The Question That Started It
A question I was asked recently: “What’s the future of OT?”
My answer was clear—AI. Not the hype-cycle AI that fills conference slides, but intelligence embedded deep within our process layers—Levels 1 and 2—where control logic and physics intersect. And that, precisely, is where the next major gap in our industrial architecture will appear.
AI is already here, but our plants weren’t built for it.
Two Dreams, Two Scars
OT has been sold two big dreams—and both have left scars.
First came asset discovery. It promised instant visibility and risk assessment “out of the box.” Reality proved harsher: incomplete detection, misclassified assets, and false confidence that sometimes made plants less secure.
The truth asset owners discovered later was simple: Visibility demands re-architecting networks, normalizing protocols, and accepting that passive discovery alone can never capture everything. Even when data was gathered, it wasn’t actionable without human interpretation.
Then came GenAI and LLMs. They were marketed as the “silver bullet” that would automate everything—from documentation to risk scoring. But again, hype met the hard walls of process engineering. AI models trained in cloud environments struggled with air-gapped systems, latency constraints, and the sheer heterogeneity of control platforms. Startups raced to build “magic” copilots; most built demos.
Engineering Reality: AI in the Plant We Actually Have
AI in industry is long past the hype cycle, yet the operational landscape remains brownfield. Legacy DCS and PLC networks, serial links, and 20-year lifecycles define reality. Under those conditions:
- Edge comes first. Cloud is optional.
- Offline is normal.
- Latency is a constraint, not a metric.
Intelligence must move toward the process, not away from it. Adaptors, not overhauls. Incremental improvement, not architectural fantasy.
Private 5G, Autonomy, and Control Boundaries
Private 5G has its place—in mobility, telemetry, and video systems—but it will never replace deterministic wired control. Closed-loop, safety-critical functions demand timing, not throughput. Keep control on wires or TSN. Use wireless for visibility, not actuation.
Autonomy itself should mature through staged authority:
Autonomy must be earned—loop by loop, audit by audit.
Model Selection: Ask the Right Question First
Before choosing an LLM or SLM, ask: Will a model even move this KPI?
Many decisions in OT can still be solved by control theory, not neural networks. If AI is justified, apply engineering discipline:
- What class of decision is this—advisory or control?
- What’s the latency budget?
- Can data leave the site?
- What’s the tolerance for wrong answers?
- What’s the lifecycle cost—tuning, drift, retraining?
Use the smallest model that improves reliability or efficiency without compromising safety. SLMs suit local inference at the edge—interpreting alarms, logs, SOPs. LLMs fit enterprise-scale search and RCA when latency and data-sovereignty allow.
Protection Before Prediction
For all the excitement around AI detection, protection remains the harder—and more important—discipline. I predicted in 2024 that industrial cybersecurity would shift from detection-first to protection-first by 2026. That transition is happening.
Asset owners are realizing that no algorithm can compensate for a flat network or a missing safety interlock. AI may detect deviation, but only architecture prevents consequence. Protection must precede prediction—always.
Reference Architecture: Edge-Native, Governed, and Safe
Every autonomous or AI-assisted control function must be designed for failure, audit, and rollback.
- Architecture: Agents at L1-L3, validated in digital twins before deployment, with store-and-forward and last-known-good states.
- Governance: Signed policies, kill switches, staged promotion (sandbox → supervised → autonomy).
- Safety/Cyber: ISA/IEC 62443 zoning, tamper-evident logs, least-privilege service accounts, strict SIS/SIF segregation.
Autonomy without verifiability is negligence.
Field Reality: Maturity Is Rising
Across industries, asset-owner maturity has improved:
- Fewer “tool-only” deployments; more architecture reviews.
- Expanded cooperation between engineering, IT, and cybersecurity teams.
- Movement from compliance-driven to consequence-driven controls.
This maturity comes from experience—not marketing—and it signals a quiet cultural shift: from quick fixes to sustained engineering.
Hype Cycles and Hard Lessons
Both hype waves—asset discovery and GenAI—taught the same lesson: there are no shortcuts in OT.
OT security will remain engineering-led and consultancy-driven for the next decade. The diversity of systems, long equipment lifecycles, and safety dependencies ensure it. Solutions that endure will be those that reduce field time, respect process realities, and integrate with plant operations—not those that chase trends.
An asset in OT is not defined by its IP address; it’s defined by its function. And expertise is not defined by vocabulary; it’s proven by time spent inside plants.
Procurement and Trust
Procurement processes must evolve to favor transparency over promises:
- Ask vendors for configuration export, change journals, and full audit trails.
- Tie payment to operational KPIs, not licenses.
- Require demonstrable safety boundaries and rollback.
- Interview partners on real site experience, not slide decks.
Trust in OT isn’t built in sales—it’s built in shutdowns, startups, and night shifts.
Closing Reflection: The Architecture of Reality
The marketing cycles will continue. Silver bullets will keep reappearing—asset discovery, GenAI, quantum, whatever comes next. But the physics of plants will not change.
Hype fades. Engineering endures. Protection before prediction. Design before deployment.
AI will have a place in OT—but only when it behaves like a great operator: conservative, prepared, transparent, and measurable.
The future of OT will be intelligent, yes—but it will still belong to engineers who understand the plant we actually have.