anomaly scoring, demand forecasting	Believing tabular data models can handle vision, audio, and vibration data
DL	Image/Voice recognition, Autonomous driving	AOI defect detection, hard hat/zone violation detection, acoustic anomaly detection	Focusing only on accuracy rates while omitting corrective actions and audit trails
GenAI	ChatGPT, Midjourney, DALL-E, Diffusion models	Standard Operating Procedure (SOP) drafts, anomaly report drafts, training content	Directly using generated text as live factory execution commands
LLMs	GPT-4/5, Claude 3, Gemini, Llama 3, Qwen	SOP and FDC log queries, cause hypothesis drafting	Assuming the model already knows our specific processing chamber
RAG	Enterprise chatbots, Internal doc Q&A, Knowledge bases	Answering based on Work Permits, LOTO, recipes, and historical tickets	Believing that adding search completely eliminates hallucinations and liability
Agentic AI	Autonomous agents, Tool use, Workflow automation	Patrol routing, recipe adjustment suggestions, equipment shutdown recommendations	Equating autonomy with unverified automated execution

This is where the diagram is honest. ChatGPT is an example of LLM/GenAI, not a factory’s Agentic layer. Enterprise document Q&A is an example of RAG, not a system that issues commands to a PLC. Pricing the "body" of a humanoid or robot dog on the open market and making the judgment for that body to halt within our internal regulations belong to completely different layers.

Source: Palladium Dynamics Smart factory dashboards are already up on the screen. The real missing piece isn't the dashboard—it's knowing which layer produced those numbers and who approves the execution.

3. Why This Hierarchy Becomes Sharper in the Factory

· Understand the layers and their relationships: Buying an LLM does not mean you have purchased RAG, and deploying RAG is not the same as running an Agent.

· Select the right layer for the problem: Hard hat detection can be fully handled by DL vision. Citing historical anomaly logs from identical recipes to document a root cause requires RAG. Recommending a flow rate adjustment within allowed thresholds requires an Agent and a validation gate.

· Build value-generating products: Demos end at the generation layer. Yield, defects, and serious safety incidents are resolved at the action and audit layer.

· Apply RAG and Agents appropriately for accuracy and autonomy: Accuracy comes from retrieved evidence; autonomy operates within an authorized action space. Lumping them together under a single buzzword obscures the single points of failure.

· AI continuously evolves; future-proofing is layer architecture, not model swapping: Deep learning gaining the ability to read images is fundamentally different from a system determining whether that image complies with current SOPs.

Source: promptengineering.org AI ⊃ ML ⊃ DL is a textbook diagram. The next circle a factory must question is not GenAI, but how its output closes the loop with safety regulations and equipment control commands.

4. The Structural Gap — Converting Language Model Output into Factory Input

Moving a generic tech stack directly onto the shop floor leaves five crucial lines blank.

It exposes the exact same kind of gap seen between general-purpose Robotics VLA (Vision-Language-Action) models and real-world industrial environments.

Key Factor	General-Purpose LLM / Demo Agent	Industrial Site Requirement
Action Space	Tokens, tool calls, chat responses, 6DoF poses	Recipe parameters, equipment control commands, operating envelopes
Knowledge	Pre-trained corpus + web (if connected)	SOPs, Work Permits, LOTO, FDC data, internal failure modes
Sensor Modality	Text ± static images	FDC time-series + video streams + audio/vibration/thermal telemetry
Control Rhythm	Conversational turns, cloud round-trips	Wafer/batch events, low-latency on-premises execution
Safety & Audit	Basic policy filters, black-box tolerance	Full evidence trail for every decision; complete auditability for yield, defects, and serious safety incidents

It is not because the LLM is weak—it is because the object of control is a process, not a sentence. It is not because RAG is weak—it is because the search target is unexportable fab data, not Wikipedia. It is not because Agents are a fantasy—it is because factory autonomy must mean that only Typed Actions are permitted to pass through validation gates, rather than a system that "does whatever it wants."

