OEE for CNC machines: the complete guide

OEE answers one question: of the production time you planned, what fraction turned into good parts at the intended pace? It compresses three different kinds of loss — machines not running, machines running slow, and machines making scrap — into a single percentage.

That compression is both its power and its danger. As a single number it lets an owner compare this month to last month at a glance. But the number itself is not actionable — a 55% OEE tells you nothing about what to fix. The three components underneath do. Treat OEE as the headline and the components as the story.

Note what availability deliberately excludes: time you never planned to produce. A machine idle on an unstaffed night shift is a utilization question (see OEE vs utilization), not an availability loss. Mixing the two is the fastest way to make OEE meaningless.

One shift on one machining center. Shift length 480 minutes, with a 30-minute planned lunch your plant excludes from production time — so planned production time = 450 min.

• Stops during the shift: one 35-minute changeover, one 20-minute tool-change overrun, assorted small stops totalling 15 minutes → run time = 450 − 70 = 380 min
• Availability = 380 ÷ 450 = 84.4%
• Ideal cycle time for the part is 1.4 min; the machine produced 240 parts → ideal time for that output = 336 min. Performance = 336 ÷ 380 = 88.4%
• 232 of 240 parts were good → Quality = 96.7%
• OEE = 84.4% × 88.4% × 96.7% = 72.1%

Try your own shift’s numbers in the OEE calculator — it shows the same arithmetic live.

The number quoted everywhere is that 85% is world class. It comes from mature TPM literature describing lines improved continuously for years — it is a horizon, not an entry benchmark.

Plants measuring honestly for the first time typically land between 40% and 60%. That first number usually feels insultingly low — it isn’t. It means the measurement is finally honest, and the gap between it and where you thought you were is capacity you already own, hiding in stops nobody logged and cycles nobody timed.

The comparison that matters is never against another plant’s number (whose availability rules you don’t know) — it is your machine against itself, month over month.

The pattern worth knowing: on most CNC floors the two performance losses are the least visible and collectively large — a machine that looks busy from the aisle can be losing an hour a shift in micro-stops and slow cycles that no logbook will ever capture.

• Ideal cycle time set wrong. Set it slower than reality and performance flatters everyone; losses vanish into a number above 100% that gets quietly capped. Use the best demonstrated cycle, per part program.
• Availability rules that drift. Breaks counted against one shift and not another, maintenance excluded when convenient. One written policy, applied identically everywhere, or trend lines mean nothing.
• Averaging across machines. A floor-level OEE hides the one machine at 35% behind six at 80%. Report per machine; aggregate only for trend.
• Chasing the number, not the loss. OEE can be gamed — inflate planned downtime, reclassify stops, slow the ideal cycle. The moment OEE becomes a target people are punished on, the data quality dies. Reward finding losses, not hiding them.
• Quality counted at the end of the line.A part scrapped at final inspection three days later never makes it back into that shift’s OEE. Record rejects at the machine, when they happen.
• Reconstructed data. Logbooks filled at 5:55 PM record what people remember, which is never the twelve 4-minute stops. Honest OEE needs stops captured as they happen.
• Comparing across plants. Two plants with different break policies can differ 10 OEE points on identical physical performance. Compare a machine to itself.

They answer different questions and mixing them causes real damage: a plant running one shift can have 75% OEE and 24% TEEP simultaneously — excellent discipline, enormous unused capacity. The capex conversation (“do we need another machine?”) is a utilization/TEEP question; the shift-improvement conversation is an OEE question.

The uncomfortable truth: you don’t improve OEE by working on OEE. You improve it by fixing the largest individual loss, once, then repeating. The sequence that works:

• Measure honestly for a month. No targets, no blame — the goal is a trustworthy baseline and classified stops.
• Build the loss Pareto. Minutes by cause, per machine. On most floors, three causes carry more than half the loss — see the downtime tracking guide for running this well.
• Fix the top bar, only.If it’s changeover, work setup discipline; if tooling, tool-life management; if starvation, take the evidence upstream. One cause at a time, four weeks each.
• Let the trend line be the scoreboard. 5-point OEE moves from single fixes are common on floors measuring for the first time; the gains compound.

Everything above works on paper — plants ran OEE on clipboards for decades. But manual measurement has a floor: it cannot see micro-stops, it records what people remember, and it consumes supervisor hours compiling what a system should compile itself.

Automatic measurement reads state, cycle times, and part counts directly from the CNC controller — the machine’s own account of what happened. Operators add the one thing machines can’t know: why a stop happened. That split — automatic detection, human context — is the design behind ThingConnect’s OEE software, and it’s why its availability and performance numbers hold up under argument.