Last year we got a call from a food manufacturer in the Midlands. Their main packaging line had gone down on a Thursday afternoon — a gearbox seizure on a palletiser that nobody saw coming. By the time a replacement gearbox arrived (ten days, air freight from Germany, because the UK distributor was out of stock), they'd lost north of £180,000 in production and had to pay a contract packer to fill the gap. The gearbox itself cost £2,400.
That kind of story is not unusual. If you run a manufacturing business or any operation with critical physical assets, you probably have your own version of it. The maths is always the same: the part is cheap, the downtime is not.
Predictive maintenance — using sensors and machine learning to spot equipment problems before they become equipment failures — is how you stop that happening. It has been an enterprise-grade capability for a while now. But the cost and complexity have come down enough that it makes sense for SMEs too, especially if you're already in the Microsoft ecosystem. At Veriland Consulting, we've implemented predictive maintenance solutions using Microsoft Dynamics 365 — both Business Central and Finance & Operations — with Azure IoT across manufacturing plants in the North West and Midlands. This article is an honest walkthrough of what's involved, what works, and where people get tripped up.
Most businesses we talk to are doing one of the first two. Reactive maintenance — run it until something goes bang — is the default for a lot of SMEs, not because anyone thinks it's a good idea, but because nobody's had the bandwidth to set up anything better. It's the cheapest approach right up until it isn't. Emergency call-outs, air-freighted parts, overtime to catch up on lost production. One bad failure and you've spent more than a year's worth of planned maintenance would have cost.
Preventive maintenance is the sensible middle ground that most well-run plants end up at. Service the machines on a schedule. Replace wear parts at fixed intervals. It works, broadly. But it's wasteful in ways that are hard to see from the outside. You're pulling perfectly good bearings because the calendar says so. You're scheduling shutdowns for machines that are running fine. And you still get caught out by the failures that don't respect the schedule — which, in our experience, is most of them.
Predictive maintenance is the third option. You fit sensors to the equipment, collect data on how it's actually behaving — vibration, temperature, current draw, that sort of thing — and use machine learning to spot when something starts drifting away from normal. The difference is you're not asking "when is this machine due a service?" You're asking "is this machine telling me something's wrong?"
The practical distinction: Preventive maintenance says "replace this motor's bearings every 2,000 hours." Predictive maintenance says "the vibration signature on this motor changed last Tuesday — the outer race bearing is starting to pit." One costs you money whether there's a problem or not. The other only costs you money when there is.
We've done this on CNC machines, packaging lines, HVAC plant, pumping stations, a couple of food processing setups. The specifics change but the shape of the project doesn't. Here's how it normally goes.
Not every machine. This is where people overcomplicate things. You pick the three to five assets that would hurt the most if they went down unexpectedly — your bottleneck machine, the one with the eight-week lead time on the main bearing assembly, the ageing compressor that your maintenance manager loses sleep over.
The sensors themselves are nothing exotic — off-the-shelf, you're looking at vibration sensors bolted to motor housings and bearing blocks, temperature probes on known hotspots, and current clamps on drive feeds. They're wireless, battery-powered (typically 3-5 year battery life), and you can retrofit them to running equipment without stopping the line. With off-the-shelf hardware, a five-machine pilot typically comes in between £5,000 and £8,000 depending on how many measurement points you need.
However, this is one area where Veriland has a genuine edge. We employ mechatronic engineers in-house who specialise in designing custom sensor circuits tailored to the specific assets and conditions on your shop floor. By developing purpose-built boards rather than buying off-the-shelf modules, we've brought the sensor hardware cost for a typical five-machine pilot down to around £2,500 — roughly half what you'd pay with commercial sensors. The custom circuits also tend to be more reliable in harsh manufacturing environments because they're designed for exactly the measurement profile each asset needs, with no unnecessary components adding points of failure.
The sensors push readings over Wi-Fi or LoRaWAN (useful if your plant has spotty Wi-Fi coverage in the production areas — most do) into Azure IoT Hub. IoT Hub is the bit that handles the plumbing: device registration, security, making sure readings don't go missing if the connection drops briefly. It's not glamorous work, but getting this layer right matters a lot. We've seen projects from other consultancies where they skipped this step and went straight to the analytics, then spent months debugging data gaps.
Once connected, data flows in continuously. A vibration reading every couple of seconds, temperatures every minute. You need four to eight weeks of this before the ML side is useful — the model needs enough history to learn what "normal" actually looks like for each specific asset under your specific operating conditions.
Once you have a decent baseline dataset, you feed it into anomaly detection models in Azure Machine Learning. The model learns the healthy vibration profile of a specific motor, what a normal temperature curve looks like on a particular bearing housing, how much current the drive should be drawing at different loads. Each asset gets its own model because — and this is important — no two machines behave identically, even if they're the same make and model. Age, installation, alignment, load profile: all of it matters.
