Only in a narrow form. Full AI prediction needs years of failure history and continuous sensor data that a workshop with five to fifteen machines rarely has. What is realistic is monitoring the one or two machines whose breakdown stops everything, using wireless sensors from about $26 a sensor a month or free alerts from newer machine controls.
The deciding number is what a day of unplanned downtime on your bottleneck machine costs, multiplied by how often it breaks. If a lost day costs under about $1,000 and the machine fails once a year or less, a disciplined service schedule and a logbook will do more for you than sensors. If a day costs several thousand and failures come with warning signs such as noise, heat or vibration, monitoring can pay for itself with one avoided breakdown.
What "predictive" needs that a small workshop usually lacks
Vendors use "predictive maintenance" loosely, so it helps to separate three things:
- Preventive maintenance: servicing by the calendar or by run-hours. Change the spindle oil every 500 hours whether it looks dirty or not.
- Condition-based maintenance: measuring something (vibration, temperature, motor current, oil condition) and acting when it crosses a limit you set.
- Predictive maintenance: a model that learns normal behaviour and forecasts a failure before any fixed limit is crossed.
True prediction needs three ingredients. First, many examples of the same failure, because a model learns from patterns and one seized bearing is an anecdote, not a pattern. Second, continuous data from before those failures. Third, someone who acts on an alert within hours. A factory with 200 identical pumps has all three. A workshop with a CNC mill, two lathes, a press brake and a compressor has a different machine for each job and perhaps one serious failure every couple of years.
That is why the realistic target for a small workshop is condition-based monitoring, with AI used for the parts it genuinely helps: learning what "normal" looks like for each machine so you don't have to set every limit yourself, and reading your maintenance records for patterns you hadn't noticed.
The downtime sum that decides it
Before looking at any product, work out what a breakdown costs on each machine. The formula:
Cost of one breakdown =
(hours down x lost contribution per hour)
+ overtime to catch up
+ repair parts and engineer call-out
+ expedited freight or late-delivery penalties
Yearly risk = cost of one breakdown x breakdowns per yearLost contribution per hour means the margin you lose, not the sales price: what the machine would have earned over its materials. If other machines can pick up the work, the lost hours are smaller than they look.
A quick illustrative sum for a small fabrication shop's press brake: a hydraulic pump failure took it out for three days while parts arrived. Lost contribution was about $150 an hour over 16 hours a day, so $7,200. Overtime to catch up added $1,100, the pump and engineer $3,400, and one late order cost a $500 penalty. One breakdown: $12,200. If that happens once every two years, the yearly risk is $6,100. A monitoring setup for that one machine only has to cost a fraction of that, and warn early even half the time, to be worth considering.
Run the same sum on the compressor, the saw and the lathes. In most workshops one or two machines carry nearly all of the risk. Those are the only ones worth monitoring.
Five rungs, from a logbook to AI alerts
Climb only as far as the downtime sum justifies. Each rung is cheaper than the one above it, and most workshops stop at rung two or three.
Rung 1: a run-hours logbook and a real service schedule
Cost: nothing but discipline. Record run-hours, services, part changes and every odd noise in one place: a shared spreadsheet or a sheet on the machine. Picture a small joinery workshop whose dust extractor motor burned out twice in three years. When the owner finally logged it, both failures came within weeks of a filter being left unchanged past the maker's interval. A reminder on the filter change fixed a problem that no sensor would have predicted better. If years of history sit on paper job cards, turning handwritten forms into spreadsheet data gets them into a form you can analyse.
Rung 2: the alerts your machines already have
Cost: often free. Many modern CNC controls record alarms, spindle load and cycle status. Haas Automation, for example, includes its HaasConnect remote monitoring on machines with its Next Generation Control, sending machine status alerts by email, text or its app. Switching on alerts for spindle overload, coolant level and unexpected stops takes an afternoon and costs nothing extra on machines that support it. Check what your own controls offer before buying anything.
Rung 3: periodic handheld readings
Cost: from an infrared thermometer to a professional vibration meter. A monthly round with a thermometer on motor casings and bearing housings, written into the logbook, catches overheating trends. A handheld vibration meter adds a lot: a professional unit such as the Fluke 805 FC was listed at roughly $2,700 to $3,200 by retailers in September 2026, and cheaper vibration pens exist. The value is in taking the reading at the same marked spot, at the same speed, every month.
Rung 4: wireless sensors on subscription
Cost: from about $26 a sensor a month. VibeCloud, for instance, prices its vibration sensor service from $26 per sensor per month including the hardware, remote analysis by its engineers and online access. Tractian sells cellular-connected sensors with AI failure diagnostics but prices by quote. Subscriptions like these put the analyst on the vendor's side, which suits a workshop without a maintenance engineer. One or two sensors per critical machine is typical: one on the motor, one on the main bearing or gearbox.
