DataData in ManufacturingManufacturing

The same shift scores 73% or 85% OEE, and only the definitions changed

Most OEE disputes between corporate, the plant and the monitoring vendor are not measurement errors. They come from three unratified data definitions (planned time, ideal cycle time and good unit), and the CDO is the person who should own them.

Three supervisors measure the same factory press with rods of different lengths while an executive watches
Listen · 11 min

Overall equipment effectiveness looks like the simplest metric in manufacturing. OEE = (Good Count x Ideal Cycle Time) / Planned Production Time, which you can also break out as availability times performance times quality. Even so, a typical multi-plant group can hold three OEE figures for the same line in the same week. The corporate dashboard has one, the plant manager's morning meeting has another, and the new edge-sensor pilot has a third. Each team defends its own figure, and the argument eats up the operations review.

My view is that these gaps are almost never instrument errors. Each OEE number is the result of three data definitions that nobody signed off on: what counts as planned time, which cycle time is "ideal", and what counts as a good unit. The fix is to treat those three inputs as governed master data with named owners. A new dashboard won't solve it, because a dashboard only displays whichever definitions it was fed.

Why OEE definitions drive capex, bonuses and plant rankings

The OEE definition matters to a CDO because the number gets used to allocate money, and differences in definitions show up as gaps in performance that aren't real. A data glossary vendor, Dawiso, puts it this way: executives who rank sites, set bonuses, or target investment on raw OEE comparisons are, without a shared definition, comparing measurements taken with different rulers. In manufacturing, the cost lands in particular places. A plant whose OEE looks inflated doesn't get the maintenance budget it needs. A plant whose OEE looks low gets a second line approved, when the missing capacity was really lost to unlogged micro-stops. In unionised plants where OEE feeds a gainsharing scheme, a change of definition also changes pay, so it is a labour-relations question as well as a data question.

The gaps are large. Industrial software vendors list the usual causes: some plants calculate their OEE based on theoretical production time, others on actual presence time, while others exclude changeover times. One vendor, iFactory, published a customer quote that says it plainly: "Our OEE was reported at 74 percent from the manual shift logs our operators filled in. When iFactory started capturing directly from the PLC, the real number was 58 percent." The vendor sells the monitoring that found the gap, so read that as a sales anecdote and not as a benchmark. Another monitoring vendor, TeepTrak, says in its 2026 benchmark that manual logs overstate true OEE by 8 to 15 points. That also comes from a company with something to sell, so check it against your own pilot.

The AI copilots that plants are now putting on top of MES data add a new risk. In a dbt Summit 2026 session written up by dbt Labs (a vendor that sells the semantic layer it was promoting), two AWS architects showed an agent-style query summing gross order amounts returned $14.86 million, while actual revenue was $11.36 million, an overstatement of 30.8%. That example used revenue, but it maps directly onto OEE. If an agent is asked "which plant has the worst OEE?" and the definition isn't governed, it will pick a column, apply some exclusion and give a confident answer.

How do availability, performance and quality inputs drift apart?

They drift because each factor depends on a choice that the formula leaves open, and every site makes that choice locally. ISO 22400-2 sets out the structure: it defines availability as actual ÷ planned production time, effectiveness as quantity × ideal cycle time ÷ actual time, and quality as good ÷ produced. The standard gives you the ratios. It doesn't decide which events go into the numerator and which go into the denominator, and that's where the drift comes from.

Availability: the planned-time exclusion list

Under the standard, time you never meant to run doesn't count against OEE. The disagreement is about which events qualify. Vorne, which sells production monitoring and runs OEE.com, counts changeovers as a loss: it defines stop time as all time where the manufacturing process was intended to be running but was not due to unplanned stops (e.g., breakdowns) or planned stops (e.g., changeovers). Many plants do the opposite, and that's where availability gets inflated. Even the vendor guides point out that teams which exclude changeovers from OEE by classifying them as "planned downtime" inflate availability and hide a real loss.

