The eras of monitoring, compressed
Every operations centre carries its history in layers. The oldest layer is the clipboard: manual rounds, hourly log sheets, an operator's judgement about whether a bearing sounded wrong. SCADA replaced the clipboard with telemetry, with remote terminal units polling field devices, a master station raising alarms, and a human acknowledging them. Then came the data historian, which stopped throwing telemetry away, and the dashboard era, which put that history on screens. More recently, analytics layers began computing KPIs, efficiency curves, and availability figures on top of the historian.
Each era added visibility. None of them, on its own, added judgement. A SCADA screen tells you a value is out of range; it does not tell you which of today's forty excursions actually threatens revenue, or what to do about it. The current transition, from analytics to AI-driven operations, is the first one that changes the nature of the output: from numbers a person must interpret to ranked, explained recommendations a person can act on. Getting there is less about model selection than about a handful of architecture decisions made well before any model runs.
The data layer: historians meet time-series stores
The classic process historian was built for one job: compress high-frequency tag data and play it back for trending. It does that job well, and ripping it out is rarely the right move. The problem is what historians were never built for: cross-site queries, joining telemetry with maintenance and market data, feeding model training pipelines, serving thousands of concurrent reads.
The pragmatic pattern in most estates is coexistence. The historian remains the plant-level system of record for raw tags, while a modern time-series platform, cloud or on-premises, receives a curated, contextualised replica. The decision that matters most here is not which database to buy, but the data model. Raw SCADA tags named PLC3_AI_0047 are useless to an analytics team. An asset model that maps tags to equipment, equipment to systems, and systems to sites, increasingly delivered through open standards such as OPC UA information models or a unified namespace built on MQTT, is the single highest-leverage investment in the whole stack. Teams that skip it spend the next three years re-deriving context in every project.
Edge or cloud? In emerging markets, deliberately both
Cloud-first architectures assume bandwidth that many operating environments simply do not have. A solar plant on a constrained rural link, or a gas asset on satellite backhaul, cannot stream every tag at one-second resolution to a cloud endpoint, and should not try.
The workable pattern pushes intelligence to match connectivity. At the edge, do the time-critical and bandwidth-heavy work: local buffering through outages, high-frequency anomaly scoring, protocol translation, and store-and-forward so no data is lost when the link drops. In the cloud (or a central data centre), do the work that benefits from scale: fleet-wide benchmarking, model training, long-horizon degradation analysis. Models train centrally and deploy back to the edge for inference.
Emerging-market operators hold one genuine advantage here. Where there is no thirty-year brownfield SCADA estate to defend, greenfield assets can leapfrog straight to open protocols, publish-subscribe data movement, and edge-native designs, skipping the integration debt that older fleets are still paying down. The constraint becomes a forcing function for cleaner architecture.
Security is not a separate workstream in any of this. Every path between OT and IT should cross a controlled boundary: one-way data flows out of the control network where feasible, no analytics platform with write access to control systems by default, and segmentation aligned with recognised frameworks such as IEC 62443. An intelligence layer that weakens the control layer has negative value.
Alarms are not intelligence
The alarm system is where the dashboard era shows its limits most painfully. Industry standards are blunt about human capacity: EEMUA 191 puts a manageable steady-state load at fewer than six alarms per operator per hour, and ISA-18.2 defines an alarm flood as ten or more alarms in ten minutes, the point at which an operator can no longer process each one. Unrationalised sites routinely run far above these thresholds in normal operation, and orders of magnitude above them during upsets, which is precisely when attention matters most. The predictable result is fatigue, suppression, and missed events.
Alarm rationalization, reviewing every configured alarm for purpose, priority, and required operator action, remains unglamorous, essential groundwork. But it also exposes the deeper issue: threshold alarms fire on symptoms, one tag at a time. A machine-learning layer works differently. Multivariate anomaly detection learns an asset's normal operating envelope across dozens of correlated signals and flags deviation before any single tag crosses a limit. Instead of forty simultaneous alarms describing one root cause, the operator sees one event, with the contributing signals attached. The AI layer does not replace the alarm system; it sits above it and restores the signal-to-noise ratio that alarm inflation destroyed.
From dashboards of numbers to decisions of record
The final shift is organizational, and it is the one most modernization programmes underestimate. A dashboard estate makes watching the job; an intelligence-led operation makes deciding the job. The difference shows up in what the operations centre actually produces. A mature recommendation is not a chart. It is a structured object:
Ranked against other open issues by expected impact on energy delivered, cost, or risk
Explained, with the driving signals and confidence stated, so an engineer can interrogate it rather than take it on faith
Actionable, tied to a specific intervention and owner
Closed-loop, capturing what was done and what happened, so every decision trains the next one
That last property is what turns recommendations into decisions of record: an auditable operating history that compounds. It also demands discipline that no vendor ships. It takes clear thresholds for when a human must review, honest tracking of false positives, and a willingness to retire recommendations nobody acts on. In our framing, this is the top of a four-layer build: reliable data capture, contextualised infrastructure, analytics that explain themselves, and an operating discipline that converts intelligence into value outcomes. Skip a layer and the ones above it wobble.
What the next control room looks like
The direction of travel is visible now. Fewer wall-sized mosaics of trend lines; more queues of ranked, explained interventions. Edge inference as standard equipment rather than pilot project. Asset models rich enough that a new analytics use case is a query, not an integration programme. And operators whose scarce attention is spent on the ten decisions that matter, not the ten thousand points that don't. SCADA gave the industry its nervous system. The work of this decade is giving it judgement, and keeping a record of every judgement made, so the system, and the people running it, get measurably better at the job each quarter.
