A UK-based water utility called us last autumn after their third pumping station failure in six months. Each time, the same pattern: a bearing or seal degrades over weeks, nobody spots it until the pump trips, and then it's an emergency call-out, a portable pump hired at weekend rates, and a compliance report to Ofwat explaining what happened. The equipment cost to fix each failure was under £3,000. The combined cost of the three incidents — call-outs, hire equipment, penalties, overtime — was north of £140,000.
That story is not unique to water. If you run any operation with critical physical assets — manufacturing lines, building services, fleet, utilities infrastructure — you have your own version. A food manufacturer in Cheshire whose palletiser gearbox seized mid-shift. A facilities manager in Manchester whose chiller compressor failed the week before a heatwave. The economics are always the same: the part is cheap, the downtime is ruinous.
The sensor-and-ML side of predictive maintenance — fitting vibration sensors, collecting telemetry, training anomaly detection models — has been well understood for years. It works. But it only solves half the problem. Knowing that a bearing is degrading is useful. Having an AI agent that detects the degradation, assesses the severity, checks your spare parts stock in Dynamics 365, finds the best maintenance window in the production schedule, creates a work order, assigns the right engineer, and captures the resolution so the system gets smarter over time — that is a different capability entirely.
At Veriland, we call this agentic predictive maintenance. Sensors and ML are the foundation. The AI agent layer is what turns detection into autonomous action. This article is an honest walkthrough of how the whole system works, what it costs, and where people get tripped up.
Most businesses we talk to are stuck on one of the first two approaches. Reactive maintenance — run it until something fails — is the default for a lot of UK operations, not because anyone thinks it's wise, but because nobody has had the bandwidth to set up anything better. It is the cheapest approach right up until it is not. Emergency call-outs, air-freighted parts, overtime to catch up on lost production or service levels. One bad failure and you have spent more than a year of planned maintenance.
Preventive maintenance is the sensible middle ground that most well-run operations end up at. Service the equipment on a schedule. Replace wear parts at fixed intervals. It works, broadly. But it is wasteful in ways that are hard to see: you are pulling perfectly good bearings because the calendar says so, scheduling shutdowns for machines running fine, and still getting caught out by the failures that do not respect the schedule — which, in our experience, is most of them.
Predictive maintenance — fitting sensors and using ML to spot when equipment drifts from normal — fixes the detection problem. You know something is wrong, weeks before it becomes a crisis. But here is what most implementations miss: detection is only the first half. The alert still lands in someone's inbox. Someone still has to interpret it, check if the spare part is in stock, find a maintenance window that does not disrupt production, assign the right engineer, and write up what was done afterwards. If your best maintenance engineer is off sick, on holiday, or has left the company, that knowledge gap can turn a predicted failure into a reactive one anyway.
The step most vendors skip: Sensors tell you a bearing is degrading. ML models tell you how fast. But who checks the parts stock? Who finds the maintenance window? Who assigns the engineer? And who captures what was done so the system learns for next time? That orchestration layer — what we call agentic predictive maintenance — is the difference between a monitoring system and an autonomous maintenance capability.
We have implemented this across manufacturing plants, pumping stations, building services, and processing facilities. The specifics change but the architecture does not. There are two distinct layers, and both matter.
Not every asset. You pick the three to five that would hurt the most if they failed unexpectedly — your bottleneck machine, the pump with an eight-week lead time on the main seal assembly, the ageing compressor that your maintenance manager worries about. The sensors are straightforward: vibration on motor housings and bearing blocks, temperature probes on known hotspots, current clamps on drive feeds, pressure transducers where relevant. Wireless, battery-powered, retrofitted without stopping the line or shutting down the asset.
Off-the-shelf sensor hardware for a five-asset pilot typically costs between £5,000 and £8,000 depending on measurement points. Veriland also has in-house capability to design purpose-built sensor circuits for specific asset types, which can reduce the hardware cost and improve reliability in harsh environments — but standard commercial sensors work well for most pilots.
Sensors push readings over Wi-Fi or LoRaWAN into Azure IoT Hub, which handles device registration, security, and reliable delivery even when connectivity is intermittent. Azure Stream Analytics sits behind it, applying immediate threshold rules while routing everything into storage for the ML models. Getting this plumbing layer right matters — we have seen projects from other consultancies that skipped it and spent months debugging data gaps.
You need four to eight weeks of baseline data before the ML side is useful. The models need enough history to learn what "normal" actually looks like for each specific asset under your specific operating conditions.
Azure Machine Learning learns the healthy signature of each asset — vibration profile, temperature curves, current draw at different loads. Each asset gets its own model because no two machines behave identically, even if they are the same make and model. Age, installation, alignment, and load profile all matter.
