Lean Manufacturing & Production Systems: The Complete Engineering Guide — Little's Law and the Kingman Equation Worked on a Job Shop, Why 90 % Utilisation Triples Lead Time, OEE Decomposed from Shift Data to 78.4 % Against the 85 % World-Class Standard, Takt Time and Line Balancing, the Square-Root Law of Batch Size and Shingo's SMED, Safety Stock and Cpk Worked in Numbers, the Theory of Constraints, Kanban Sizing, and a 2026 Rupee Ledger from ₹500 VMC-Hours to Inventory Carried at 22 % a Year

Lean manufacturing and production systems in numbers: Little's Law and the Kingman equation on a job shop, OEE worked from shift data and benchmarked at 85%, takt time, EOQ and SMED batch economics, safety stock, Cpk, Theory of Constraints, kanban sizing, and a 2026 rupee ledger for Indian machine shops.

·

Lean Manufacturing & Production Systems: The Complete Engineering Guide

Every guide in this series so far has been about a machine. The steam boiler, the pressure vessel, the CNC spindle, the conveyor — each one is a solution to a physics problem, and each one can be specified, calculated and inspected until its behaviour is known to the third decimal. This guide is about the other half of manufacturing, the half that has no spindle and no datasheet: the flow of work through the factory that owns those machines. It is the half where most of the money is won and lost, and — for an astonishing number of otherwise excellent shops — the half that runs on folklore.

Watch two machine shops with identical equipment lists. Both have three VMCs and a turning centre, both have competent operators, both can hold ±0.05 mm all day. Shop A quotes three weeks and delivers in two; Shop B quotes two weeks and delivers in five. Shop A's machines sit at about 75 % loaded with a lean queue in front of each; Shop B's are at 95 %, "fully utilised", with forty jobs stacked on pallets waiting their turn. Shop A's owner can tell you the shop's OEE, its on-time delivery, and the value of its WIP; Shop B's owner knows the machines are "busy" and little else. The difference between the two shops is not capability. It is production systems engineering: queueing theory, capacity arithmetic, batch economics and measurement discipline — and the arithmetic in this guide is the reason Shop A wins the repeat orders.

The stakes are large because the audience is large. India has roughly 1.6 crore manufacturing enterprises registered on the Udyam platform — small and micro workshops, most of them job shops — inside a sector of some 7.6 crore registered enterprises contributing 31.1 % of GDP, 35.4 % of manufacturing output and 48.58 % of exports as of January 2026 [1]. It is also a sector under margin pressure that would make a process engineer wince: in Coimbatore, one of India's densest machining clusters, CNC job-work rates that should support ₹350 per hour are being quoted at ₹150, and VMC work that needs ₹550 per hour is being bid at ₹250 [17]. Meanwhile the average discrete-manufacturing plant runs at about 60 % OEE — Overall Equipment Effectiveness — against an 85 % world-class benchmark, a gap worth 41 % more output from the same machines, people and shifts [4][5]. Shops squeezed on price while carrying 40 % of their capacity in unproductive time are not short of effort. They are short of the arithmetic that tells them where the capacity and the money actually go.

This guide works production systems the way the rest of the series works machines — in numbers, end to end:

1. The Factory Is a Queueing Network, Not a Machine

1.1 Little's Law — the only equation you cannot argue with

John Little proved in 1961 what every dispatcher has felt in their bones: for any system in steady state, the average number of items in the system equals the average arrival rate times the average time each item spends inside [7]:

where WIP is work in process (jobs, parts, orders — whatever unit you choose), TH is throughput (the same unit per unit time) and CT is the average time from start to finish (cycle time, in the plant's real calendar). The equation is not an approximation, does not assume any distribution, and does not care about priorities, ERP systems or good intentions. It is a conservation law, and it says something brutal about manufacturing: in steady state, if you hold WIP and complete TH per day, the lead time CT = WIP/TH is already decided. The only question is whether anyone has told the customer.

Run the numbers on a real job shop. Forty-five jobs in process across the shop; the shop completes nine jobs a day. Then

Not three. Not "two if we really push". Five, until either the numerator falls or the denominator rises. This is the first hard lesson of production systems: promising shorter lead times is impossible without changing WIP or throughput — and since throughput is capped by the bottleneck (§7), in the short term the lever is almost always WIP. Shops that "expedite" are not reducing CT; they are moving one job's wait onto other jobs, keeping the average exactly where Little's Law put it.

