Search for OEE improvement and the first things you find are the formula and a target number, usually the “world class 85%” figure. Yet most plants that actually start measuring get stuck in the same place. The number comes out higher than expected, and then it refuses to move no matter what countermeasures are launched. The cause is rarely a shortage of improvement ideas. It is that the way the plant measures quietly drops a large part of what happens on the floor. This article works through the definition of overall equipment effectiveness OEE, the benchmarks available in 2026, the mechanism that makes self-reported figures run higher than measured ones, and how to put equipment uptime visibility and downtime cause analysis into practice, with Japanese-owned plants in Thailand in mind.
Overall Equipment Effectiveness OEE Explained | What the Three Factors Measure
OEE stands for Overall Equipment Effectiveness. It expresses, as a single ratio, the share of the output a machine could have produced that it actually produced. ISO 22400-2 defines it as a KPI for manufacturing operations management, and calculating in line with that definition is what makes comparison across plants and across countries meaningful.
The OEE Formula and Its Three Components
OEE is the product of three ratios.
OEE = Availability x Performance x Quality
Each of the three captures a different family of loss.
- Availability is the share of the time you intended to run that the machine was actually running. Breakdowns, changeovers, waiting for material and waiting for people all cut into it.
- Performance is the share of that running time in which the machine held its proper speed. Speed losses, minor stops and idling cut into it.
- Quality is the share of total output that can be shipped without rework. Defects, rework and start-up scrap cut into it.
Put numbers in. For a machine at 85% availability, 90% performance and 98% quality, OEE is 0.85 x 0.90 x 0.98 = 0.7497, or roughly 75%. None of the three components is below 85%, and taken one at a time they all look acceptable. Multiply them together, though, and a quarter of the potential output has disappeared.
That gap between how the components feel one by one and what they produce when multiplied is the single most important property of the metric. Because the three are multiplied rather than added, improving one component dramatically will not move the total if another is low. Conversely, lifting the lowest of the three by 5 points often produces a much larger effect than intuition suggests. Whenever you decide what to work on first, break the number into its three components and look at the smallest one.
One note on vocabulary. Availability, performance and quality are given slightly different names from one source to the next, and the local-language terms used on the floor multiply the variants further. When a plant uses two names for the same ratio, nobody can tell whether two reports refer to the same figure. Which term you standardize on matters far less than writing down, in plain sentences, exactly what goes into the numerator and the denominator of each.
How Equipment Uptime Differs from OEE
The term most often confused with OEE on the shop floor is uptime, or utilization. In most plants, uptime means the share of the intended running time during which the machine was running. That corresponds to Availability alone, one of the three components.
Watch uptime by itself and the number stays high as long as the machine is turning. If it runs at 80% of its proper speed, uptime does not notice. If 3% of what it makes is scrap, uptime does not notice. The machine was running, after all. What makes OEE fundamentally different is that it also asks what happened inside the running time.
There is a second problem with the word. “Machine uptime” and “plant loading” tend to blur together. A day when a line was stopped because there were no orders says nothing about the machine’s capability, yet depending on how the denominator is defined it will pull uptime down. If you intend to use OEE as an internal KPI, decide how planned downtime is treated before you start. Otherwise the monthly figure swings with the order book and no one can read the effect of improvement work.
OEE Benchmarks 2026 | The Distance Between a 60% Median and 85% World Class
Once your own number exists, the next question is whether it is good or bad. Here is where the published benchmarks stand as of 2026.
| Level | Typical OEE | What it represents |
|---|---|---|
| Industry median | About 60% | Where most plants actually sit |
| Top quartile | About 75% | Plants running continuous improvement |
| World class | About 85% | The ideal derived from the TPM benchmark |
Several surveys, including TeepTrak’s State of OEE 2026, converge on a similar distribution. A median of 60% means the average plant is losing four tenths of the output it could have produced. That sounds exaggerated the first time you see it, but the arithmetic above explains it. If the three components each sit in the low-to-mid 80s, multiplying them lands you somewhere between the 50s and the low 60s.