You don't need data scientists for this, which is the bit that surprises most people. Microsoft's AutoML can pick and tune algorithms automatically. We handle the configuration, the feature engineering (deciding which sensor readings matter and how to combine them), and the tuning to reduce false positives — which, honestly, 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 on a hot day.
With the models running, the system continuously compares incoming sensor data against the learned baselines. When a bearing starts to develop a fault, the vibration signature shifts — subtly at first, often weeks before anyone on the floor would notice. The model picks this up and scores it: how far has it drifted, how fast is it deteriorating, how does this compare to known failure patterns.
This is fundamentally different from a simple threshold alarm. A threshold alarm tells you the vibration has exceeded 7.1 mm/s — at which point the bearing is probably already damaged. The ML model tells you the vibration trend has shifted and the bearing is degrading. One gives you hours. The other gives you weeks.
This is the bit that separates a useful system from a fancy dashboard that nobody checks. When the model flags a problem, it doesn't just fire off an email (we all know what happens to emailed alerts). It raises a work order directly in your Dynamics 365 environment, assigns it to the right engineer, checks spare parts stock, and suggests a maintenance window based on the production schedule. If the predicted time-to-failure is short, it escalates.
How this works depends on which Dynamics 365 platform you're on. Business Central does not have native asset management or work order capabilities — so we use MaxWAM, our own asset management product, to provide the full work order lifecycle, asset register, and planned maintenance that BC lacks out of the box. For Finance & Operations customers, F&O already has an Asset Management module — but MaxWAM extends it with Remaining Useful Life (RUL) tracking. The ML models calculate how much operational life each asset has left, and that RUL figure is surfaced directly inside the application so maintenance planners can schedule work orders based on predicted condition rather than fixed calendar intervals. It turns F&O's asset management from time-based to condition-based.
Getting this integration right from the start is something we insist on, even when clients want to "just do the monitoring first." A predictive maintenance system that doesn't connect to your maintenance workflow is a monitoring system. It tells you something is wrong but doesn't help you do anything about it efficiently.
A real example, warts and all: 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 about nine days, but it wasn't dramatic — their maintenance team initially dismissed the alert as a false positive because the line was running fine. We had to talk them into pulling the bearing at the weekend. When they did, the outer race was visibly pitted. Their maintenance manager reckoned it had maybe another week before it would have locked up mid-run. A new bearing: £85 and forty minutes of labour. If it had seized on a weekday, they were looking at two days of lost output on that line while they sourced a replacement assembly. The point isn't that the system worked perfectly from day one — it's that even a disputed alert prevented a far worse outcome.
Five years ago, building a predictive maintenance system meant stitching together hardware from one vendor, a data platform from another, ML tooling from a third, and then writing custom integration code to connect it all to your ERP. For an SME, the integration cost alone killed the business case.
What's changed is that the whole pipeline now lives inside one ecosystem. If you're already running Dynamics 365 — whether Business Central or Finance & Operations — and Azure (or even just Microsoft 365), you're not starting from zero.
Azure IoT Hub is the front door for sensor data. It handles device registration, security certificates, and makes sure readings arrive reliably even when your factory Wi-Fi is having one of its moments. Azure Stream Analytics sits behind it, processing data as it arrives — applying immediate rules (like "if temperature exceeds X, alert now") while also routing everything into storage for the ML models to chew on later.
Azure Machine Learning is where the predictive models live. The AutoML capability means you don't need to hire a data scientist to pick the right algorithm — the platform tests multiple approaches against your data and selects the one that performs best. We still do manual tuning on top of that (especially for false positive reduction), but the baseline is surprisingly good out of the box.
Power BI handles the dashboards. Your maintenance manager gets asset health scores, trend charts, and the ability to drill into individual sensor readings. Useful for the humans who want to understand why the model flagged something, not just that it did. And Dynamics 365 is where it all lands operationally — work orders, spare parts inventory, purchase requisitions, engineer assignments. For Business Central customers, our MaxWAMproduct provides the asset management and work order layer that BC doesn't include natively. For Finance & Operations customers, MaxWAM extends the built-in Asset Management module with Remaining Useful Life (RUL) tracking so maintenance planners can schedule based on predicted condition. Either way, when the AI says "this bearing needs attention," a work order lands in front of the right engineer.
On cost: Because these are all consumption-billed Azure services, the running cost scales with what you're monitoring. A typical five-to-ten machine deployment runs at a few hundred pounds a month. We had one client agonise over this for weeks before someone pointed out they'd spent more than that on a single emergency call-out the previous month.
We can talk about numbers — and we will — but the thing that strikes us most across these projects is how the maintenance team's working life changes. Reactive maintenance is stressful. You're always waiting for the next crisis. Predictive maintenance turns the job into something planned, skilled, and frankly more interesting. Two clients have told us unprompted that they've found it easier to retain maintenance engineers since the switch. Take that as anecdotal, but it comes up often enough that we think it's real.