1.2 Critical WIP — how little work the shop needs to be full

The throughput ceiling is the bottleneck rate r_b: the completions per hour (or day) of the slowest rate-limiting resource, measured in effective hours, not nameplate hours (§3 will show why that distinction is worth 25 % of the capacity). The second fundamental constant is the raw process time T_0: the total of all the touch times along a job's route, scaled to the bottleneck's pace — the time a single job would take if it flowed through an empty shop behind it. Multiply them:

W_0 is the critical WIP, and it has a beautiful interpretation: it is the least work in process that still keeps the bottleneck busy. Below W_0 the shop starves the bottleneck and throughput falls; at W_0 the shop achieves its maximum throughput at the minimum possible cycle time T_0; above W_0, throughput stays flat at r_b and every additional job goes straight into cycle time.

Work it. A six-machine shop; the bottleneck is a VMC cell with an effective rate of r_b = 3 jobs/hour (six-minute average effective cycle — again, effective, not ideal). A typical job's routing sums to T_0 = 4 hours of machine-plus-handling touch time at the bottleneck pace. Then W_0 = 3 \times 4 = 12 jobs. Now trace the policy:

WIP in shop · Throughput (jobs/h) · Cycle time · What the shop feels like

6 (half of W_0) · 1.5 · 4 h · Starved: machines idle, "we need orders"

12 (W_0) · 3 · 4 h · Full throughput, minimum lead time

18 · 3 · 6 h · Comfortable, one queue per machine

30 · 3 · 10 h · "Busy"; two-week promises starting to slip

60 · 3 · 20 h · Gridlock; expediting everywhere; OTD < 60 %

Read the third and fifth rows again: doubling WIP from 30 to 60 adds zero throughput and adds ten hours — 2.5 working days — of lead time to every job in the shop. This table is the mathematical form of the difference between Shop A and Shop B in the introduction. Shop B at 95 % utilisation is not producing more; it is producing the same output with twice the lead time, twice the WIP capital and a fraction of the delivery credibility (§10 prices that difference).

1.3 The three levers, and only three

Everything in production systems engineering pulls one of three levers on Little's Law and the critical-WIP pair:

  1. Cap the WIP (a CONWIP card count, a kanban system, a "don't release what you can't finish" discipline — §8): lead time becomes a decision — CT = WIP/TH — instead of an outcome;
  2. Raise the bottleneck rate r_b (SMED, OEE improvement, debottlenecking, a second shift on the constraining resource — §3, §5, §7): the only lever that raises throughput, because a non-bottleneck's improvement cannot;
  3. Cut the raw process time T_0 (faster programs, better tooling, lights-out unattended time, fewer operations — the domain of the machine guides in this series).

Every lean tool ever invented is a vehicle for one of those three. Everything else — the whiteboards, the prioritisation meetings, the ERP with a "rush" flag — is negotiation about which job suffers, not engineering of the system.

2. The Variability Tax — Kingman's Equation, Worked

2.1 Why queues exist even at 80 % load

A shop at 80 % utilisation has, arithmetically, 20 % spare capacity: in a perfectly deterministic world, work would never wait. Real shops queue constantly, and the reason is variability — in when jobs arrive, how long they take, which machine breaks, whether the customer's drawing was readable. The science of quantifying that queues-up-around-variability comes from queueing theory, and for manufacturing it was packaged into its most useful form by Hopp and Spearman in Factory Physics [7][8] as the VUT equation (also known as Kingman's equation for the G/G/1 queue):

Unpack it symbol by symbol because every one of them is a management lever:

Three facts fall straight out of it. Queues scale linearly with variability — halve the squared CVs and you halve the wait. Queues scale hyperbolically with utilisation — the u/(1-u) factor. And queues scale with the process time itself, which is why the same policy that lets a 30-second station tolerate 95 % load will drown a 40-minute CNC operation at 95 %: the absolute wait is t_e-multiplied.