Where the 85% Target Actually Comes From
The world class figure of 85% is not the measured average of any real group of plants. It comes from Nakajima’s benchmark in the TPM literature. Multiply the three target values of 90% availability, 95% performance and 99.9% quality and you get 0.90 x 0.95 x 0.999 = 0.8541, or about 85%. In other words, 85% is what you get when an ideal value is placed on each of the three components, and it is a number to work backwards from rather than to adopt as your own target.
Fabrico’s 2025 study of 250 European plants reconfirms 85% as a reference level for what is achievable. The same work notes that plants running many products across several lines rarely sustain 85% as an all-plant average, and that the figure sits close to a theoretical ideal. For a single-product line with almost no changeovers, 85% is a reasonable target. For a high-mix, low-volume plant that adopts 85% as a site-wide average, missing the target becomes the normal state of affairs and the metric itself loses credibility.
Realistic Targets Vary by Industry
The industry-level guidance compiled by Symestic puts discrete manufacturing at 85% and above, pharmaceuticals at 75% and above, and aerospace components at 72% and above. The same 85% is a completely different proposition depending on what you make and how tightly it is regulated.
Pharmaceutical and aerospace targets are lower not because those plants are managed loosely. Validation and record-keeping requirements, material behavior and small lot sizes structurally cap what availability and performance can reach. When you set your own target, do not borrow another industry’s 85%. Place an achievable ceiling on each of the three components given your own constraints, and make the result of multiplying those three your target. If that result falls short of 85%, it is still your perfect score.
One more piece of practical advice. For the first year, set the target as a gain rather than a level. A plant at 60% that announces 85% has no way to choose its next action, whereas “plus 5 points in six months” can be worked backwards into which losses have to shrink and by how much.
Why Self-Reported OEE Runs 10-18 Points Above Measured OEE
Before you compare your number with any benchmark, there is one thing to check. Self-reported OEE tends to come out 10-18 points higher than a figure based on actual measurement. This pattern, which TeepTrak and Fabrico both raise repeatedly, is not a story about people massaging numbers. Systems built on manual recording have omissions built into them.
The size of the gap varies. Some studies put it at 8-15 points, depending on how the surveyed plants record data and which industries they represent. What matters is not the exact width but the direction. As long as recording is manual, the reported figure is biased upward.
A plant that believed it was running at 75% and then found 60% once measurement was installed is an entirely ordinary outcome. It looks like a disappointment, but it is where improvement actually begins. Those 15 points were always there. They were simply invisible, and now they are available.
Minor Stops Never Reach the Record
The largest omission is minor stops. A stoppage that lasts anywhere from tens of seconds to a few minutes never appears in the downtime column of a shift report. Operators judge it as not worth writing down, and they are right on their own terms. Interrupting other work to log a three-minute stop is not a sensible use of the shift.
But if a three-minute stop happens 15 times a shift, that is 45 minutes, roughly 9% of an eight-hour shift. The machine really was down for 45 minutes, and the manual record shows not one line. Worse, most minor stops surface in performance rather than availability, so even when something is written down it can only be observed as slower cycle time. In a manual world this loss disappears almost without exception.
Which component should catch minor stops, and where to set the threshold, is worked through in Minor Stoppage Countermeasures 2026 | Catching Stops That Never Show Up in Availability. If the gap between your reported and measured figures is large, start your suspicions here.
Changeovers Get Parked in Planned Downtime
The second omission is how changeovers are classified. Book changeover time as planned downtime and it leaves the denominator, so OEE goes up. The reasoning sounds plausible at first hearing. It is planned, therefore it is planned downtime.
Changeover, however, is the textbook improvement target. Both the ISO 22400-2 framing and the TPM loss taxonomy count setup and adjustment as a loss in its own right. Removing it from the denominator is the same as removing the area with the most improvement headroom from measurement altogether. The number gets prettier and one of your best opportunities vanishes.
On shortening the changeover itself, Changeover Time Reduction 2026 | Why Gains Slip Back When the Unit of Measure Is Too Coarse covers the mechanism by which improvements fail to stick when the measurement unit is too coarse. Read alongside the denominator question here, it makes concrete why changeovers should not be parked in planned downtime.
