Production bottleneck: how do you tell the real constraint from its symptoms?

A line is running, the operators are at their stations, the material is there, and yet the day's target is missed. Nobody really understands why. On paper, the theoretical capacity would be more than enough.

Ask where the problem is, and everyone will point somewhere different. The end-of-line operator points to the work in progress piling up in front of him. The shift leader blames a breakdown that morning. The planner suspects a material shortage.

They are all seeing something real. But only one of these points limits the throughput of the whole line. The rest are just consequences, reflections of a constraint located elsewhere.

That is the trap: the place where things pile up is almost never the place where you need to act.

What is a bottleneck, and why is the real one so often hidden?

A bottleneck is the station whose pace caps the throughput of the whole. A line moves no faster than its slowest constraint, whatever the power of the neighbouring stations.

Intuition tends to look for the constraint where the material piles up. That is logical, and it is often misleading. Work in progress accumulates in front of a station that cannot keep up, but the reason for that slowdown is frequently upstream.

Take the concrete mechanism. A station takes in material in fits and starts because the previous station feeds it irregularly. Work in progress swells at the entry of the second station. The eye blames this second station. The cause lies with the first.

• The symptom is visible, local, easy to point at.
• The constraint is often discreet, located higher up the flow.
• Starvation downstream, a station waiting for want of material, sometimes betrays a bottleneck far upstream.

Confusing the two leads to reinforcing a station that did not need it, while the real constraint stays intact.

What signs betray a bottleneck?

No single sign is enough on its own. Each one points towards a family of possible causes, to be confirmed afterwards by the data. The table below links observable symptoms to the leads they open up.

Table scrolls horizontally on small screens.

The same symptom opens several leads. A build-up of work in progress can come from the station itself as much as from a jagged, uneven feed. The table serves to formulate hypotheses, not to conclude.

Which data should you analyse to decide?

Gut feeling points the way, the data decides. Several systems each hold pieces of the answer, each with its own point of view.

• The MES traces the real pace station by station, cycle times, micro-stoppages.
• The ERP carries the production orders, the quantities, the planned production sequences.
• The CMMS records breakdowns, interventions, equipment availability.
• Quality checks flag the scrap and rework that send material round again.

At the centre of the analysis sits OEE, overall equipment effectiveness. It measures the share of time an item of equipment is actually producing good parts, at the right pace, compared with what it could have produced. It combines three losses: stoppages, pace slowdowns and poor quality.

A low OEE catches the eye, but it does not say whether the station is the constraint or merely a victim: a piece of equipment can post a poor OEE because it is waiting for material that never arrives. It is one more clue, to be cross-checked.

To go further on reading these histories, see how to exploit CMMS data without drowning in the volume.

What AI brings to the search for the bottleneck

Lean and the theory of constraints have long known how to locate a bottleneck. On a simple line, a good observer and a stopwatch are enough. The problem changes in nature when the data multiplies: dozens of stations, thousands of orders, stoppage histories over several months, several references running in parallel.

That is where AI-assisted analysis gains ground. It does not replace the method, it speeds up its execution over volumes a human cannot scan by eye.

• Detect correlations between a downstream slowdown and an upstream event that occurred earlier.
• Analyse several variables together, pace, availability, quality and sequencing, instead of looking at them one by one.
• Tell a recurring constraint apart from a one-off fluke, separating the pattern that keeps returning from an isolated chance event.
• Simulate scheduling scenarios to anticipate where the bottleneck will move if the current one is lifted.

You have to stay honest about what AI does here. It creates no truth: it links and compares what your systems have already recorded. If the recorded times are wrong or the stoppage causes badly coded, the analysis will point to the wrong place with confidence. Data quality sets the ceiling on the result.

The final reasoning stays human. The tool reads, links, aggregates and flags correlations. The production manager decides, with the context the data ignores. This division of labour structures any approach to AI in industrial production.

• Adding capacity where things pile up.You reinforce the saturated station that was only a symptom, and throughput does not budge.
• Treating the symptom rather than the constraint.Clearing the visible work in progress relieves the eye for a moment, without lifting the real constraint located elsewhere.
• Forgetting that the bottleneck moves.You optimise today's constraint and consider the matter closed, when a new one has just appeared.
• Trusting gut feeling rather than the data.The place where the team points to the problem is rarely the place where you need to act.

Why does the bottleneck never stay in the same place?

A bottleneck that is lifted does not disappear, it moves. As soon as you unblock the constraint, another station becomes the new ceiling on throughput, and the line moves as fast as its second-slowest point. This is not a failure: it is the normal mechanics of a flow, where every improvement reveals the next constraint.

Searching for the bottleneck is therefore not a one-off operation but a continuous loop, locate, lift, measure, start again. The same data serves again to track where the constraint migrates.

An untreated bottleneck also costs downstream: the fits and starts it causes generate unstable settings, rework and scrap, and feed the cost of poor quality. The same logic applies to reducing unplanned downtime, often concentrated on the constrained station.

Building this loop assumes a clean, queryable database, fed by your existing systems. That is the purpose of Assets 4.0, which builds AI systems around your data rather than the other way round.

What should you take away?

The bottleneck limits the whole flow, but its most visible symptom, the accumulated work in progress, almost always misleads about its real position. Lean and the theory of constraints give the method. The MES, ERP, CMMS and quality data give the proof, and AI cross-references them faster than a human eye over large volumes. Once the constraint is lifted, another appears: it is a loop, never a single blow.