You do not need data scientists for this. Microsoft's AutoML selects and tunes algorithms automatically. We handle the feature engineering and the false-positive tuning — which is the part that takes the most iteration. Nobody wants their maintenance manager getting pinged at 3am because a sensor misread an ambient temperature spike.
This is where the architecture diverges from a standard IoT monitoring setup. When the ML model flags an anomaly, it does not just fire an alert. It triggers an AI agent built on Microsoft Copilot Studio and Azure AI that orchestrates the entire maintenance response.
The agent assesses the anomaly against the asset's history and known failure patterns. It checks spare parts availability in Dynamics 365 — whether you are on Business Central with MaxWAM or Finance & Operations with its native Asset Management module. It identifies the best maintenance window by reading the production schedule or operational calendar. It creates a work order with the right priority, assigns the appropriate engineer based on skills and availability, and attaches the full diagnostic context — sensor trends, model confidence, recommended actions — so the engineer arrives prepared.
If the anomaly falls outside the agent's confidence threshold — an unfamiliar pattern, a high-cost decision, a safety-critical asset — it escalates to a human with full context rather than acting autonomously. The human's decision is captured and fed back into the system, so the agent handles similar situations itself next time. This is how institutional maintenance knowledge accumulates in the system rather than existing only in the heads of your most experienced engineers.
Every maintenance event — whether handled autonomously or escalated — feeds back into the system. What the fault was, how it was diagnosed, what was done to fix it, how long it took, which parts were used. Over time, this builds up a structured maintenance knowledge base that is independent of any individual engineer.
This matters more than most people realise. The average age of maintenance engineers in UK manufacturing and utilities is climbing. When an experienced engineer retires, decades of "I know that motor sounds wrong" knowledge walks out the door. Agentic predictive maintenance captures that knowledge systematically — not as documents nobody reads, but as decision patterns the agent uses actively.
A real example: A packaging client in Cheshire had a conveyor drive that the model flagged on a Tuesday. The vibration pattern on the output shaft bearing had been trending upward for nine days. Their maintenance team initially dismissed the alert — the line was running fine. The agent escalated it a second time with updated severity scoring and a parts-availability check showing the replacement bearing was in stock. They pulled the bearing at the weekend: outer race visibly pitted, maybe a week from seizure. Cost: £85 and forty minutes. If it had seized mid-run: two days of lost output. The agent captured the entire sequence — initial dismissal, re-escalation trigger, resolution — so that vibration pattern is now handled at higher confidence without needing human confirmation.
The sensing and ML pipeline — Azure IoT Hub, Stream Analytics, Machine Learning — is mature, well-documented, and offered by every Microsoft partner who does IoT work. That is table stakes. What changes the capability is what sits on top.
Azure IoT Hub is the front door for sensor data. Device registration, security certificates, reliable delivery over intermittent connections. Azure Stream Analytics processes data as it arrives — applying immediate threshold rules while routing everything into storage for the ML models. Azure Machine Learning runs the anomaly detection, with AutoML handling algorithm selection and Veriland managing the feature engineering and false-positive tuning.
The orchestration layer is where Veriland's architecture differs. We build maintenance agents using Microsoft Copilot Studio for the reasoning and decision logic, extended with Azure AI Foundry services for more complex scenarios. The agent connects to Dynamics 365 — Business Central with MaxWAM for asset management and work orders, or Finance & Operations with its native Asset Management module extended by MaxWAM's Remaining Useful Life tracking. Power BI provides dashboards for the humans who want to understand the trends, and Microsoft Teams is where escalation notifications and human-in-the-loop decisions happen.
The reason this matters is that the entire chain — from sensor reading to anomaly detection to autonomous work order creation to engineer notification to resolution capture — lives inside one ecosystem. No integration middleware between the IoT platform and the ERP. No separate ticketing system. No manual step where someone has to copy an alert from one screen into another. When you add an agent layer on top of a fragmented stack, you spend half the project budget on glue code. When the stack is unified, the agent just talks to the systems it already lives in.
On cost: Because these are consumption-billed Azure services, running cost scales with what you are monitoring. A five-to-ten asset deployment — including IoT Hub, ML inference, and agent execution — typically runs at a few hundred pounds a month. The agent layer adds minimal incremental cost because the heavy compute is in the ML models, not the orchestration logic.
The operational numbers are what you would expect from any well-implemented predictive maintenance system: unplanned downtime drops by a third to a half in the first year, maintenance spend becomes predictable rather than spiky, and equipment life extends by 20-30% because you catch cascading damage early. Those results come from the sensing and ML layers. They are real and they matter.
But the changes that come from the agent layer are different in kind, not just degree.