2.2 The hockey stick, in minutes

Assume a "normally messy" cell — c_a^2 = c_e^2 = 1, so V = 1 — processing jobs with a mean effective time t_e = 40 minutes. Then CT_q = 40 \times \frac{u}{1-u} minutes:

Utilisation u · U = u/(1-u) · Queue time CT_q · Throughput margin left

50 % · 1.0 · 40 min · Huge — machines look idle

70 % · 2.33 · 93 min · Comfortable; deliveries safe

80 % · 4.0 · 160 min · Workable; lead times stable

85 % · 5.67 · 227 min · Fragile; one breakdown cascades

90 % · 9.0 · 360 min · 6 h of queue per 40 min of work

95 % · 19.0 · 760 min · Gridlock: 12.7 h of wait per job

Trace the central comparison with a finger: 80 % → 90 % utilisation, a ten-point change at the load dial, triples the queue — 160 minutes to 360 minutes. Nothing about the machine changed. No new jobs were added (utilisation is a ratio). The shop simply accepted more work relative to its effective capacity, and the price is paid in lead time and WIP, which by Little's Law is the same statement wearing different clothes: at 3 jobs/hour throughput, the extra 200 minutes of queue multiplies out to about ten extra jobs standing on the floor at any moment — a lakh-plus of WIP on light machined work, several lakh on heavy fabrication — plus the delivery slippage.

2.3 The variability sources, ranked by what you can fix

The V term is where engineering lives, because c_a^2 and c_e^2 are measurable and attackable:

Source · Typical c^2 effect · The honest fix

Job release bursts (all orders released Monday) · c_a^2 1 → 4 · Level the release (CONWIP, §8); a queue cap, not discipline

Batch arrivals from upstream · c_a^2 up · Smaller transfer batches (§5)

Machine breakdowns · c_e^2 up 50–200 % · TPM, spares for the bottleneck, §3

Rework loops · c_e^2 up · First-pass yield programs (§6)

Setup-time variance ("20 minutes, sometimes 3 hours") · c_e^2 up · SMED standard work; the variance is often worse than the mean

Missing tooling, drawings, programs · c_e^2 up · 5S + kitting; the cheapest single fix in most shops

Operator-dependent cycle times · c_e^2 up · Standard work, tool presetters, probing

Two practical rules follow. First, if you want short lead times, plan the shop to run at 75–85 % of effective capacity, and treat the remaining 15–25 % as the buffer you chose — because the law of variability buffering (Hopp & Spearman) says the only question is which buffer you pay for: inventory (WIP), capacity (spare hours), or time (long lead times) [7][8]. Every shop pays in one of the three, and capacity is usually the cheapest by a factor of several; §10 does that arithmetic in rupees. Second, variability reduction has the same effect as capacity, at a fraction of the cost — halving V is worth more than ten points of utilisation anywhere in the 80–95 % danger zone.

2.4 Variability propagates — and so do slowdowns

A warning for anyone planning a multi-station line: the VUT equation applies per resource, and queues compose. If a job crosses six work centres, the total cycle time is T_0 + \sum CT_{q,i}, and the variance arriving at station n+1 reflects everything that happened upstream (the linking equation: c_a^2(n+1) \approx variability out of station n, amplified by each queue) [7][9]. The composed service levels are equally unforgiving: six resources that each deliver their part on time 90 % of the time, with independent failures, give an overall promise of 0.9^6 = 53\% — almost half of all orders late, without a single negligent operator. This is why the delivery promise of a long routing must be built from per-station capacity provision, not from end-of-line heroics, and it is why manufacturers running job shops in the FabFlow network price and quote lead times from capacity math (§10) rather than from optimism.

3. OEE, Decomposed — The Machine's Report Card, Done Honestly

3.1 Nakajima's instrument

Overall Equipment Effectiveness was built at Nippondenso — a Toyota group supplier — in the 1960s, formalised by Seiichi Nakajima and the Japan Institute of Plant Maintenance (JIPM) as the measuring instrument of Total Productive Maintenance, and standardised for the software age in ISO 22400-2 [3][20]. Its structure is deliberately multiplicative:

Availability A = \dfrac{\text{run time}}{\text{planned production time}} — the machine was scheduled and did not run (breakdowns, changeovers, waiting for material, waiting for an operator) is subtracted.

Performance P = \dfrac{\text{ideal cycle time} \times \text{total count}}{\text{run time}} — the machine ran, but slower than its design speed, or in micro-stops nobody logged.

Quality Q = \dfrac{\text{good count}}{\text{total count}} — parts were made, but some were scrap or rework (including the first-piece scrub after every changeover).

The multiplication is the point. OEE does not let excellence at one factor hide failure at another: three strong-looking numbers — 90 % availability, 95 % performance, 99.9 % quality — multiply to 85.4 %, which is why the world-class benchmark is conventionally 85 % and is composed of exactly those three component targets [5][6]. Nakajima's framework also names the six big losses the three factors catch: breakdowns and setup/adjustment (availability); minor stops/idling and speed loss (performance); startup reject defects and production defects/rework (quality).