The stop threshold also moves time between factors. Any stop shorter than the threshold is considered a small stop (a Performance Loss) and is not recorded. A typical stop threshold is five minutes for manual systems and two minutes for automated systems. So if a plant logs by hand and a sister plant has PLC capture, the same jam can land in availability at one site and in performance at the other.

Performance: whose "ideal" cycle time

This is the input that gets gamed most quietly. According to Vorne's guidance, Ideal Cycle Time is the theoretical minimum time to produce one part (it is NOT a 'budget' or 'standard' time). The preferred method is to use Nameplate Capacity (the design capacity specified by the equipment builder). In practice, a plant with ageing presses often swaps in the routing's standard rate, and that's the rate finance uses to cost the product. Performance jumps immediately and no equipment has changed. The ideal cycle time needs to be stored for each product on each machine, so it is a master-data field. A clean equipment hierarchy tied to the routing decides whether the field can be audited at all.

Quality: first pass or final yield

The quality factor depends on when a unit is counted as good. In the ISO 22400-aligned reading, rework counts as a quality loss, not as good parts. Plants that report from the ERP goods receipt, after rework, count reworked parts as good. Plants that report from the station's first-pass inspection don't. In a regulated sector this also affects recall exposure. If the line produced a part that needed rework, your batch records already say so, and an OEE figure that hides it contradicts your own quality system.

A worked example: one shift, three numbers

This example starts from a published shift calculation from Machdatum, a manufacturing software vendor: 480 min shift, downtime of 45 min breakdown plus 15 min changeover, availability (480, 60) / 480 = 87.5%, ideal output 420 min x 2 pieces/min = 840 pieces, actual output 720 pieces, performance 85.7%, OEE 0.875 x 0.857 x 0.975 = 73.1%. The variations after the first version are my own hypothetical arithmetic, and the physical shift doesn't change in any of them.

  • Corporate version (changeover counted as loss, nameplate rate, first-pass quality): 73.1%.
  • Plant version (changeover moved to planned time): planned time falls to 465 minutes and availability becomes 420/465 = 90.3%. OEE is 0.903 x 0.857 x 0.975 = 75.5%.
  • Line version (adds a standard rate of 1.8 pieces/min, and 12 of the 18 rejected pieces are reworked and counted as good): performance becomes 720/756 = 95.2% and quality becomes 714/720 = 99.2%. OEE is 0.903 x 0.952 x 0.992 = 85.3%.

That last figure passes the "world-class" mark. In his 1984 book *Introduction to TPM*, Nakajima suggested availability of 90 percent, efficiency of 95 percent, and quality of 99 percent, yielding 85 percent (world-class) OEE. The line moved from ordinary to world-class through three definition choices, and none of them was recorded anywhere.

The decision rule: the could-have-decided test

This is the rule I would write on a whiteboard:

If anyone on site could have decided otherwise, it is a loss. Only what nobody on site could change leaves the denominator.

Under this rule, no scheduled shift, no customer demand, contractually fixed meal breaks and legally mandated cleaning can be excluded. Changeovers, warm-ups, waiting for material and preventive maintenance the plant scheduled itself all stay in as losses. Ideal cycle time comes from nameplate or a time study and can only be changed through a logged change request. Good units are counted at first pass. Any site that wants an exception publishes it as a named variant, such as "OEE-local, changeover excluded", and the variant never takes over the standard figure's name.

When a single standard OEE helps and when it misleads

A standard definition is worth the political cost when you are comparing like with like: sister plants running similar equipment, or capex decisions between sites. Hutchinson is the most cited case. According to TeepTrak, the vendor that supplied the sensors (so these are vendor-published results), every site measured OEE the same way, standardized on the ISO 22400-2 methodology, so results were comparable across plants and countries, and the Tier-1 automotive supplier went from 42 to 75 percent across 40 sites in 12 countries. An independent check is a peer-reviewed MDPI case study of a plastic injection moulding plant, which reported that average OEE increased to 89% compared with 64% in January, September 2025. The authors found the earlier losses were constrained primarily by downtime and suboptimal cycle times, which means availability and performance, the same two factors that drift most.