Response time collapses. In a traditional setup, an alert might sit in an inbox for hours before someone interprets it, checks parts, and creates a work order. With an agent, the work order exists within minutes of the anomaly being flagged — with parts checked, engineer assigned, and maintenance window identified. The detection-to-action gap goes from hours or days to minutes.
Knowledge stops being a single point of failure. Every resolution, every human decision, every override feeds back into the system. When your most experienced maintenance engineer retires, the knowledge they built over decades does not retire with them. It lives in the agent's decision patterns and the structured knowledge base. New engineers have access to the full diagnostic history of every asset from day one.
The maintenance team's work changes character. Instead of firefighting and paperwork, engineers spend their time on the genuinely skilled work — the edge cases the agent escalates, the complex diagnostics, the root cause analysis. Two clients have told us unprompted that they have found it easier to retain maintenance engineers since the switch. When the job shifts from reactive crisis management to planned, skilled intervention, people stay.
Compliance becomes a by-product. For utilities, regulated manufacturing, or any operation with audit requirements, every maintenance action is logged with full provenance — what triggered it, what the agent recommended, what was done, and by whom. Audit trails that previously required manual documentation happen automatically.
Start with a five-asset pilot. We have had clients come to us wanting to sensor 40 machines or every pump in their network on day one, and we talk them down every time. Pick the assets that would hurt the most if they failed — your bottleneck machine, the pump with the long lead time on spares, the compressor your maintenance manager loses sleep over.
Budget four to eight weeks for the baseline period after sensors go in. The ML models need this time to learn what normal looks like under your specific operating conditions. You cannot shortcut this — models trained on too little data produce false positives, and false positives erode trust faster than anything else.
Connect it to Dynamics 365 from the start, and configure the agent layer alongside the ML, not as a later phase. A predictive maintenance system without workflow integration is a monitoring dashboard. A monitoring dashboard without an agent layer is something your team checks for the first fortnight and then forgets about. We learned this early: the projects where detection and orchestration were delivered together saw adoption and results within months. The ones where "phase two" was the workflow integration often never got to phase two.
If you are already on Business Central or Finance & Operations with Azure, you can realistically go from scoping conversation to live agentic monitoring on pilot assets in about ten weeks. If you are not on Dynamics 365 yet, our Business Central Accelerator or F&O Accelerator gets you onto the platform at a fixed price, and the predictive maintenance layer can follow once your core ERP is bedded in.
Traditional IoT monitoring detects anomalies and sends alerts — a human still has to interpret the alert, check parts, find a maintenance window, and create a work order. Agentic predictive maintenance adds an AI orchestration layer that handles the entire response: interpreting the anomaly, checking spare parts in Dynamics 365, identifying the best maintenance window, creating and assigning a work order, and escalating to a human only when the situation falls outside its confidence threshold.
Any operation with critical rotating or mechanical equipment. Veriland has implemented this across manufacturing plants, water and utilities infrastructure, facilities with critical building services, and operations with high-value mechanical assets. The sensing principles are the same — vibration, temperature, current, pressure — but the agent layer is configured for each industry's maintenance workflows and compliance requirements.
A typical five-asset pilot including sensors, Azure IoT and ML configuration, agent setup, and Dynamics 365 integration costs between £15,000 and £25,000. Ongoing Azure consumption runs at a few hundred pounds per month. Most clients see breakeven within six months from a single prevented unplanned stoppage.
No. Microsoft's AutoML handles algorithm selection. Veriland manages the agent configuration, the Dynamics 365 integration, and the ongoing tuning. Your maintenance engineers interact with the system through Dynamics 365 work orders, Power BI dashboards, and Teams notifications — not ML notebooks.
Agents are configured with confidence thresholds. When an anomaly pattern falls below the threshold — an unfamiliar combination, a high-cost decision, a safety-critical asset — the agent escalates to a human with full context. The human's decision is captured and fed back so the agent handles similar situations autonomously next time. This is how institutional knowledge builds up in the system rather than existing only in individual engineers' heads.
We are biased, obviously. But the question has shifted. Five years ago, the debate was whether predictive maintenance was mature enough for mid-market manufacturers. It is. The sensing and ML technology is proven, the cost has come down, and the Microsoft ecosystem means you are not starting from scratch if you are already on Dynamics 365 and Azure.
The question now is whether you want a system that tells you something is wrong, or one that does something about it. A monitoring dashboard that sends alerts is useful. An agentic system that detects, decides, acts, and learns is a different order of capability. Every client we have done this for has expanded the programme beyond the initial pilot — not because we sold them on it, but because their maintenance team asked for it.
Get the agentic pipeline diagram, ROI timeline, and key stats in one shareable PDF.
Book a discovery call. We will look at your critical assets, your maintenance pain points, and tell you honestly whether agentic predictive maintenance makes sense — and what a pilot would involve.
Or call us directly: 01625 569 777