3.2 A worked shift, harvested from the floor

Here is a single eight-hour shift on a production VMC, tracked properly — one machine, one part number, numbers recorded at the machine rather than reconstructed by the ERP:

Now the three factors:

And because OEE is a ratio of times, we can render the whole shift as a waterfall of minutes, which is where OEE stops being a score and becomes a to-do list:

Loss bucket · Minutes · Share of shift · Where it actually came from

Planned production time · 450 · 100 % · —

Availability loss · −60 · 13.3 % · 38 min breakdown + 22 min changeover

Performance loss · −19.2 · 4.3 % · 16 pieces short of the ideal, at 1.2 min each

Quality loss · −18 · 4.0 % · 15 defect units × 1.2 min

Productive minutes (good parts) · 352.8 · 78.4 % · 294 \times 1.2 min

Check the integrity: 450 - 60 - 19.2 - 18 = 352.8 min \Rightarrow 294 good pieces. The waterfall is conservative in one direction and generous in another: it does not count the 4 reworked pieces' rework time as a loss beyond their first pass (a fuller accounting, OEE's cousin Overall Resource Effectiveness, would), and it does not count the material scrapped. But it converts a fuzzy "decent shift" into an engineering statement: on this shift, the machine lost 60 minutes to availability, 19 minutes to speed, and 18 minutes to defects — start with the 38-minute breakdown. That is what the instrument is for.

3.3 The benchmark ladder — and where Indian MSMEs sit

Do not evaluate 78.4 % in a vacuum. The industry reference points:

OEE · Reading · Where it's found

85 %+ · World class (A ≥ 90, P ≥ 95, Q ≥ 99.9) · Top-tier, heavily automated lines [3][5][6]

75–85 % · Above average · Competitive shops with real measurement discipline [4]

65–75 % · Typical discrete manufacturing · Tier-1 automotive supplier territory: 70–80 % [6]

60 % · The industry median (MESA) · Where the average plant in a survey actually lives [4][18]

40–55 % · Common in small shops without data · MSMEs often run 40–50 % and don't know it [18]

Sector ranges run from pharma at 40–60 % (batch validation, cleaning) and food at 55–70 % (allergen changeovers, sanitation) up through packaging 60–75 %, discrete 65–75 %, automotive 70–80 %, with mining 50–70 % [6]. The single most useful sentence about all of this: the median shop at 60 % OEE is leaving 41 % of its output capacity on the floor (85/60 = 1.417), and in a market where job-work rates in some clusters are being pushed 40–50 % below sustainable levels [17], that capacity is not a nice-to-have — it is the difference between the ₹4.50/minute that covers a small shop's fully loaded cost and the ₹2.50/minute that does not (§10).

3.4 What ten OEE points are worth — the capacity arithmetic

Take a 20-machine shop running one shift, and suppose the shop measures honestly at 78.4 % (the worked shift). Five improvement projects — TPM on the two worst machines, SMED on the changeover-heavy cells (§5), first-pass-yield work (§6), and data capture to find the losses — plausibly lift it to 85 %. What has actually been bought?

On 20 machines that is 1.68 machines' worth of capacity — created without a single capital purchase. At mid-range Indian VMC prices of ₹12–30 lakh per machine installed (₹18 lakh is a fair centre, §10), the replacement value of that capacity is roughly ₹30–50 lakh; counted as billable hours, 20 machines × 4,000 effective hours/year × 8.4 % ≈ 6,700 machine-hours recovered annually, which at even ₹500/hour of contribution is ₹33 lakh a year. The improvement projects typically cost ₹2–5 lakh in fixtures, training and data tools. This ratio — capacity for five percent of the capex per point — is why OEE work is the highest-return engineering activity in most small shops, and why the shops in the FabFlow network that track it win the price conversations: they know their true cost per hour and their true capacity, so they bid ₹400 at a profit while the folklore-driven shop bids ₹400 at a loss.