There are three places where one number misleads.

First, comparing across sectors or process types. A trade analysis notes that a regulated pharma line at 60% may be world-class for its context, because much of its downtime is mandated rather than wasted. The same applies to high-mix shops: applying Nakajima's benchmark to a changeover-heavy job-shop and concluding the plant is failing is a category error. For these sites, compare loss trees, not headline OEE.

Second, averaging across a facility. The vendor FlowFuse makes a fair point: averaging OEE across an entire facility produces a meaningless number. Report OEE at the constraint, and give everything else as loss minutes.

Third, the drop after standardising. When a group moves to strict definitions, reported OEE falls even though nothing on the floor has changed. If plant managers are paid on the old figure, they will fight the change. Rebaseline incentives on the same day you switch definitions, or don't switch.

For long-range capacity planning, OEE is the wrong ratio anyway. TEEP extends the denominator to the full calendar day or week, so it shows the capacity locked away in unscheduled shifts, which OEE leaves out by design.

What to do on monday

Each of these takes a week or less:

  • Collect the OEE calculation actually used at every site, whether it lives in MES configuration, a PI tag script or an Excel macro. Put the planned-time exclusions side by side in one table. Most groups find they have at least three variants.
  • Pick one constraint line and recalculate last week's shifts under the corporate, plant and strict definitions, as in the worked example. Present the spread to the COO as a definitions gap, not a performance gap.
  • Query the routing master: for each product-machine pair, does the ideal cycle time come from nameplate, a time study or the costing standard? Count the costing-standard pairs.
  • Set one group stop threshold and one shared reason-code list, and record which sites log by hand and which capture from the PLC, so threshold effects can be traced.
  • Register OEE and its three inputs as a governed metric with a version and an owner, and route any AI copilot pilot through that metric only, never through raw MES tables.

Your next OEE dispute will most likely come down to whether a changeover sits inside or outside planned time. Write that decision down, version it, and tie it to the incentive plan before anyone compares plants again.

The full course on this sector: Data in Manufacturing.

Frequently asked questions

Should changeovers count as planned downtime in OEE?

Under a strict reading, no. Changeovers are time the line was meant to be running, so they count as an availability loss. Vorne's OEE guidance treats changeovers as stop time. Plants that move them into planned downtime inflate availability, and in the worked example that change alone raised OEE from 73.1% to 75.5% for the same shift.

What is the difference between ideal cycle time and standard cycle time in OEE?

Ideal cycle time is the theoretical fastest time to make one part, preferably taken from the equipment builder's nameplate capacity. Standard cycle time is a budget or costing rate. Using the standard rate makes performance look better without any change to the machine, which is why ideal cycle time should be governed master data that can only be changed through a logged change request.

Does ISO 22400 make OEE comparable across plants?

ISO 22400-2 standardises the ratios: availability as actual over planned production time, effectiveness as quantity times ideal cycle time over actual time, and quality as good over produced. It does not settle which events each plant excludes from planned time, or which stop threshold splits availability losses from performance losses, so the group still has to write those rules down.

Why does reported OEE drop after automating data collection?

Manual shift logs usually miss short stops that operators clear without recording them. Once PLC or sensor capture starts, those minutes appear as performance losses. Vendor-published examples describe drops of more than ten points. Groups should rebaseline incentives on the day definitions or capture methods change, so plant managers don't resist an accurate number.

Go deeper

The lessons that take this article further, free to read.

  1. 1Measuring what matters with OEE and quality metricsData in manufacturing
  2. 2Data quality metrics that matter on the shop floorData in manufacturing
  3. 3Master data done right: materials, BOMs, and equipment hierarchiesData in manufacturing
  4. 4Integrating MES and ERP for a unified data backboneData in manufacturing
  5. 5The metrics & semantic layerAnalytics, BI & decision intelligence

Sources

Finished reading?

Validate your read to earn XP and feed your radar.