SEC-03 / OT Security

The Purdue model and AI: where analytics and cloud fit in your plant

The Purdue model is still the clearest way to explain where things sit in a plant network and which way data should flow. Use it to place AI and cloud workloads, but design to zones and data flows, not to the level numbers.

7 min read Updated OT engineersControls engineersCISOsIT architects

Key takeaways

  • Purdue describes functional levels from physical process (Level 0) to enterprise (Level 4/5), with an industrial DMZ (Level 3.5) between IT and OT.
  • Most AI training and cloud analytics belong at Level 4/5. Edge inference usually sits at Level 2 or 3 in its own zone.
  • The rule that matters: no direct path from the enterprise or internet into Level 2 and below. Data flows up and out through the DMZ.
  • IIoT sensors and cloud-connected devices break strict layering. Handle them with zones and conduits, not by pretending they fit a level.

Every AI project in a plant eventually hits the same whiteboard argument. The data scientist wants live sensor data in the cloud. The controls engineer wants nothing new talking to the PLCs. The Purdue model is the shared language that settles that argument. This article explains the levels, shows where AI, analytics and cloud fit, and explains why you should treat Purdue as a reference rather than a rulebook.

What the Purdue model is

The Purdue model comes from the Purdue Enterprise Reference Architecture (PERA), developed at Purdue University in the early 1990s for computer-integrated manufacturing. It became the standard way to describe industrial control networks, and ISA-95 builds on the same layering for enterprise-control integration. It was never written as a security standard. It describes function: what each layer does and how close it is to the physical process.

The levels as most plants use them today:

Level Name Typical assets Time scale
0 Physical process Sensors, actuators, valves, motors, drives Milliseconds
1 Basic control PLCs, RTUs, safety controllers, remote I/O Milliseconds to seconds
2 Area supervisory HMIs, SCADA servers, line-level engineering workstations Seconds to minutes
3 Site operations Historian, MES, batch servers, site engineering, OT domain controllers Minutes to hours
3.5 Industrial DMZ Historian replica, patch and AV staging, jump hosts, brokers Buffer, no process control
4 Business planning and logistics ERP, email, office IT, site file servers Hours to days
5 Enterprise and cloud Corporate data center, SaaS, cloud analytics and AI platforms Days and longer

Safety instrumented systems (SIS) are usually drawn alongside Level 1 as their own segment. Keep them that way.

Note: Level 3.5 was not in the original PERA. Practitioners added it as firewalls and remote access made a buffer between IT and OT necessary. It is now standard practice, and DMZ-based architectures are discussed in guidance such as NIST SP 800-82 Rev. 3.

Where AI and analytics fit

AI is not one thing on a Purdue diagram. Split each use case into its parts: data collection, inference, training, and action.

 L5  Cloud AI platform: training, fleet models, dashboards
       ^ data up (HTTPS, initiated from DMZ)
       |                       work orders / reports v
 L4  ERP, CMMS, BI  <---------------------------------+
       ^                                              |
 =L3.5= DMZ: historian replica, MQTT broker,          |
       ^      model staging, jump host  --------------+
       | outbound only from OT
 L3  Historian, MES, site AI server (inference)
       ^
 L2  SCADA, HMI, edge inference box (own zone)
       ^          | advisory to operator via HMI
 L1  PLCs, SIS (separate), drives
       ^
 L0  Sensors, vision cameras, actuators

A sensible default placement:

  • Data collection stays where the data is: Level 0 to 2 devices, collected by the historian or an edge gateway.
  • Inference that needs low latency (vision reject, anomaly detection on vibration) runs on an edge device at Level 2 or 3. Give it its own zone with defined conduits.
  • Training and fleet analytics run at Level 4/5 or in the cloud, fed by replicated data in the DMZ.
  • Actions come back as advice to an operator, a work order in the CMMS or a recipe change through normal change control. Not as a cloud service writing to a PLC.

The joint CISA and partner guidance on AI in OT, published December 2025, recommends exactly this push-based pattern: get data out of OT for AI use without persistent inbound access. See securing AI in OT networks for the controls on each part.

Two rules that matter more than the level numbers

  1. No direct path from Level 4/5 or the internet into Level 2 and below. Traffic terminates in the DMZ. A session from IT into OT is a new session started from a DMZ host.
  2. Connections are initiated from the more trusted side. OT pushes data out to the DMZ. The 2026 secure connectivity principles from the UK NCSC, ASD's ACSC and partners say to initiate OT connections from inside the OT environment so no inbound ports are exposed.

If your design breaks either rule, you need a strong, documented reason and compensating controls.

Worked example: one AI use case on the diagram

Take a predictive maintenance project on a line of extrusion presses. Placed against Purdue, it looks like this:

Part Where it runs Level Data direction
Vibration and temperature sensors On the press, wired to an I/O module 0/1 Up to PLC or gateway
Edge gateway with anomaly model Panel-mounted industrial PC in its own zone 2 Reads PLC tags; publishes results up
Historian collection Site historian 3 Pulls from gateway and PLCs
Replica and broker Industrial DMZ 3.5 Receives from Level 3; nothing passes through
Model training and fleet dashboard Cloud AI platform 5 Pulls from DMZ replica
Maintenance recommendation CMMS work order 4 Created from cloud result; a planner decides
New model version Signed package staged in DMZ 3.5 to 2 Pulled by OT under change control

Notice that the only thing moving down the levels is a new model version, and it moves as a controlled change, not as a live connection. The action on the press is taken by a person, through a work order. That is the pattern to aim for with most AI use cases in a mid-market plant.

