Table of Contents
The Purdue model sorts a plant's systems into levels, from field devices at Level 0 to business IT at Levels 4 and 5, with a DMZ between operations and IT. Predictive machine learning usually fits in the operational Levels 1 to 3. Large language models (LLMs) and AI agents belong at Levels 4 and 5, working on data pushed out of OT.
Key takeaways
- Joint guidance from CISA and eight partner agencies says predictive ML is typically used in operational Levels 0 to 3, while LLMs are typically used at Levels 4 and 5, potentially on data exported from the OT network (CISA et al., 2025).
- The same guidance prefers push-based or brokered architectures that move data out of OT without persistent inbound access.
- NIST recommends firewall rules that allow connections only between adjacent levels, so Level 4 systems never talk directly to Levels 2, 1 or 0 (NIST SP 800-82r3).
- Where a model runs and what its output can change are separate questions. AI that acts on the process without a human in the loop should be limited.
- The Purdue model is a reference, not a security standard. ISA/IEC 62443 zones and conduits are what you use to specify the controls.
What is the Purdue model?
The model comes from the Purdue Reference Model for Computer Integrated Manufacturing, published by the Instrument Society of America in 1989 (NIST SP 800-82r3). The original document runs to about 250 pages and is still free to download (PERA). The ISA-95 standard grew from the related Purdue Enterprise Reference Architecture, first published by ISA in 1992.
More than thirty years later, the 2025 joint guidance on AI in OT still calls it "a widely accepted framework for understanding the hierarchical relationships between OT and IT devices and networks" (CISA et al., 2025). That guidance uses the version described by the SANS Internet Storm Center, summarised below with MES added from the CISA table:
| Level | Name | Typical systems |
|---|---|---|
| 5 | Enterprise network | Corporate data centre services: directory, email, billing |
| 4 | Business network | Site IT: workstations, file and print servers, phone systems |
| DMZ | IT/OT boundary | Services that broker data and access between the two sides |
| 3 | Site-wide supervisory | Historians, MES, engineering workstations, analytics |
| 2 | Local supervisory | SCADA, DCS operator stations, HMIs, local control room |
| 1 | Local controllers | PLCs, RTUs, controllers |
| 0 | Field devices | Sensors, actuators, transmitters, smart field devices |
The SANS description adds that direct internet access should not be granted beyond Level 4, and that Level 2 processes should be segregated by function, type or risk (SANS ISC, 2023).
Is the Purdue model a standard?
No. It is a reference that helps people agree on where a system sits. Two standards build on the same idea.
ISA-95, also published as IEC 62264, defines Levels 0 to 4 and "primarily deals with the interface between levels 3 and 4", that is, between manufacturing operations and business systems such as ERP (ISA).
ISA/IEC 62443 is the security series for industrial automation and control systems, written for asset owners, integrators, product suppliers and service providers (ISA). Part 3-2 requires partitioning a system into zones and conduits and setting a target security level for each. It defines a conduit as a logical grouping of communication channels that share common security requirements and connect two or more zones (ISA-62443-3-2 preview).
NIST lists the Purdue model, ISA-95 levels and a three-tier IIoT architecture as options for organising OT network segmentation, "or a combination of these models" (NIST SP 800-82r3). In practice, the Purdue levels give everyone a shared map, and 62443 zones and conduits turn that map into enforceable rules.
Which kind of AI belongs at which level?
The CISA-led guidance maps example AI uses onto the levels. The table below follows its Table 1, with plant examples added.
| Level | AI uses listed by CISA | Plant example |
|---|---|---|
| 0: Field devices | Source of the OT data used to train mainly predictive models, or to spot significant deviations | Transmitter and analyser readings |
| 1: Local controllers | Some modern PLCs or edge controllers run lightweight, pre-trained models for local anomaly detection, load balancing or maintaining a known safe state | A vision classifier rejecting defective parts on a line |
| 2: Local supervisory | Models analyse SCADA or DCS data to detect early signs of equipment anomalies and alert operators | A soft sensor running in the DCS |
| 3: Site-wide supervisory | Predictive maintenance on aggregated historian data; recommendations that support operator decisions | Anomaly detection, virtual flow metering |
| DMZ | None listed | Data is staged here on its way out, not analysed |
| 4 and 5: Business networks | Workflow optimisation with AI agents and LLMs; analysis of OT data together with IT data | A document assistant, text-to-SQL, a CMMS agent |
Source: CISA et al., 2025.