This is why the KEY MODERN CONCEPTS at the bottom of the diagram are rewritten into factory language:

· Multimodal AI: Processing not just text, but video, sensor data, and system logs within a unified context. Existing CCTV and FDC setups are the true starting points here.

· RAG: Retrieving internal regulations and historical incident logs either prior to or alongside text generation.

· Fine-tuning: While general-purpose models solve "what is visible," domain-specific fine-tuning answers "what is allowed in this specific chamber and recipe."

· Prompt Engineering: The prompt is merely the entry point. The true output consists of structured, schema-compliant actions and explicit reasoning chains.

· AI Safety: Not just conversational safety filters, but explicit architectural safeguards that prevent erroneous actions from impacting yield or causing severe safety incidents.

· AI Agents: Planning, tool use, and execution—with the caveat that execution rests with the PLC, while the Agent functions as a recommendation engine.

Source: Weaviate RAG is not a trick for feeding documents to an LLM. It is a mechanism for leaving a trace of exactly which page in your knowledge base an answer originated from.

Source: JR Automation On a semiconductor line, commands do not belong in a chat window. They are wafers, recipes, and interlocks.

5. Building Layers on a Single Core — Turning Buzzwords into Products

Viewing Mithril purely as a safety company captures only a single box of the big picture. The product lineup is diverse, but the engine is unified. Mapping the diagram's hierarchy onto real-world field adapters makes the architecture immediately clear:

· NENYA: An sVLA (Specialized Vision-Language-Action) core trained directly on industrial video, sensor data, process telemetry, and physical actions. It is not an open-source VLM fine-tuned for conversational chat; it builds an industrial encoder, multimodal fusion, reasoning engine, XAI, validation gates, and PLC adapters as native, layered capabilities.

· Guardian-Alpha: Focuses on personnel, safety risks, and IoT integration. Combines DL vision + Agentic actions (PTZ control, alerts, machine signals) + audit trails.

· Optima-Alpha / Optima-Vision: Handles FDC, quality, and defect analysis. Integrates RAG-backed domain knowledge and DL-driven scratch/particle detection into a unified loop.

· Robotics & Physical OS: The physical action layer for quadrupedal robots, drones, and humanoids—the exact point where Agentic AI maps directly to physical hardware.

All four products share a single underlying sequence that mirrors the evolution diagram: Perception → Context → Evidence → Action.

Core Prerequisites for Industrial Execution

· Engine built on domain data: While general models ask "what is visible," NENYA answers "what is currently permitted for this recipe, this chamber, and this SOP."

· Overlay, do not replace: Integrates seamlessly over existing MES, SCADA, FDC, and PLC infrastructures. The AI recommends, rules validate, and the PLC executes.

· On-premises deployment over existing feeds: Utilizes existing CCTV feeds for safety monitoring and directly ingests active FDC streams without requiring dedicated sensor overhauls or letting data leave the plant network.

· Grounded decision-making: Follows Perception → Decision → Action (within authorized control envelope) → Audit (Evidence Chain). An alert popup on a screen is a feature, not a complete product.

· Field-proven at scale: Deployed across 25+ industrial sites since November 2023, leveraging over 23 million training frames. Demonstrates a 98.7% recall rate for unsafe behavior, a 0.8% false positive rate, and a 94% reduction in safety incidents compared to pre-implementation baselines (internal operational standards). Validated across diverse sectors—including Hanil Hyundai Cement, Sampyo, Sungshin Cement, SPC, AJU, Dongwon Systems, and Hankook Tire—proving the core architecture scales seamlessly from heavy industries to food processing, semiconductor tooling, and battery manufacturing.

· Tangible impact on yield: In battery coating applications, customer feedback documented a 26% defect reduction, 31% lower equipment downtime, and a 3.8 percentage point yield increase. In semiconductor equipment blind testing, NENYA Gen1 achieved a 5/5 score compared to 2/5 from legacy systems. These metrics originate from active production lines, not artificial benchmarks.