Rung 5: AI anomaly detection across the whole shop
Cost: platform pricing, usually quoted. This is where software watches every machine's power, vibration and cycle data and flags anything unusual without you setting limits. It suits workshops with many similar machines or several sites. For a single-site shop with a handful of different machines, the extra machines monitored at this rung usually have low downtime costs, so the platform rarely pays for itself.
A six-machine CNC shop with one bottleneck
Consider an illustrative subcontract machine shop with nine staff: a five-axis mill, two three-axis mills, two lathes and a compressor. The five-axis mill does all the complex work, and nothing else in the shop can take it over. Its spindle was rebuilt 18 months ago after a bearing failure that stopped it for eight working days.
The owner runs the downtime sum for the five-axis mill: $120 an hour lost contribution across 16 hours for 8 days is $15,360. The rebuild cost $14,000, and two customers were late, which cost goodwill but no penalties. Call it $29,000 for one failure. Spindle bearing failures on a machine like this might happen every four to five years, so the yearly risk is roughly $6,000 to $7,000. The other machines, which can cover for each other, come to under $1,000 a year each.
Plan, costed:
| Step | Cost | What it catches |
|---|---|---|
| Switch on control alerts for spindle load and temperature (rung 2) | $0 | Overloads and alarms while unattended |
| Two subscription vibration sensors on the five-axis spindle housing and axis motor (rung 4) | 2 × $26 × 12 = $624 a year | Bearing wear trends weeks ahead |
| Monthly thermometer round on the compressor and lathes (rung 3) | An infrared thermometer and 20 minutes a month | Overheating motors and belts |
| Owner's time reviewing alerts: 15 minutes a week | About 13 hours a year | Nothing on its own; it's what makes the rest work |
Total cash is about $650 a year against a yearly risk of $6,000-$7,000 on one machine. Catching the next bearing failure early enough to plan a rebuild for a quiet week would save most of the eight lost days, and the spindle might be saved from secondary damage. The shop doesn't need AI across the floor; it needs early warning on one spindle and discipline on the rest.
Reading vibration numbers without an analyst
Vibration readings are usually shown as velocity in millimetres per second. The reference most people meet is ISO 20816-3, which replaced ISO 10816-3 in 2021 and sorts readings into four zones: A for newly commissioned machines, B for acceptable long-term operation, C for machines that should only run for limited periods before remedial action, and D for vibration that can cause damage.
There's a catch for small workshops: that part of the standard covers industrial machines above 15 kW, and many workshop motors are smaller. So treat published zone limits as a rough guide and rely mainly on your own trend. A machine that has read 1.8 mm/s for a year and now reads 3.5 mm/s has doubled, and that change matters more than which zone the number falls in.
An illustrative monthly log for a lathe's headstock bearing, taken at a marked spot at 1,500 rpm:
| Month | Velocity (mm/s) | Bearing temp (°C) | Note |
|---|---|---|---|
| Jan | 1.7 | 41 | |
| Feb | 1.8 | 42 | |
| Mar | 1.8 | 41 | |
| Apr | 2.6 | 45 | Slight whine at high speed |
| May | 3.4 | 49 | Whine louder |
April is the moment to act: a 44% jump in one month plus a temperature rise and a new noise. That's a bearing inspection booked for a quiet day, not an emergency in June. A sensor subscription does exactly this comparison automatically every few minutes; a handheld meter does it monthly. Both beat finding out when the spindle seizes.
Letting ChatGPT or Claude read your maintenance log
The most useful AI for a small workshop may be a general assistant reading the records you already keep. Export the logbook or breakdown sheet (remove anything a customer would regard as confidential) and ask something specific. Try something like:
Below is our workshop maintenance log for the last 3 years: date,
machine, what happened, downtime hours, parts, cost.
1. Group the unplanned stops by machine and by cause.
2. For each machine, total the downtime hours and costs.
3. Point out any repeated cause, and anything that happened shortly
after a missed or late service.
4. List the three machines where better monitoring or maintenance
would have saved the most, with your reasoning.
Only use what is in the log. If an entry is too vague to classify,
list it separately rather than guessing.The kind of answer to expect (illustrative):
Repeated causes
- Compressor: 4 stops (31 hours total), 3 logged as "tripped, reset".
Two of these followed hot days in July. Possible overheating;
check ventilation and the intake filter.
- Lathe 2: 2 belt failures, 7 and 9 months apart, both within
two weeks of a heavy job noted as "long bar work".
Entries too vague to classify: 6 (e.g. "machine down, fixed").Two things to fix before acting on this. The compressor pattern is useful, but "hot days in July" is an inference from dates; confirm it with whoever reset the trips. And six vague entries is a finding in itself: they hide which machine really costs you most. The fix is a logbook column for cause, chosen from a short list (bearing, belt, electrical, hydraulic, operator, unknown). That one change makes the next analysis far more reliable, and it is the same record-keeping groundwork covered in getting business data ready for AI.
Where small-shop monitoring goes wrong
- Alerts that go to nobody. Picture a shop that set sensor alerts to a supervisor's email; he left, and the address was deleted. For four months nothing arrived, and the owner assumed all was well. Send alerts to at least two people and to a shared mailbox, and check once a quarter that a test alert arrives.
