Manufacturing AI deployment puts the model inside the plant, on the IT side of a properly segmented network, where it can reach maintenance histories, quality records, and standard operating procedures without any of that leaving the site. Two constraints drive the architecture: process knowledge is trade secret material whose value depends on never being disclosed, and anything touching operational technology has to respect the zone-and-conduit segmentation that keeps the line safe. Neither is an obstacle — both are reasons the deployment looks the way it does.
Why this is hard
Decades of process knowledge live in maintenance notes, shift logs, and the heads of people approaching retirement.
Trade secret protection depends on demonstrable reasonable measures — uploading specifications to a public service undermines exactly that.
The plant network was segmented deliberately, and most AI tooling assumes a flat network with internet access.
What the deployment gives you
Institutional knowledge, made searchable
Retrieval over maintenance histories, deviation reports, shift handovers, and equipment manuals, so a technician on nights gets the answer a thirty-year veteran would have given. This is the use case that reliably pays for the hardware.
Correctly placed in the zone model
Inference sits on the IT side, reading from the historian and maintenance system through defined conduits. The model informs people; it does not reach into control systems, and the segmentation that keeps the line safe stays intact.
Trade secrets stay demonstrably protected
Process parameters, formulations, and tooling specifications never leave the site. Keeping the data inside the plant is itself part of the reasonable-measures argument that trade secret protection depends on.
SOP and quality documentation support
Draft work instructions, summarise deviations, and answer questions grounded in the controlled document set — with the source revision cited, so a quality system that depends on document control is not undermined by an unsourced answer.
Where inference belongs in a segmented plant
The zone model already in place is the design. Inference lives in the IT zone with read access through defined conduits to the systems that hold the data; it never sits on the control network and never writes to it. The value comes from reading operational history, not from touching operations.
Read-only access to operational data through defined conduits; the model must not be able to write to any control system, directly or indirectly.
Cite the document revision on every answer — a quality system built on document control is undone by an answer from a superseded SOP.
Keep the deployment advisory. Language models support decisions about the process; they do not belong in a safety or control loop.
Treat the corpus as trade secret material: access-controlled, logged, and inside the plant boundary as a matter of network configuration.
Include the AI system in the change management process that already governs plant IT, so a model update is a controlled change.
What you have to satisfy
IEC 62443-3-3 — Zones and conduits: The system security requirements that define how plant networks are segmented — an AI deployment has to be placed in a zone and reach data through explicit conduits.
ISA-95 levels — IT/OT layering: Inference belongs at the enterprise and MES layers reading downward, not at the control layers. Getting this placement right settles most architecture arguments early.
NIST SP 800-82 — ICS security guidance: Practical guidance for securing industrial control systems, including the segmentation and access patterns an added system must respect.
18 U.S.C. §1839(3) — Trade secret — reasonable measures: Protection depends on taking reasonable measures to keep information secret; where process data is processed is directly relevant to whether that test is met.
Models that fit this deployment
Model
VRAM (Q4)
Runs on
Context
License
Qwen 3 14B Single-plant tier — 12 GB card — Maintenance and SOP retrieval for one site on modest hardware in the plant IT rack. ollama pull qwen3:14b
Qwen 3 32B Site tier — 24 GB card — The practical default: grounded Q&A over historian data, manuals, and shift logs. ollama pull qwen3:32b
20 GB
24 GB GPU (RTX 3090/4090) Mac: 32 GB unified
128,000
Apache 2.0
Llama 3.3 70B Instruct Multi-site tier — 48 GB class — Where a central deployment serves several plants and document sets. ollama pull llama3.3:70b
43.1 GB
2×24 GB GPUs or 48 GB card Mac: 64 GB unified
128K
Llama Community
Unlocking plant knowledge without exporting it?
We help manufacturers deploy AI inside the plant boundary — corpus design over maintenance and quality records, correct placement in the zone model, hardware sizing for plant conditions, and grounding that respects document control. Vendor-neutral and practitioner-level.
What is the most valuable manufacturing AI use case to start with?
Making institutional knowledge searchable. Maintenance histories, deviation reports, shift handovers, and equipment manuals hold decades of hard-won troubleshooting knowledge that is effectively inaccessible under time pressure, and that walks out the door as experienced staff retire. A retrieval system over that corpus gives a technician on a night shift the answer a veteran would have given, and it pays back faster than any predictive-maintenance project because the data already exists and needs no new instrumentation.
Can an AI system touch the control network?
It should not, and the segmentation you already have is the reason. Language models belong on the IT side, reading operational history through defined conduits and producing advice for people. Anything that writes to a control system crosses a boundary that exists for safety reasons and brings the deployment into scope of a very different assessment. Keeping inference advisory and read-only is not a limitation of the technology so much as the correct placement of it in a plant that has been designed properly.
How do we keep process knowledge from leaking into a vendor's systems?
Run the model where the data is. Trade secret protection depends on demonstrating reasonable measures to preserve secrecy, and sending formulations or process parameters to an external service is difficult to reconcile with that — regardless of the contract terms. On-premise inference with no outbound route from the model host makes the answer a matter of network configuration rather than vendor trust, which is a much stronger position if the question is ever litigated.
Does an AI assistant interfere with a quality management system?
Only if you let it answer without citing the document revision. Quality systems depend on people working from the controlled version in force, so an assistant that produces a confident answer from a superseded SOP is an audit finding. Ground every answer in the controlled document set, show the document identifier and revision alongside the answer, and re-index when documents change. Done that way, the assistant strengthens document control by making the current revision easier to find than the printout on the wall.
What hardware does a plant need?
Modest, and it fits in the IT rack you already have. A single 12–24 GB GPU running a 14B or 32B-class model handles retrieval and drafting for a site, because usage is intermittent — technicians ask questions in bursts rather than continuously. Multi-site deployments consolidating several document sets justify a 48 GB card. Environmental fit matters more than raw performance here: power, cooling, and dust in a plant IT room constrain the choice more than benchmark scores do.