Two points are easy to miss. First, models at Levels 1 and 2 are part of the control system. They are owned by control engineers and change under the same procedures as control logic. A general-purpose AI server does not belong there; our guide to integrating AI with SCADA and OT systems explains why.
Second, the level where a model runs is not the level its output acts on. A model at Level 3 or 4 can still influence a valve if its recommendation is applied at an HMI. That is why the next two questions matter.
Where should an on-premise LLM sit?
For most plants, at Level 4 or in a dedicated zone attached to it. Three reasons:
- Its users and documents are there. Engineers, planners and maintenance staff work on the business network, and the manuals, procedures, work orders and lab reports an assistant answers from live in document management, ERP and LIMS systems.
- It is an IT system to maintain. An inference server, its libraries and its GPU drivers need regular security updates. That fits the patching rhythm of the business network, not the control network.
- It does not need live control data. When it needs process context, a summary or a historian extract pushed through the DMZ is enough. It should never query a PLC or an operator station.
Running the LLM on your own servers does not make it safe by default; access control, logging and prompt injection still need work, as covered in securing an on-premise LLM. For sites where the whole system must run with no external network at all, see air-gapped AI deployment.
How should plant data reach AI systems?
These rules come from the CISA-led guidance and NIST SP 800-82r3:
- Push data out; do not pull it in. Use push-based or brokered architectures that move features or summaries out of OT without persistent inbound access, so the AI system is not a standing attack path into OT (CISA et al., 2025).
- Connect adjacent levels only. Firewall rules should prevent Level 4 devices from communicating directly with Levels 2, 1 or 0, and outbound rules from the control network should be as strict as inbound ones (NIST SP 800-82r3).
- Go through the DMZ. NIST's example architecture routes all communication between the enterprise and operations management levels through services in the DMZ, and asks that those services be monitored and protected.
- Use one-way transfer where you can. Where data must reach business networks, CISA recommends one-way transfer patterns and audited staging buffers. NIST notes that a unidirectional gateway between Levels 2 and 3 can protect Levels 0 to 2 from a cyber event at Levels 3 to 5.
- Make every read authenticated and read-only. OPC UA identifies each client and server with X.509 certificates and can encrypt messages (OPC Foundation). Give AI-related accounts read access only.
- Keep separate credentials. NIST recommends separate authentication mechanisms and credentials for users of the corporate and OT networks.
- Treat the AI system like any new OT system. Test it for safety impacts such as latency, try it on test infrastructure before production, and route any new connection through the existing DMZ or jump host (CISA et al., 2025).
What if the AI needs to change the process?
Start with a recommendation, not a write. The CISA-led guidance advises limiting active control of OT by AI without a human in the loop, citing safety and latency, and asks for failsafe mechanisms that revert to traditional automation or manual operation (CISA et al., 2025). In practice, a model at Level 3 or 4 proposes a setpoint change, and an operator applies it at Level 2 through the usual management of change. Our article on human-in-the-loop for AI agents covers where to put approval gates.
The process industry has formalised a similar separation. The NAMUR Open Architecture (NOA), described in NAMUR Recommendation NE 175, sends data for monitoring and optimisation over a second communication channel without affecting the automation system (NAMUR). NAMUR lists a separate recommendation for verifying requests sent back to the core system, NE 178, as planned. In one vendor's implementation, the NOA edge device reads HART data through a second channel and "does not allow feedback or modification of field device data" (Beckhoff).
