Bottleneck Analysis in Lean Manufacturing: A Guide
Key takeaways
- A bottleneck is the station whose cycle time exceeds takt time; confirm it with throughput, cycle time, and WIP data, not queue observation alone.
- Time-slice your data by shift and week to separate a temporary bottleneck tied to a specific cause from a chronic capacity gap.
- Apply Goldratt's five focusing steps in order: exploit and subordinate before spending money to elevate the constraint.
- Validate the constraint with the starved-or-blocked test and a one-shift perturbation, since a genuine fix moves the bottleneck to the next-slowest station.
Bottleneck analysis in lean manufacturing means collecting throughput, cycle time, and WIP data from each station, then using those numbers to confirm which step limits total system output. This guide walks the full sequence on one five-station machining cell, so each step's outputs feed the next, from data collection through constraint confirmation to the five focusing steps.
What counts as a bottleneck in lean manufacturing versus a constraint or a bottleneck symptom?
A bottleneck is the station with the least capacity relative to demand; a constraint is the single point that limits total system throughput. In most lines they are the same thing, but a constraint can also be a policy or a market, not a machine. Goldratt's five focusing steps define the constraint as whatever governs system output, which is why identification comes before improvement (lean glossary entry on the five focusing steps).
Bottleneck vs constraint
Use "bottleneck" when you can point at a work centre whose cycle time exceeds takt time. Use "constraint" when you are ready to reason about the whole system, because elevating a station only helps if it is the true limiting point.
Symptoms that mimic a bottleneck
Starved stations, batch queues, and erratic arrivals all look like capacity problems but trace back upstream. A worked example on a manufacturing line shows how WIP piling before one station can reflect an upstream changeover problem rather than that station's capacity (worked example). Confirm with cycle time data before treating any queue as the constraint.
Which metrics do you collect first to confirm where the bottleneck actually is?
Four numbers confirm a bottleneck: throughput at each station, cycle time per unit, WIP sitting in front of each station, and demand. The station with the highest cycle time relative to takt time and the largest WIP queue is your candidate constraint. If you only have one of the four, you have a guess, not a confirmation. A structured walkthrough of the minimum dataset is in our guide to bottleneck analysis software.
Sample data for the worked example
Use this five-station line as the running example for the rest of the article. Demand is 90 units/day; takt time at 7.5 productive hours is 5.0 minutes.
| Station | Cycle time (min) | Throughput (units/day) | WIP in queue (units) |
|---|---|---|---|
| Cut | 4.1 | 108 | 6 |
| Drill | 5.8 | 76 | 41 |
| Weld | 4.4 | 100 | 9 |
| Paint | 4.6 | 96 | 12 |
| Pack | 3.9 | 112 | 4 |
Drill has the longest cycle time, the lowest throughput, and the deepest queue. That is the candidate.
Collection methods that fit a one-day site visit
Time 20 cycles per station with a stopwatch and log WIP counts every hour on a clipboard or phone. Pull demand from the production schedule rather than the sales forecast. Any station where cycle time varies widely needs the full 20 samples; stable stations can drop to 10.
How do you distinguish a temporary bottleneck from a chronic one?
Time-slice the data. Take the same cycle-time and WIP measurements you collected and split them by shift, by day, and by week. A temporary bottleneck shows up as a spike tied to a specific cause: one operator out sick, a batch changeover, a late material delivery. A chronic bottleneck shows up as a persistent gap between that station's capacity and demand, repeated across nearly every slice.
A bottleneck is the station with the lowest capacity in a defined system, and it only holds that title while the system and time window stay the same (Bottleneck (production))). Change the window and the bottleneck can move, which is why a single shift's data will mislead you.
Run both views through the Constraint Impact Calculator with the same inputs. If the constraint moves between slices, you are looking at a scheduling problem, not a capacity investment, and the scoping conversation should price accordingly.
How do you analyze a confirmed bottleneck using the five focusing steps?
Step 1-2: identify and exploit
Goldratt's first two steps are cheap and fast. Identification you already have: the press in Cell 2 runs a 14-minute cycle against a 12-minute takt time, sits at 92% utilization, and has 18 units of WIP stacked in front of it. Exploitation asks what you can squeeze from that press without spending capital. In the worked example, that means cutting its changeover from 25 to 15 minutes, moving inspection off the press operator, and making sure the press never waits for material or an operator at shift start. None of these touch any other station, which is the point: an exploit only has to work on the constraint.
Step 3-5: subordinate, elevate, repeat
Subordination is where most shops flinch. The upstream lathe should stop producing to takt and produce to the press's pace, even though that drops the lathe's utilization from 78% to around 65%. A partially idle non-constraint is not waste; running it flat out just converts cash into WIP the press cannot consume. If exploitation and subordination close the demand gap, stop. If they do not, elevate: add a shift, offload one part number to the Cell 4 press, or buy capacity. Then re-measure, because the constraint moves, and the five steps start again at step 1.
A scoping report that shows the client which step you stopped at, and what the constraint will become if you continue, is what turns a day of analysis into a quoted engagement. Tools like Improve First's Constraint Impact Calculator (https://improvefirst.com/account) are built for exactly that handoff: model the exploit, show the projected throughput gain, and name the next constraint before the client asks. Where it falls short: it will not simulate minute-by-minute dispatching the way a dedicated simulator would, so treat its output as a scoping estimate, not a schedule.