Nameplate Cycle Time in the Denominator
The third is the reference cycle time used to calculate performance. Use the nameplate cycle time from the equipment maker’s catalog and you are usually using a figure slower than what the machine can do under its best conditions, which flatters performance.
What you should use is the best cycle time that machine has actually recorded for that product. Switch to the measured best and performance drops, but the gap against “the speed it reaches on a good day” becomes visible as loss. That gap is the improvement target. As long as the nameplate figure is in use, nobody in the room actually knows how fast the machine can run.
None of these three is malice or laziness. They are limits inherent in manual recording. That is exactly why the first step in OEE improvement is not a countermeasure but a review of how you measure.
Where Downtime Cause Analysis Starts | Translating the Seven Big Losses
Once measurement is fixed, the next job is classifying where the lost output went. TPM calls the taxonomy the seven big losses, and in the OEE literature it usually appears as the Six Big Losses. The names differ, but they look at nearly the same things. The six-item version generally folds tooling changes into setup and adjustment, and that is roughly the extent of the difference.
| Loss type | Typical content | Component mainly affected |
|---|---|---|
| Breakdown loss | Sudden failures, equipment trouble | Availability |
| Setup and adjustment loss | Product changeover, die change, first-piece check | Availability |
| Tooling loss | Tool and blade replacement and readjustment from wear | Availability |
| Minor stop and idling loss | Material jams, false sensor trips, short stops | Performance |
| Speed loss | Running below design speed, running on reduced settings | Performance |
| Defect and rework loss | In-process defects, sorting, reprocessing | Quality |
| Start-up loss | Scrap at shift start and immediately after changeover | Quality |
Using this table is simple. Identify the lowest of your three components and look only at the losses that act on it. A plant with weak availability that launches a defect reduction program will barely move its OEE. If the table did nothing but keep the order of work straight, it would still be worth pinning to the wall.
The 2026 Finding That “No Orders” Is the Largest Cause of Downtime
The tricky part of downtime cause analysis is how to treat causes the plant cannot control. In the benchmark Guidewheel published in 2026, the largest single loss category across all monitored machines was “no orders”, observed on 37.6% of machines.
You cannot carry that number straight into your own improvement targets. An absent order is a sales and demand question, not a statement about equipment capability or maintenance quality. The finding does carry two practical implications, though. First, if OEE is to be used to evaluate equipment, downtime from a lack of orders must be excluded from the denominator explicitly, and the exclusion has to be recorded. Second, once you know which machines sit idle for long stretches, you can design what that time is used for, such as pulling changeover practice or preventive maintenance forward onto those machines. Downtime cause analysis exists to decide how time is used, not to assign blame.
Keep Reason Codes to What an Operator Can Pick in Five Seconds
Whether downtime analysis works at all depends on whether a reason gets recorded when a machine stops. And whether a reason gets recorded depends almost entirely on how many choices there are.
In plants that set up 50 reason codes, operators select “other”. Data in which “other” dominates tells you nothing. Start with something like 8 to 12 codes, run them for three months, then interview the floor about what is hiding inside “other” and add only what is needed. Keeping the record going matters more than a fine-grained taxonomy.
The second trick is to name the codes in the words the floor already uses. Not “material supply system anomaly” but “out of material”. Not “quality confirmation activity” but “waiting for first-piece approval”. Align the codes with management vocabulary and every selection requires a mental translation, which slows the choice down and pushes people back to “other”.
How to Start Equipment Uptime Visibility | Begin by Choosing the Unit of Measure
Everything above depends on making uptime visible. “Visibility” covers a lot of ground, though, and the difficulty changes completely depending on where you start.
Where to Take the Signal From
There are broadly four ways to capture machine state.
- Read the colors of the stack light with an optical sensor. It requires no modification to the machine and retrofits onto old equipment. What you get is state only, such as running, stopped or alarm.
- Measure current with a current sensor. In many cases a clamp works without touching the wiring, and you also see load level. Separating running from idling requires threshold tuning.
- Take data directly from the PLC. You can reach stop reasons and process data, but you need the equipment maker’s cooperation and protocol support.
- Pick up pulses from the production counter. Count and cycle time come directly, which suits the performance calculation.
