Industrial AI for SMEs: energy, vision, predictive maintenance. In what order?

Almyros, 06:50. The maintenance manager hears the compressor before he sees it, and it isn’t the sound he knows. In the office, this month’s electricity bill lies open with the total circled twice. At the loading bay, a pallet comes back from a customer with a note: “Same again.”

Three AI ideas for the same plant: a meter, a camera, a sensor. Which do you start with?

Where Greek manufacturing stands

Among firms with 10+ staff, 7.77% of Greek manufacturers use some form of AI, against 17.27% across the EU (Eurostat 2025, official statistics). Inside the production process itself, AI is used by 1.88% of Greek firms across all sectors, against 4.14% in the EU. Large Greek firms went from 24.27% to 37.58% in a single year; small ones, with 10–49 staff, slipped from 8.20% to 7.33% (Eurostat, official statistics).

In a survey of Greek SMEs, 69% said they need predictive maintenance and only 3–4% use it (AI Readiness Index 2025, survey co-authored by a vendor). The need is there. What’s missing is an order that doesn’t put the line at risk.

Three tests for the order

1. What data you already have. Every plant has an electricity meter. Not every plant has a breakdown log with causes.
2. How much it touches the line. Energy reads meters. Vision sits on the line. Predictive maintenance puts sensors on the machines.
3. How fast it shows up in euros. The electricity bill arrives every month. A breakdown you avoided shows up nowhere, precisely because it never happened.

Four steps, in order · pilot lengths: typical range

Step	What you already have	What gets added	Pilot	What we measure	Touches the line?
1. Energy	Bills, the main meter, PLCs	Sub-meters only where missing	6–10 weeks	kWh per unit, peak kW, €/MWh	No
2. Vision on one line	Often IP cameras	A camera where there isn’t one, and proper lighting	6–8 weeks	Defects reaching customers (ppm), scrap %, credit notes	Rejects only within limits you set
3. Predictive maintenance	Work orders, a breakdown log	Vibration, current and temperature sensors on 3–5 machines	8–12 weeks	Unplanned downtime hours, MTBF, MTTR	Reads, never sends commands
4. Planning and operator copilot	The planning spreadsheet, manuals, past work orders	A link to the ERP	6–10 weeks	Planner hours, changeover time, MTTR	No

1. Energy: start with the bill you already pay

The bill arrives once a month, as one number. Inside it are things nobody ordered: a compressor running all weekend to feed leaks, three dryers that start together at 07:00 and set the month’s peak.

We start from your bills, your main meter and your PLCs, and add a sub-meter only where a big load isn’t measured. We tie consumption to output per shift, so “how much power” becomes “how much power per unit”. The system forecasts tomorrow’s load, suggests which loads can move out of the peak, and flags anything that doesn’t add up. If your plant has to carry out an energy audit or run an energy management system under Directive (EU) 2023/1791, you’ll be asked for this same data anyway.

The limit: if output isn’t logged per shift, kWh per unit is guesswork. On our plant energy page you can move loads out of the peak in a typical plant.

2. Vision on one line: so the customer stops finding the defect first

A camera looks at every part, not one sample an hour. There’s often an IP camera over the line already. What’s nearly always missing is proper lighting, which matters as much as the camera.

Before any pilot, we ask for 20 photos of the defects on your line and tell you whether the defect shows up on camera. In the pilot, anything above a confidence threshold is rejected; anything in the grey zone is judged by a person.

The limit: you won’t find an accuracy figure here. Wherever you see one close to 100%, ask: on how many images, for how many defect types, in a lab or on a line? Your number comes from your own photos, mistakes included, such as a reflection read as a crack. You’ll find examples on our vision quality control page.

3. Predictive maintenance: only on the machines that stop everything

We don’t put sensors everywhere. We pick the 3–5 machines that take the line down with them when they stop: a compressor, a motor, a pump, a press. Vibration, current and temperature show when their behaviour changes, so the stop is planned for Thursday at 06:00 instead of landing on you on Friday night.

