How to Find and Open a Production Bottleneck with Data

On many plant floors, the loudest, busiest station gets the blame for missed shipments. Supervisors see a crowd of people, a pile of parts, and a machine that never seems to stop—and they conclude: that must be the bottleneck. Often it is not. A busy station can be overfed by an upstream step that is running too fast, or it can be reworking scrap created elsewhere. The real constraint is the step that sets the pace of the whole chain: when it slows, finished goods slow; when it idles, the plant’s throughput idles with it.

This article is a practical, data-first playbook for finding that constraint and opening it—usually without buying a new machine. We stay close to the Theory of Constraints (TOC), simple shop-floor measures, and one afternoon spreadsheet exercise you can run this week. For broader process and architecture work on radsap.com, start from the English homepage or the articles library.

More videos on my YouTube channel ↗

What a bottleneck really is

In the Theory of Constraints, popularized by Eliyahu Goldratt, a system’s throughput is limited by its scarcest resource—the constraint. Goldratt’s Five Focusing Steps, summarized in accessible overviews such as the Lean Production TOC guide, are deliberately simple:

1. Identify the constraint.
2. Exploit it (get more from what you already have).
3. Subordinate everything else to that decision.
4. Elevate the constraint (invest) only if exploitation is not enough.
5. Repeat—because after you elevate, a new constraint appears.

A bottleneck in production terms is usually a process step whose available capacity is less than the demand placed on it over a meaningful horizon. It is not “the place where people look stressed.” Stress can be a symptom of poor sequencing, missing kits, or local efficiency targets that push non-constraints to overproduce. Practical manufacturing guidance on bottleneck performance—for example from Goldratt UK’s bottleneck performance discussion—emphasizes protecting the constraint’s productive minutes before adding capital.

Data signals that reveal the constraint

You do not need a full MES rebuild to start. Three linked measures, taken in the same time window for every major step, already separate noise from signal:

• WIP (work in process) — units or jobs waiting or being worked before/at the step.
• Throughput — units completed per hour or per shift at that step.
• Lead time — how long work spends in that segment of the flow.

Little’s Law connects them. In a stable system, average lead time is approximately WIP divided by throughput (with consistent units). Operations primers such as the Umbex manufacturing lead-time framing use this relationship to show why growing queues are not “just inventory”—they are time. If WIP at a step climbs for three or more consecutive weeks while throughput stays flat, you are watching a constraint form (or a release policy that is feeding it too aggressively).

Complementary signals:

• Downstream idle time that coincides with stops at the suspect step.
• Rising changeover or micro-stop minutes on one resource while others wait.
• Full-kit readiness below a healthy threshold at the gate (kits arriving incomplete).

Treat a single bad day as weather. Treat a three-week queue trend as climate.

Plant-floor walk + afternoon spreadsheet diagnostic

Before you trust the spreadsheet, walk the flow with a simple map: receiving → prep → core processes → test/pack → ship. Mark where physical queues sit. Ask operators what starves them and what blocks them. Often the data and the walk agree; when they disagree, investigate definitions (is “throughput” counted at start or at good-unit completion?).

Then spend one afternoon on a diagnostic table:

Sort by lead time and by queue growth. The step that ranks high on both is your prime suspect. Add two columns many teams forget: changeover minutes and % jobs arriving full-kit. A “slow” machine that loses two hours a day to incomplete kits is often a policy problem, not a capacity problem.

Open the constraint: exploit before you elevate

Buying a second machine feels decisive. It is also expensive, slow, and sometimes unnecessary. TOC’s exploit/subordinate steps usually come first:

1. Protect uptime on the constraint — priority maintenance windows, critical spares, and skilled coverage so the gate does not wait on a fuse or a missing technician.
2. Full kit / full information — do not release work into the bottleneck until materials, tools, drawings, and quality checks are ready. Starving the constraint mid-run burns its scarce minutes.
3. Reduce changeovers and micro-stops on the gate first — sequence similar jobs, stage tools, apply a light SMED mindset where the minutes are most expensive.
4. Subordinate non-constraints — stop local efficiency theater. If Step A can run at 140% of the bottleneck, running it flat out only grows the queue in front of the gate and hides problems elsewhere.
5. Quality at the gate — scrap and rework that consume bottleneck time are double losses: lost good output and lost capacity.

Vendor and practitioner case write-ups sometimes claim large output lifts from bottleneck focus alone; treat those as case claims, not universal constants. Your plant’s gain depends on how much hidden protective capacity you recover.

When to elevate (invest) versus keep exploiting

Elevate when, after two to four weeks of disciplined exploit actions, re-measured throughput still cannot meet credible demand. Elevation can mean people, tooling, fixtures, a process redesign, overtime policy—or yes, capital equipment. After you elevate, re-scan the chain. The old bottleneck may vanish and a new one will surface. That is success, not failure: the system’s limiting factor moved.

Do not elevate a false bottleneck identified from one busy Friday. Do not elevate a station whose queue is fed by uncontrolled release from upstream.

Illustrative example: a four-step plant

Illustrative example only — not real client data. Imagine a small discrete plant with four steps: Prep, Machine, Assemble, Pack. Weekly demand is 400 good units.

• Prep throughput: 120 units/shift; WIP stable; queue not growing.
• Machine throughput: 70 units/shift; WIP before Machine rising for four weeks; downstream Assemble often waits.
• Assemble throughput: 110 units/shift when fed; idle time high when Machine stalls.
• Pack throughput: 130 units/shift; rarely limiting.

Little’s Law on Machine shows long lead time driven by high WIP ÷ modest throughput. The walk confirms a physical queue at Machine and incomplete kits arriving twice a day. Exploit actions: kit completeness gate, changeover pairing on Machine, and slowing Prep so it does not flood Machine. After three weeks, Machine effective throughput rises (same asset), Assemble idle time falls, and on-time shipments improve—still without a new CNC. Only if demand then grows beyond the exploited Machine would elevation (extra shift or machine) be justified.

If you already use optimization and learning models for planning decisions, the same constraint logic pairs well with data-driven decision tools—see also our related article on neural networks and stochastic programming for optimal decision-making.

Conclusion and next step

Busy is a feeling; a bottleneck is a measured limit on system throughput. Use WIP, throughput, and Little’s Law lead time; watch queues for at least three weeks; walk the floor; then exploit—protect uptime, enforce full kits, cut gate changeovers, and subordinate non-constraints—before you spend capital. When exploitation plateaus against real demand, elevate deliberately and expect the constraint to move.

Download the printable bilingual checklist and run the afternoon diagnostic with your team:

Download the bilingual Bottleneck Finder checklist (PDF)

More process-engineering and data articles: radsap.com/articles · Home: radsap.com

References (external): Theory of Constraints overview — Lean Production · Bottleneck performance — Goldratt UK · Manufacturing lead time & Little’s Law framing — Umbex