Which is right depends on the age of the equipment and the plant’s engineering resources, but for the very first machine the realistic choice is whatever retrofits easily and installs without stopping production. The purpose of machine number one is not perfect data. It is to create one measured figure that internal discussion can stand on.
For the wider investment picture that starts with sensor selection, meaning where the money goes and which cost layers get overlooked, Factory IoT Implementation Guide 2026 | Five Cost Layers and How to Start Uptime Monitoring breaks the cost structure into layers. What you need at the approval stage is collected there.
One Machine Is Enough, but Collect a Full Month First
The classic reason visibility projects stall is scoping them too broadly at the outset. A plan covering every line carries a large number, drags the approval process out, cannot settle on a design because every machine has its own quirks, and burns six months before anything is installed.
The recommendation is to take a single machine at the bottleneck process and collect one month of data. A month is enough to reveal the shape of a day, the pattern across weekdays, the differences between products, and above all the total volume of stoppage that manual records never showed. With one measured figure in hand, the next approval request is a different conversation. “Other companies apparently see results” and “our machine number three had 168 hours of downtime last month, and 62 of those hours never appeared in the shift reports” do not get approved at the same speed.
There is one thing to do without fail at this stage. While you look at the data, walk the floor and ask the operators what was happening during a given window. Data tells you what happened. It does not tell you why. Build the habit of matching data against the words of the people who were there during that first month and the quality of every later analysis changes.
The OEE Improvement Loop | Measure, Analyze, Improve, Control
OEE improvement is not a one-off project but a loop of four stages, and it should be designed that way. Measure, Analyze, Improve and Control is the usual framing, and the 2026 trend is for the measure and analyze stages to be made real-time with IoT and AI.
The measure stage is where you define the numerator and denominator of each of the three components and make them collectable automatically. Document the definitions here or you will not be able to compare sites later. The definition sheet should state what counts as planned downtime, how the reference cycle time is set, and at which process good product is judged. Think of it as replacing the recording path itself, from handwritten shift reports piling up in a binder to sensors on machines feeding a dashboard on their own.
The analyze stage is where you identify the lowest of the three components and break out the losses that feed it. The important discipline is to look at frequency and duration together rather than sorting by duration alone. One eight-hour breakdown a month is 480 minutes. A one-minute minor stop 20 times a day across 25 working days is 500 minutes. As monthly lost time the two are almost identical, but the countermeasures have nothing in common. The first is a maintenance planning problem, the second is a process design or fixturing problem.
The improve stage should target one or two of the losses that analysis identified. Launch five countermeasures at once and you will never know which one worked, so the next cycle learns nothing. Keep the hypotheses under test in a single cycle to one or two.
The control stage is the most neglected and the most common cause of failure. Without a mechanism to confirm that an improved state still holds three months later, the number quietly returns to where it started. What works as a control mechanism is usually something light, such as spending three minutes at the daily morning meeting on yesterday’s OEE and its largest loss. A monthly report notices the backslide far too late.
As for cycle length, one month is realistic to begin with, moving to roughly two weeks once the routine is established. A weekly cycle gives you too little data to analyze and leaves you chasing seasonality and product-mix noise.
What OEE Improvement Actually Delivers | Cases and Estimates
Any investment discussion will ask for the expected return. It is worth separating published cases from the method you use to recalculate for your own plant.
NGK Insulators has set OEE as a common KPI across 19 overseas plants in 11 countries and is visualizing it on a BI platform. Within that program, one plant is reported to have improved OEE by 20%. Lining up plants in several countries under one metric requires standardizing the definition first, and the substance of this case is that standardization.
In a case presented by Canon IT Solutions, automatic generation of OEE reports cut the time spent producing shift reports by 90%, and the company standardized its OEE definition across the business at the same time. Notice that definition standardization appears here too. Automating a report leaves you no choice but to settle on a single formula. Put the other way around, automation works as a device that forces definitions into alignment.