Does cloud AI break the Purdue model?
It adds paths the model did not foresee. A cloud AI service needs plant data to leave the site, which runs against the SANS advice to stop direct internet access at Level 4. The CISA-led guidance does not rule cloud out, but asks owners to define AI security responsibilities in cloud contracts and to keep control over critical functions whether AI is hosted locally or in the cloud. NIST's mention of a three-tier IIoT architecture reflects the same shift.
Whatever the hosting, the method stays the same: put each component on the map, list every data flow and assign each one to a conduit with an owner. For the hosting choice itself, see sovereign cloud vs private cloud vs on-premise.
What we have seen at KABAS AI
DALILE runs completely offline on the client's own servers and connects to systems such as ERP, LIMS, MES, SQL databases, SAP and Microsoft 365. Placing it in a plant network raises the questions above, so we recommend three steps before a pilot.
Draw the levels first. For each source the assistant or model will use, write down its Purdue level, its owner and the path its data will take. A source with no approved path is a finding, not a detail.
Reuse the paths you already have. If the site already sends historian data or reports to business users through the DMZ, the AI system should use that path rather than a new one.
Decide where predictive models run with the control team. A model that only advises can run at Level 3 or 4. A model that feeds the control system belongs to the control system's change process.
Our information systems team maps and integrates these systems, and our data science team builds predictive models and deploys them on your infrastructure.
Sources
- CISA, ASD's ACSC and partners, Principles for the Secure Integration of Artificial Intelligence in Operational Technology, 3 December 2025
- NIST SP 800-82r3, Guide to Operational Technology (OT) Security, September 2023
- Manuel Humberto Santander Pelaez, Controlling network access to ICS systems, SANS Internet Storm Center, July 2023
- PERA, The Purdue Reference Model for CIM
- ISA, ISA-95 standard: Enterprise-Control System Integration
- ISA, ISA/IEC 62443 series of standards
- ISA, ANSI/ISA-62443-3-2-2020 preview
- OPC Foundation, Unified Architecture
- NAMUR, NAMUR Open Architecture
- Beckhoff, NAMUR Open Architecture (NOA)
Frequently asked questions
What is the Purdue model in simple terms?
It is a reference model that sorts a plant's systems into levels by function: field devices at Level 0, controllers at Level 1, supervisory systems such as SCADA and DCS at Level 2, site-wide systems such as MES and historians at Level 3, and business IT at Levels 4 and 5. A DMZ usually separates the operational levels from business IT. Teams use it to decide which systems may talk to which.
Is the Purdue model still relevant with IIoT and cloud services?
Yes, as a shared map, though not as the only option. The December 2025 joint guidance led by CISA calls it still a widely accepted framework, and NIST SP 800-82r3 lists it alongside ISA-95 levels, a three-tier IIoT architecture, or a combination. Cloud and IIoT services add connections the original model did not foresee, so each new path still needs to be mapped and controlled.
Can an AI model write setpoints directly to a DCS or PLC?
Technically it can, but current guidance advises against it. CISA and its partner agencies recommend limiting active control of OT by AI without a human in the loop, and having failsafe mechanisms that revert to traditional automation or manual operation. The common pattern is a recommendation that an operator reviews and applies through the existing control system.
Where should an on-premise LLM be installed in a plant network?
In most plants, on the business side, at Level 4, or in a dedicated zone attached to it, not in the control network. That is where its users and documents are, and where it can be patched like other IT systems. If it needs process data, that data should be pushed out of OT through the DMZ rather than queried from control systems.
What is the difference between the Purdue model, ISA-95 and ISA/IEC 62443?
The Purdue model is the original reference architecture that sorts systems into levels. ISA-95, also published as IEC 62264, grew from it and defines the interface between business systems and manufacturing operations, mainly between Levels 3 and 4. ISA/IEC 62443 is a series of security standards that partitions a system into zones and conduits and sets a target security level for each.