If you’ve logged very few breakdowns, which is common in well-maintained plants, we don’t start with failure prediction but with anomaly detection: the system learns each machine’s normal behaviour and alerts you when it changes.

For scale: a large plant loses 27 hours a month to unplanned downtime on average (Siemens, 2024, vendor data). The steelmaker BlueScope avoided around 2,000 hours of downtime and 53 interruptions over three years with predictive maintenance (Siemens, vendor data). Those are big plants. Yours will get its own numbers from its own predictive maintenance pilot.

When to break the order

• The customer finds the defect. If credit notes and returns cost you more than you could save on electricity, start with vision.
• One machine holds the line hostage. If a single machine causes most of your stops, start there, and only there.
• The problem is paper, not machinery. 3.1 certificates, delivery notes, HACCP files: a plant’s fastest first step often isn’t industrial at all. It’s document reading.
• Everything hangs on one person. “When Nikos is on leave, line 2 stops.” Then the operator copilot comes first, built on your manuals and past work orders.

Five questions before anyone connects to your line

They apply to every supplier. Us included.

1. Read-only, or able to send commands? Will it read from PLCs and meters, over Modbus, OPC UA or MQTT, or will it be able to send commands too?
2. What runs inside the plant and what runs in the cloud? And what happens when the internet goes down?
3. How is the production network kept apart from the office network? The reference standard for industrial control system security is IEC 62443. If your plant falls under Law 5160/2024 (NIS2), the question carries legal weight too.
4. Who gets the alert at 03:00? And what do they do once they have it?
5. Who owns the data and the models when the contract ends?

How we’d start

The Plant X-Ray runs for 10 working days on site: stopwatch readings at the workstations, energy, downtime and quality data, and a camera test on one line. It ends with an opportunity map showing value, feasibility and payback, and a pilot proposal with targets signed before anything is built. One of our two offices is inside the Almyros industrial area, and our engineers complete your plant’s safety induction before they set foot on the line.

The Plant X-Ray starts from €4,900, excluding VAT, and is credited in full against a pilot signed within 60 days.

Not sure AI fits your business at all? Read where a business starts with AI and the formula we use to measure lost hours.

Frequently asked questions

Do we need to buy new equipment?

Usually very little. Energy starts from the meters and PLCs you already have, adding sub-meters only where they’re missing. Vision uses existing IP cameras where it can, but nearly always needs proper lighting. Predictive maintenance needs sensors on the 3–5 critical machines. The Plant X-Ray lists all of it, with costs, before anything is bought.

Can the AI stop the line on its own?

Only if you decide it should, within limits you sign off. In the pilots we propose, the system alerts and suggests. The reject gate removes only what sits above a confidence threshold, a person judges anything uncertain, and every maintenance alert goes to a named person on shift.

What if we’ve logged very few breakdowns?

Then we don’t start with failure prediction, because there aren’t enough examples for a model to learn from. We start with anomaly detection: the system learns each machine’s normal behaviour and alerts when it changes. Alongside that, we organise the breakdown log so it becomes useful.

Is there funding?

Sometimes. The Plant X-Ray includes an annex listing the programmes that could cover your project, subject to each call’s current text and dated when last checked. We don’t sell grants, and we don’t design projects that only stand up if one is approved.

Sources

1. Eurostat, Artificial intelligence by NACE Rev.2 activity (isoc_eb_ain2), 2025 · official statistics · firms with 10+ staff
2. Eurostat, Artificial intelligence by size class and purpose (isoc_eb_ai), 2024–2025 · official statistics · small: 10–49 staff, large: 250 and over
3. AI Readiness Index 2025, ENTERSOFTONE and AUEB/ELTRUN, via powergame.gr (in Greek) · survey co-authored by a vendor
4. Siemens, The True Cost of Downtime 2024 · vendor data
5. Siemens, BlueScope and predictive maintenance (in German) · vendor data
6. European Commission, Energy Efficiency Directive (EU) 2023/1791 · legislation