NENYA’s generational roadmap aligns directly with the evolutionary layers of the diagram: Gen 1 focuses on understanding and explanation (LLM/RAG layers); Gen 2 introduces actionable execution coupled with validation gates (bounded Agentic space); Gen 3 establishes full PLC integration with continuous feedback re-learning. The target remains autonomous manufacturing intelligence, grounded strictly by the evolutionary sequence the diagram dictates:

AI proposes, rules validate, and the PLC executes.

Source: Korea JoongAng Daily A robot dog at a POSCO steel mill. When the machine enters the physical floor, the remaining question is not about the patrol itself—it is about which layer of the diagram generated the judgment, and where the underlying evidence is stored.

ource: Communications of the ACM The task is not replacing the monitoring screens that are already up. The real work is deciding which layer to stack on top of them.

6. The Three Lines Enterprise Procurement Must Draft Today

Translating the diagram's directive—"Choose the right approach for your problem"—into three concrete clauses for RFPs and procurement documents:

1) Define procurement units by layers and tasks, not model names: Replace "LLM adoption" or "Agent deployment" with explicit parameters: the specific task on this line, required layers (Perception / Evidence / Generation / Action), target accuracy rates, fail-safe interlocks, and designated responsibility for logging reasoning chains. The evaluation dataset must be owned and controlled by the customer, not the vendor.

2) Draft evaluation criteria directly from internal safety and operational regulations: Standard benchmarks have their place, but unless they test against Confined Space Permits, LOTO procedures, fall hazards, conveyor pinch points, and IATF quality traceability, those scores are meaningless on your shop floor. This is precisely why the Evidence Chain must sit as the final step in the operational loop.

3) Do not treat lower layers as legacy enemies: As high-level models grow more capable, the underlying value of existing CCTV networks, FDC pipelines, PLCs, and SOP records increases exponentially. The better general-purpose GenAI becomes at generating text, the more valuable RAG and fine-tuned domain cores become for retrieving and enforcing site-specific recipes and prohibited actions. In factories where system replacement costs hinder adoption, a retrofittable intelligence layer delivers immediate ROI.

7. The Blank Space Left by the Diagram

The attached chart is a family tree of models: AI encompasses ML; ML encompasses DL; DL unlocks GenAI; LLMs handle natural language; RAG attaches evidence; and Agentic AI moves toward specific goals.

What the diagram truly demonstrates is not merely that seven distinct terms exist, but that completely different layers of inputs, outputs, and operational accountability sit under the exact same umbrella term, "AI."

What the factory needs is the blank space right next to it: In this process, on this equipment, under this SOP—which layer do we start from?

Do we need generation, retrieved evidence, or bounded execution? When a failure occurs, what triggers the halt, and how is that judgment logged?

If this space remains blank, ChatGPT remains a PR talking point, while the system deployed on the shop floor defaults to a DL model that merely sounds alarms, an ungrounded generative engine, or an unverified agent. This is precisely why physical bodies that cannot wait for cloud latency co-exist alongside factories that cannot export their data.

Mithril’s response is to unpack the hierarchy of this diagram directly into its product lines. The core reasoning engine is NENYA, safety is handled by Guardian, yield by Optima, and physical motion by Physical OS. The same decision-making framework flows seamlessly from fixed cameras alongside workers to chamber telemetry and autonomous patrol robots. This architecture is backed by strategic R&D with NVIDIA Inception, TIPS, Chungnam C-STAR, and the Pattern Recognition & Machine Learning Lab at Korea University, with global expansions reaching Sunway, TOA, and Elsewedy/AOI.

All that is needed to begin is your existing operational logs, video feeds, the specific incidents you need to prevent, and the yield improvements you aim to capture. With these assets, we can systematically identify whether your immediate operational requirement is Perception, Evidence, or Action.

We invite you to reach out to Mithril to map out where the blank spaces on your shop floor belong within this architectural framework.