For the scale of improvement, TeepTrak reports a typical OEE gain of 4-12 points 90 days after go-live, based on more than 450 deployments from 2018 through the second quarter of 2026. The same source models a mid-sized plant with five lines running at 60% OEE and estimates an annual net benefit of USD 1.2 million-2.4 million with a payback period of 1.2-2.4 months. Divided across the five lines in that model, it comes to roughly USD 240,000-480,000 per line per year, or USD 120,000-240,000 per line in half a year. The underlying estimate is published by the vendor itself, and it should be read with an allowance for the possibility that favorable projects are over-represented. If you want it as a basis for your own decision, do not adopt it as it stands. Rebuild it with the method below.
The basic form of an in-house estimate is simple. A 5-point gain in OEE means that the same machines in the same hours yield that much more output. Suppose one line runs 20 hours a day for 25 days a month. Monthly loading time is 500 hours. At 60% OEE, 300 hours of that is genuinely producing value. At 65% it becomes 325 hours. The difference is 25 hours a month, and across five lines 125 hours a month, which is 1,500 hours a year.
Multiply those 1,500 hours by your own value added per hour and you have a monetary figure. Put value added at 3,000 THB per hour and the year comes to 4,500,000 THB, though this rate varies enormously between plants, so use your own actuals. The point is not the size of the number but the fact that the capacity it represents arrives without adding a single machine. An approval request for a new machine and an approval request to recover the same capacity from existing equipment are not judged on the same grounds.
The estimate rests on assumptions worth stating. There has to be enough demand to absorb the extra capacity, and people and materials have to keep up. When orders are flat, the benefit shows up as less overtime and fewer weekend shifts rather than more output. That changes the form the benefit takes, not whether it exists, but choose the form you write into the approval request to match your own situation.
Issues Specific to Running OEE Improvement in Thailand
Everything so far is general. Japanese-owned plants in Thailand face a few conditions of their own.
First, the shift to high-mix, low-volume. Plants that once handled large lots flowing down from head office now carry more product variants in smaller lots. More variants means more changeovers, which structurally depresses availability. Announce an 85% target in that environment and the floor receives it as unreachable. If the product mix has changed, the target has to be rebuilt.
Second, workforce mobility. Experienced operators change jobs or move roles more often than in Japan, so “it runs fast when that particular person sets it up” shows up in OEE as a risk. Overlay the monthly performance trend on staffing changes and you will sometimes find the graph dropping exactly when one operator transferred. Part of the value of measuring OEE is making that dependence on individuals visible as a number. Use it not to evaluate a person, but as the raw material for moving that person’s judgment into the system.
Third, reporting to the parent company in Japan. When head office lines up OEE from several sites, a difference in definitions reads as a difference in performance. Put a site that books changeovers as planned downtime next to one that does not and the first looks superior. Before any cross-site comparison begins, do the work of merging the definition sheets into one. It is unglamorous work, and it is vastly cheaper than realigning after comparisons have started.
Fourth, the relationship with carbon reporting. Thai plants increasingly receive requests from customers for GHG and CO2 accounting and reporting. What tends to happen is that data collection for carbon and IoT investment for OEE improvement are set up as two separate projects. In reality, energy consumption, yield and machine state come from the same equipment over the same path. Put them on one data platform and you invest once and answer both demands. In terms of getting an approval through, a design that satisfies two objectives with one investment is also the stronger case.
Four Stumbles Common to Plants Where OEE Improvement Stalls
Finally, here are the failure modes that tend to appear after measurement starts.
First, using OEE to evaluate individuals or shifts. The moment it is used for evaluation, the floor starts behaving so the number looks good. Stop reasons get reassigned to milder categories, incidents go unrecorded, denominators get adjusted. The measurement system you just built degrades to the quality of manual records. Make it explicit at rollout that OEE reflects the state of equipment and processes, and is not a scorecard for people.
Second, starting comparisons while denominator definitions still differ between sites and processes. A meeting held over numbers that were not defined the same way never gets past arguing about whether the numbers are right, and never reaches the improvement discussion.
Third, treating the dashboard as the finish line. When the screen is complete, the project looks complete. The screen is only the deliverable of the measure stage, though, and analyze, improve and control have yet to begin. It is not unusual for the project owner to return to normal duties the day the dashboard goes live and for nobody to be looking at it six months later. Decide who looks at the screen, when, and for how many minutes, at the same time as you build it.
