AI Adoption Trends
In a world where more than 50% of U.S. employees are already using AI in their jobs, with most reporting positive impacts in productivity and efficiency, businesses are taking AI strategy seriously more than ever.
However, adoption in the more traditional sectors, especially the process industry, is lagging behind.
Privacy and Cybersecurity
Widely recognized Language Model (LLM) providers such as OpenAI, Anthropic, and Google offer the easiest way to utilize AI. Their state-of-the-art proprietary models are accessible with little more than a browser and a prompt.
But as the saying goes, there is no such thing as a free lunch.
While convenient and powerful, these platforms raise legitimate concerns around data privacy and ownership, as addressed by various ongoing high-profile lawsuits.
Data privacy is uncompromisable. Product wheel, recipe formulation, throughput data, and trade secrets are all sensitive data that, if leaked, could provide significant leverage to competitors and adversarial attackers.
Cyber security is equally important. Plants allocate significant resources in enforcing cybersecurity — justifiably so, considering the industry's complicated history with cyberattacks and ransomware. As such, every new connection is scrutinized to prevent another Stuxnet or Triton incident.
Unfortunately, the security and privacy considerations involved often leave the OT layer extremely restricted, hampering AI adoption.
What if there was a way to get the benefits of LLMs without external connections and without risking your data? Introducing small language models (SLMs).
Small Language Models
SLMs are typically defined as language models with fewer than 40 billion parameters. Once quantized (a technique used to reduce model size at the cost of some performance), these models can fit into mid-to-high-end consumer GPUs and run locally without reliance on data-center-grade hardware.
Some smaller and Mixture of Experts (MoE) models can even run entirely on CPUs and RAM with acceptable latency.
This makes them particularly well-suited for the process industry and their process control networks: affordable enough for edge deployment, and capable of operating entirely within secure networks.
Intelligence vs. Size
Today's frontier models are undeniably large. For example, leading open-source models GLM-5.1, Kimi-K2.5, and MiniMax-M2.7 have 754, 1,000, and 230 billion parameters, respectively.
To put that into perspective, an RTX 6000 PRO GPU featuring 96 GB of VRAM costs about $10,000, and each billion parameters of an LLM requires approximately 0.5 to 2 GB of VRAM, depending on the quantization.
In comparison, most SLMs can be run using readily available workstations such as the DGX Spark for less than $5,000.
You might think that larger models are more capable — and you would be right.
However, the latest SLMs are reaching what is known as the "efficiency frontier" — the point where flagship-level reasoning is achieved with significantly streamlined architectures. This efficiency frontier can be understood as the top-left quadrant in the figure below (taken from Artificial Analysis).
The Artificial Analysis Intelligence Index combines various LLM benchmarks into a single (albeit imperfect) metric to facilitate model comparison.
As expected, there is a general correlation between model size (total parameters) and the intelligence index. However, we are starting to see SLMs — particularly the recent Qwen3.5 and Gemma4 families — creep into the top-left quadrant (i.e., the efficiency frontier).
Zooming into one of the benchmarks making up the intelligence index — MMLU Pro — we see a similar trend. The MMLU Pro benchmark is designed to evaluate language understanding models across broader and more challenging tasks, comprising over 12,000 curated questions spanning 14 diverse domains including both liberal arts and STEM.
You might notice a dip in performance for smaller variants like Qwen3.5-4B and 9B. The reason is straightforward: smaller models generally have less embedded world knowledge than their larger counterparts.
Rising Atop Despite Small Size
Despite this limitation, smaller models can still be highly effective in enterprise and industrial environments.
Rather than acting as a static repository of knowledge, modern systems augment models at inference time using techniques like Retrieval-Augmented Generation (RAG). This allows the model to access relevant, up-to-date information before generating a response.
In practice, this means:
- Reduced hallucinations
- Improved accuracy
- Seamless integration with internal document systems
- No reliance on stale training data
By offloading knowledge storage to external systems, SLMs can focus on reasoning and synthesis — where they increasingly excel.
Atlas Advisor
Atlas Advisor is our proprietary intelligence solution designed specifically for secure process control environments where privacy and cybersecurity are non-negotiable.
By combining SLMs, RAG, and agentic AI, Atlas Advisor:
- Unlocks operational insights
- Supports day-to-day decision-making
- Democratizes access to institutional knowledge
- Reduces time-to-first-action
All while orchestrating data across control loop monitoring, advanced process control, and other critical systems commonly found in the control room.
Learn more about it on our Atlas Advisor page!