On the quantitative side, based on Veriland Consulting implementation data across North West and Midlands manufacturing plants: unplanned downtime typically drops by somewhere between a third and a half in the first year. That is not because machines stop failing. It's because you catch failures early enough to deal with them on a planned shutdown instead of a Tuesday afternoon crisis. The failures still happen — you just control when and how you respond to them.
Maintenance spend gets more predictable, too. Reactive budgets are spiky: one bad month with a couple of major failures and you've blown the quarter. With predictive, you're ordering parts weeks in advance at normal prices instead of paying premium for next-day air freight. Based on our project data, the overall spend tends to come down by 20-30%, but almost more importantly, it becomes plannable.
And equipment lasts longer. This is the one that surprises people. When you catch a bearing that's starting to wear and replace it early, you prevent the cascading damage — the worn bearing that then damages the shaft, which then misaligns the coupling, which then takes out the gearbox. A £85 bearing replacement versus a £12,000 gearbox rebuild. Across our implementations, we see clients routinely extending asset life by 20-30%, which for capital equipment costing six figures, is meaningful.
Don't try to do your whole plant at once. Seriously. We have had clients come to us wanting to sensor 40 machines on day one and we talk them down to four every single time. Start with a small pilot, prove it works on your most critical assets, then expand once you have data and confidence.
Pick your pilot machines by asking your maintenance manager one question: "Which machine keeps you up at night?" It's usually the bottleneck asset, or the one with an eight-week lead time on a critical spare, or the ageing piece of kit that's had three expensive failures in the past two years. Those are your candidates.
Budget four to eight weeks for the baseline period after sensors go in. This is the training phase — the model is learning what normal looks like for your equipment, in your environment, under your operating patterns. You can't shortcut this. We've tried. The models produce too many false positives if the baseline period is too short, and false positives erode trust faster than anything else.
Connect it to Dynamics 365 from the start. Not later. Not as a phase two. If alerts don't flow into your maintenance workflow as actual work orders with parts checks and engineer assignments, they'll get ignored. We learned this on our second project, when a client had a beautiful monitoring dashboard that nobody opened for three weeks because everyone was too busy fighting fires on the shop floor.
If you're already on Business Central or Finance & Operations with Azure, you can realistically go from initial scoping conversation to live monitoring on pilot assets in about eight weeks. If you're not on Dynamics 365 yet, our Business Central Starter Pack or F&O Essentials Pack gets you onto the platform at a fixed price, and the predictive maintenance layer can follow once your core ERP is bedded in.
Off-the-shelf sensor hardware for a five-machine pilot typically costs between £5,000 and £8,000, but Veriland's in-house mechatronic engineers design custom sensor circuits that bring hardware costs down to around £2,500 for the same scope. Ongoing Azure consumption costs — covering IoT Hub, Stream Analytics, and Machine Learning — run at a few hundred pounds a month for a five-to-ten machine deployment. Implementation and configuration by Veriland Consulting is scoped and quoted per project, but most SME pilots complete within eight weeks.
No. Microsoft's AutoML capability selects and tunes algorithms automatically. Veriland handles the configuration, feature engineering, and false-positive tuning so your team does not need specialist data science skills. Your maintenance engineers interact with the system through familiar tools — Power BI dashboards and Dynamics 365 work orders — not ML notebooks.
Sensor data flows from your shop floor into Azure IoT Hub, which routes it through Azure Stream Analytics for real-time processing and into Azure Machine Learning for anomaly detection. When the model flags a potential fault, a work order is created directly in Dynamics 365 — Business Central or Finance & Operations — via the Dataverse connector, complete with asset details, recommended actions, spare parts checks, and suggested maintenance windows based on the production schedule.
You need four to eight weeks of baseline data after sensors are installed. During this period, the ML models learn what "normal" looks like for each specific asset under your operating conditions. After that training phase, the system begins scoring anomalies and raising alerts. Most Veriland clients see their first actionable alert within the first few weeks of live operation.
We're biased, obviously. But here's the thing: every client we've 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. Once people see the alerts working — once they've had that first experience of catching a failure before it causes a stoppage — the question stops being "does this work?" and starts being "which machines do we add next?"
The technology is mature. The cost is proportionate to SME budgets. And the alternative — continuing to absorb the cost of reactive failures and over-servicing healthy equipment — has a price tag too. It's just one that most businesses have stopped noticing because they've been paying it for so long.
Get all the stats, pipeline diagram, and ROI timeline in one shareable PDF.
Book a discovery call. We'll look at your equipment, your pain points, and tell you honestly whether predictive maintenance makes sense for your operation — and what a pilot would involve.
Or call us directly: 01625 569 777