Fourth, not placing an improvement owner on the floor. Data collects in the office, but losses happen on the floor. In a structure where analysis results are simply handed over, the context of why a given stop occurred is never filled in. Whether the person doing the analysis can protect time to spend on the floor several times a week decides whether this loop turns at all.
Frequently Asked Questions (FAQ)
What is overall equipment effectiveness OEE?
Overall equipment effectiveness OEE expresses the share of the output a machine could have produced that it actually produced. It is calculated by multiplying Availability, Performance and Quality. ISO 22400-2 defines it as a KPI for manufacturing operations management, and calculating in line with that definition makes it usable for comparison across plants and across countries. Because the three are multiplied, components sitting around 85% each can still produce an OEE in the 60s.
What is a good OEE target or world standard?
Published data as of 2026 shows an industry median of about 60%, a top quartile of about 75% and world class at about 85%. The 85% figure, however, is the theoretical product of the TPM benchmark values of 90% availability, 95% performance and 99.9% quality, and it is rarely sustained as an all-plant average where many products are run. Industry-level guidance puts discrete manufacturing at 85% and above, pharmaceuticals at 75% and above, and aerospace components at 72% and above. Rather than borrowing another company’s level, place an achievable ceiling on each of the three components under your own constraints and make their product your target.
How does equipment uptime visibility differ from measuring OEE?
The uptime most plants track is the share of intended running time during which machines ran, which corresponds to Availability alone. Uptime stays high as long as machines are turning, so speed losses and defect losses stay invisible. OEE differs decisively in that it evaluates what happened inside the running time, meaning speed and quality as well. Starting from equipment uptime visibility is the right sequence, but stopping there means missing about half of the available improvement.
What does OEE improvement cost and how long does it take?
It depends heavily on scope, but at the scale of instrumenting one machine at the bottleneck and collecting a month of data, there are options that install without stopping production. On the size of the gain, TeepTrak cites 4-12 points at 90 days after go-live across more than 450 deployments, though as vendor data it is better recalculated using your own loading hours and value added per hour. On timing, allow two to three months from the start of data collection through analysis and countermeasures to a confirmed effect.
Where should downtime cause analysis begin?
Identify the lowest of the three components and look only at the losses that act on it. Weak availability points to breakdowns, setup and tooling changes. Weak performance points to minor stops and speed losses. Weak quality points to defects and start-up losses. Do not build a detailed reason code list at the start. Hold it to roughly 8 to 12 codes so an operator can choose in five seconds. After three months of operation, interview the floor about what ended up in “other” and add only what is needed, and the recording habit survives.
Summary
The first thing to work on in OEE improvement is not a countermeasure but the measurement itself. Self-reported OEE tends to run 10-18 points above measured OEE, and that gap comes from three structural sources, namely minor stops that never reach the record, changeovers parked in planned downtime, and nameplate cycle time used in the denominator. Inspect those three, replace them with measurement, and the targets for improvement finally become visible.
From there, run the loop. Identify the lowest of the three components, narrow the countermeasures to the losses that feed it, and stay with them until the gain is embedded. The benchmarks give a median of 60%, a top quartile of 75% and world class at 85%, but 85% is a TPM theoretical value and not a number a high-mix plant should adopt as its site-wide average. Place an achievable ceiling on each component under your own constraints and make the product your target.
In Thailand, four further issues stack on top, namely more changeovers from a broader product mix, workforce mobility, definition gaps in reporting to head office, and coexistence with carbon reporting. On the last point in particular, putting energy consumption and machine uptime on one data platform makes it possible to answer two demands with one investment.
Half of OEE improvement is decided the moment you decide where to measure. Before selecting a sensor, start by writing the definitions of your numerators and denominators onto a single sheet of paper.
Where to measure on your own equipment, which signals are available, and how it connects to the systems you already run are questions whose answers change with the age of the machines and the layout of the control panel. Through uptime monitoring and traceability work at Japanese-owned plants in Thailand, TOMAS TECH is often asked to think these conditions through with the plant, case by case. Even if the direction is still open inside your company, you are welcome to describe your current equipment and concerns through our contact page. We can start from the practical end, such as how to carve out the measurement scope and where to put the first machine.