- Too many false alarms. Limits set too tight fire every time a heavy job runs. After two weeks of noise, people ignore them. Let the system learn normal running for a fortnight before you trust its alerts, and ask the vendor how it separates a heavy cut from a failing bearing.
- A sensor in the wrong spot. A sensor stuck on a sheet-metal guard measures the guard rattling, not the bearing. It needs bare metal on or close to the bearing housing.
- No signal inside a steel building. Cellular and Wi-Fi sensors can struggle behind steel cladding. Ask for a coverage test at the exact machine before signing.
- Nobody acting on the warning. Monitoring gives you a date to plan work. If spare bearings take three weeks to arrive and nobody orders them when the alert fires, the warning is wasted. Decide in advance who orders parts and who books the downtime.
A compressor is a good example of the last point. An illustrative fabrication shop fitted a temperature sensor to its screw compressor and got a clean warning of rising oil temperature. The fix was a free, ten-minute cooler clean, but the job sat on a list for three weeks and the compressor tripped on the hottest day of the year anyway. The sensor worked; the process didn't.
A logbook entry that makes later analysis possible
Whatever rung you stop at, the records you start keeping today decide whether any AI can help you in two years. Most workshop logs fail because entries are written in a hurry with no fixed fields. A filled-in example of an entry worth keeping:
| Field | Example entry |
|---|---|
| Date and time noticed | 14 May, 10:20 |
| Machine (fixed name) | Lathe 2 |
| Symptom | Squeal at start-up, belt smell |
| Cause (from list) | Belt |
| Planned or unplanned | Unplanned |
| Downtime hours | 3.5 |
| Parts and cost | V-belt set, $64 |
| Run-hours at the time | 11,420 |
| What was running | Long bar work, 40 mm steel |
Fixed machine names matter more than they look: if the same lathe appears as "Lathe 2", "L2" and "the old one by the door", no analysis will group it correctly. The same goes for causes. A drop-down list in a shared spreadsheet or a simple form on a tablet by the door keeps entries consistent, and it takes less time to fill in than a free-text note.
Questions to ask a sensor vendor
- What exactly is included in the monthly price: hardware, gateway, cellular data, analysis by a person or only software?
- What is the minimum number of sensors and the minimum contract length?
- Who owns the data, can I export it, and what happens to it if I cancel?
- How long does the system take to learn normal running before alerts are trustworthy?
- How are alerts delivered, to how many people, and can I set quiet hours?
- How long do batteries last, and who replaces them?
- Can you show me an alert from a comparable machine and what the customer did about it?
Treat the answers as part of the price. A cheap sensor that needs you to interpret raw charts costs more in your time than a dearer one that tells you "bearing wear, plan inspection within four weeks". And weigh the payback claims on vendor sites against your own downtime sum, not theirs: the tutorial on calculating ROI before you build shows how to discount optimistic assumptions. If a camera for quality problems is also on your list, what AI visual inspection costs uses the same approach for inspection hardware.
Follow-up questions from workshop owners
Can I fit vibration sensors to older machines with no electronics?
Yes. Wireless vibration and temperature sensors mount magnetically or with adhesive or a stud on the bearing housing or motor casing, so the machine's age and control don't matter. What matters is a flat, bare metal spot close to the bearing and a signal path from the sensor to the internet, either Wi-Fi, a gateway or cellular. Steel buildings can block signals, so ask the vendor to test coverage before you commit.
Is predictive maintenance the same as a maintenance schedule?
No. A schedule services a machine by calendar or run-hours whether it needs it or not. Condition-based maintenance acts when a measured reading crosses a limit. Predictive maintenance goes further and forecasts when a failure is likely from trends in the data. Most small workshops get the biggest gain from a good schedule plus a few condition checks, long before true prediction.
Will my machine insurer or supplier give me a discount for monitoring?
Some insurers and machine dealers take monitoring into account, but it varies and I can't promise it. Ask your broker directly whether condition monitoring on named machines affects the premium or excess for breakdown cover, and ask your machine dealer whether monitoring data helps with warranty claims. Get any answer in writing before counting it as a saving.
Further reads
- Where Should a Small Manufacturer Start With AI? — Where maintenance fits among a manufacturer's first AI projects.
- AI Quality Control for Small Manufacturers: Where to Start — Catching the quality drift a worn machine causes.
- Do You Need a Lot of Data to Use AI in Your Business? — How much data AI really needs, in plain terms.
- AI Maintenance Costs: What You Pay After an Automation Goes Live — The running costs once any AI system is live.
- What If Your AI Vendor Shuts Down? Checks Before You Commit — Checks before relying on a sensor subscription.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: VibeCloud predictive maintenance pricing page; Tractian plans page; Haas Automation HaasConnect page; retailer listings for the Fluke 805 FC vibration meter; published summaries of ISO 20816-3 vibration severity zones. Checked September 2026.