What data does a bottleneck analysis require before you can trust its conclusions?
Three checks separate a defensible analysis from a guess. First, observation window: measure long enough to cover at least one full demand cycle, typically several shifts, so a lunch break or a changeover does not masquerade as a constraint. Second, variation: record individual cycle times, not just averages, because a station averaging 9 minutes with a range of 6 to 14 behaves differently from one locked at 9. Third, utilization threshold: a station above roughly 90% utilization relative to takt time is a candidate constraint, and takt itself depends on a correct demand figure in the standard calculation (takt time formula, Lean Enterprise Institute glossary). If any of the three fails, flag the conclusion as provisional and widen the sample. The Constraint Impact Calculator applies the same thresholds, but it inherits whatever data you feed it: garbage inputs still produce a confident-looking report.
How do you validate that your bottleneck analysis found the true constraint and not a downstream effect?
Three checks separate a real constraint from a symptom: where WIP accumulates, which stations starve or block, and what happens when you perturb the system.
The starved-or-blocked test
At the true constraint, WIP piles up upstream and the station itself never waits for work. Downstream stations starve; upstream stations block. In the worked example, if the mill is the constraint, the lathes upstream should show queue growth while packing downstream sits idle waiting for the mill's output. If instead a downstream station shows WIP building in front of it, that station is the constraint and your analysis stopped one step short.
Testing whether the constraint moves
Run a small intervention: add capacity or reduce setup time at the suspected constraint for one shift, with everything else held constant. If system throughput rises, you found it. If throughput stays flat, the limiting factor sits elsewhere, likely in material release or demand. Then re-run the WIP and starved-or-blocked checks, because a genuine fix moves the constraint to the next-slowest station, and your scope should follow it.
When does a bottleneck analysis justify shifting resources versus elevating the constraint?
Use the demand gap as your decision rule. If throughput at the constraint still covers current demand, the gap is a scheduling or WIP problem: exploit first. Little's Law says cycle time equals WIP divided by throughput (Little's Law), so cutting WIP at the constraint shortens lead time without buying capacity. In the worked example, the mill's 42-minute cycle against a 38-minute takt means you recover the gap through exploitation, not new spend.
If throughput sits below demand even after exploitation, elevate. That means adding capacity at the constraint: an extra shift, a second machine, or offloading one product family. Shift resources only when a non-constraint station starves the constraint; otherwise moving people just relocates the backlog and wastes the analysis.
Which tools fit a scoping-stage bottleneck analysis, and where does each fall short?
Criteria for a scoping tool
A scoping tool should confirm the constraint from throughput, cycle time, and WIP data you already collected in the steps above, cost little enough to justify before a quote, and produce a summary you can hand the client. A spreadsheet capacity model meets the cost bar and stays flexible, but it will not show where the constraint moves after you act. DO.Impact and FlowVisual cover constraint identification at scoping depth. ProcessModel runs full simulation, which is more than a scoping stage needs and usually means a bigger commitment. KaiNexus supports ongoing improvement programs, not a pre-quote analysis.
Where Improve First's calculator falls short
The Constraint Impact Calculator does three things a scoping stage needs: it finds which part of a process limits total output, tests whether an improvement changes the whole system, and shows where the constraint is likely to move next. It does not simulate a full production schedule or handle detailed resource scheduling the way ProcessModel does, and it will not replace the deeper engagement you quote afterward. If the client needs takt-level scheduling or a validated simulation model, scope that as phase two and say so in the report.
FAQ
Can a bottleneck have utilization below 100% and still be the constraint?
Yes. A constraint can be a policy or a market rather than a machine, so no station may run near 100%. Also, if your observation window is too short, a true constraint can look underutilized. Treat roughly 90% utilization relative to takt time as the candidate threshold, then confirm with WIP and starved-or-blocked checks before concluding.
How long should I observe a process before calling a station the bottleneck?
Measure long enough to cover at least one full demand cycle, which typically means several shifts. A single shift can hide or invent a constraint: a lunch break, a changeover, or a late delivery can masquerade as one. Time at least 10 cycles per station, 20 where cycle time varies widely, and log WIP hourly across the window.
Does takt time change where the bottleneck is?
The station with the least capacity stays the same, but takt time depends on demand, so a demand change shifts which stations exceed takt and how large the gap is. A station comfortably above takt at 90 units/day can fall below it at 120. Recalculate takt from the production schedule before re-running your comparison.
What is the cheapest way to run a bottleneck analysis for a small workshop?
A stopwatch, a clipboard or phone, and a spreadsheet. Time 20 cycles per station, log WIP counts hourly, pull demand from the production schedule, and compare each cycle time against takt. A spreadsheet capacity model confirms the constraint and tests an exploit at near-zero cost, though it will not show where the constraint moves after you act.
How do I present bottleneck findings to a client before quoting a full engagement?
Deliver a short scoping report that names the confirmed constraint, shows the data behind it, and states which of the five focusing steps you stopped at. Model the exploit and its projected throughput gain, name the likely next constraint, and flag any provisional conclusions. A calculator output works as a scoping estimate, not a schedule.