Why Purdue is a reference, not a dogma

Purdue was drawn for a world of hierarchical, wired networks. Modern plants do not look like that.

IIoT sensors with their own cloud. A wireless vibration sensor kit that reports straight to the vendor's cloud over LTE skips every level. It lives at Level 0 physically and Level 5 logically.

Unified namespace and MQTT. A unified namespace publishes data from many levels into a shared broker. Data still moves up and out, but not strictly level by level.

Cloud historians and SaaS MES. Functions that used to live at Level 3 now run in the cloud. See OT data to cloud historian for the data side.

Virtualized and converged infrastructure. One hypervisor cluster may host Level 2 SCADA and Level 3 MES. The levels still exist logically, but the boundaries are now virtual switches.

None of this makes Purdue useless. It makes it a map, not the territory. The fix is to pair it with IEC 62443 zones and conduits:

Question Purdue answers Zones and conduits answer
What does this asset do? Yes, by level Partly
How bad is it if this is compromised? Loosely Yes, by risk and target security level
Which paths are allowed? Implied by adjacency Yes, every conduit is defined
Where does a cloud-connected sensor go? Awkwardly Its own zone with an external conduit
Can two lines at Level 2 be separated? Not addressed Yes, separate zones

Plant-floor reality: In brownfield plants, Level 1 to 3 are often one flat network with a firewall to the office. Do not wait for a perfect Purdue redesign before doing anything. Put a DMZ between IT and OT first, then segment the most critical lines, then refine.

What good looks like

  • One current drawing of the plant network showing levels, zones and every external connection, including OEM and AI vendor links.
  • An industrial DMZ where all IT and cloud traffic terminates.
  • AI inference devices in their own zones with defined conduits.
  • No cloud service with an inbound path below the DMZ.
  • IIoT and vendor cloud devices documented as exceptions with their own zone and a risk owner.

Effort and cost

Drawing an accurate Purdue and zone diagram for one site takes a few days of an OT engineer's time plus a walk-down, and it always finds surprises. Building a basic DMZ with a pair of firewalls and a few servers is a modest capital project for a mid-market site, usually scheduled around a shutdown. Segmenting individual lines is incremental and can be done line by line during maintenance windows.

Common mistakes

  • Drawing a perfect Purdue diagram that does not match what is actually cabled. Validate with traffic captures or your OT asset inventory tool.
  • Putting the AI server "at Level 3" on the same flat network as the PLCs.
  • Using the DMZ as a routed pass-through instead of a place where sessions terminate.
  • Forcing an LTE-connected IIoT kit into the Purdue diagram and ignoring its cloud path.
  • Treating Purdue compliance as the goal. The goal is controlled risk.

Monday-morning checklist

  1. Pull the latest network drawing for your main site. Mark each device's Purdue level.
  2. Highlight every connection that crosses from Level 4/5 into Level 3 or below without stopping in a DMZ.
  3. List every AI, analytics or IIoT system and draw its data path on the same diagram.
  4. Pick the single worst path and close or broker it this quarter.
  5. Start a zones and conduits worksheet using the Purdue levels as the first draft of zones.

Sources and further reading

  1. NIST SP 800-82 Rev. 3, Guide to Operational Technology (OT) Security
  2. ISA/IEC 62443 Series of Standards (ISA)
  3. Secure connectivity principles for operational technology (ASD's ACSC, NCSC-UK and partners, Jan 2026)
  4. Principles for the Secure Integration of Artificial Intelligence in Operational Technology (CISA and partners, Dec 2025)

This guide is general information for manufacturers, not legal advice. Confirm regulatory obligations with qualified counsel.

FAQ

Frequently asked questions

What is Level 3.5 in the Purdue model?

Level 3.5 is the industrial DMZ, a buffer network between site operations (Level 3) and the enterprise (Level 4). It was not in the original Purdue reference architecture; practitioners added it. Systems such as historian replicas, patch servers, remote access jump hosts and MQTT brokers sit there so IT and cloud users never connect directly into OT. Traffic should terminate in the DMZ rather than pass through it.

Is the Purdue model obsolete because of cloud and IIoT?

No, but it is incomplete. Cloud services, wireless sensors that report straight to a vendor cloud and unified namespace architectures do not fit neatly into one level. Purdue remains useful for describing function and consequence. For security design, combine it with IEC 62443 zones and conduits, which group assets by risk and define every allowed communication path.

Where should an AI predictive maintenance system sit in the Purdue model?

Split it. Sensors and any edge gateway sit at Level 1 or 2. Inference that needs low latency can run on an edge device at Level 2 or 3 in its own zone. Training, dashboards and fleet-wide models belong at Level 4 or in the cloud. Data should move up through the DMZ, and maintenance recommendations come back as work orders, not as direct commands to controllers.

Does IEC 62443 require the Purdue model?

No. IEC 62443 uses zones and conduits based on risk, not the Purdue levels. Many plants use Purdue as a starting point to draw their first zones because it maps well to function. The 62443-3-2 process then refines those zones by consequence, criticality and connectivity, which can result in several zones at the same Purdue level.

Related guides

Know where your plants stand

Score your AI governance, OT security and data readiness in about 10 minutes each. Results stay in your browser.