3.5 TEEP — the loading question under the OEE question

OEE judges only planned time. A machine scheduled 8 h/day, 5 days inherits a flattering denominator: idle nights and weekends never show up as losses. TEEP (Total Effective Equipment Performance) fixes this by using calendar time as the base (8,760 h/year):

The worked shift's machine: one shift, 450 min of a 1,440-minute day → loading 31 %; TEEP = 0.784 \times 0.31 = 24.3\% of calendar capacity. Frame the strategic question with TEEP, not OEE: a shop at 78 % OEE on one shift and a shop at 60 % OEE on three shifts can deliver similar monthly output — but the three-shift shop spends 3× the labour and power and carries 3× the wear. The decision between "raise OEE" and "add a shift" is a TEEP-versus-cost question, and it has a clean answer every time the numbers are written down: shift addition multiplies variable cost with the fixed cost already sunk, OEE improvement multiplies neither.

3.6 How OEE goes wrong — the four classic misuses

Averaging. A cell Tuesday 92 %, Wednesday 45 % is not "68.5 % on average" — it is a 47-point stability problem that no average can express. Report minimum and spread, not just the mean; variance is the enemy (§2), and hiding it inside an average is how the enemy survives audits.

Gaming. Once OEE is a target, it becomes a market. Classic plays: shrink the denominator by declaring "planned downtime" for everything inconvenient; declare micro-stops "not downtime, just slow running" (they shift from A to P and dilute across counts); reclassify rework as "good, for now". The defence: the six big losses must be logged individually, at the machine, by the operator — OEE is a data-collection discipline first and a score second.

Comparing across machines. A 45 % OEE on a 4-minute-cycle press and a 90 % on a 40-second-cycle wrapper are not comparable, and neither is a machine that runs one high-runner part number against one running forty. Benchmark against a machine's own history and its own design limits; use the industry ladder only for gross triage.

Optimising OEE off the constraint. The most expensive misuse, and the bridge to §7: raising OEE at a non-bottleneck does not add one rupee of throughput — it adds WIP, shortening nothing. On the constraint, every OEE point is throughput; everywhere else, it is inventory. The shop-appropriate behaviour: measure OEE everywhere, act on it at the bottleneck first, always.

4. Takt, Capacity and Line Design — Demand, Turned Into Seconds

4.1 Takt time — the heartbeat arithmetic

Takt time (German for the conductor's beat) is the production pace at which supply exactly meets demand — no more, no less:

Work it on a real product line: demand 15,000 pieces/month, 25 working days, one shift of 450 minutes per day. Then daily demand is 600 pieces and

Every station in the line must complete its work on a unit within 45 seconds on average. Takt is not a target or an aspiration — it is demand expressed in seconds, and the line's design cycle time must sit below it with margin, or the shortfall is arithmetic, not effort. Contrast the job-shop world of §1, where work arrives in unpredictable batches and "takt" as such does not exist: there, the equivalent discipline is capacity provision per work centre plus lead-time quoting; a job shop does not need every station to hold a beat, it needs enough effective capacity that the Kingman queue stays inside the promised date (§2).

4.2 From takt to a hiring decision — the OEE chain

Takt alone says nothing about how many machines. The bridge from demand to equipment — and the place most business plans quietly cheat — runs through OEE. Suppose our line's critical operation is done on a CNC that can nominally complete a piece in 40 seconds (rated cycle), and the shop measures OEE 78.4 % (the §3.2 shift). The machine's effective pace is

The line cannot keep up — daily effective output is 27{,}000/51 = 529 pieces against a demand of 600, a 12 % shortfall that will present itself as permanent overtime, expedited shipments and an exhausted planner. The OEE chain now hands the shop three priced options, in ascending order of cost:

  1. Raise OEE to the takt — required: t_{eff} \le 45 \Rightarrow OEE \ge 40/45 = 88.9\%. From 78.4 %, that is ten points; by §3.4's arithmetic, the cheapest option by far, and achievable in the §3.2 shift by fixing the breakdown band and the changeover band alone.
  2. Buy capacity — a second machine at ₹18 lakh (plus tooling, operator, floor) to cover a 12 % gap you could have engineered away.
  3. Add time — the real requirement is 51/45 = 1.13 machines; 13 % more scheduled hours (overtime on 2 of 25 days — but overtime also multiplies defect rates and, in the queueing math of §2, buys only 13 % throughput at the price of 13 % less maintenance window).

Option (1) wins on the arithmetic every time, which is the deep point of the OEE chain: takt converts demand into seconds, and OEE converts nameplate capacity into the capacity you actually own. Quotes, hirings and capex decisions made without both numbers are guesses wearing a spreadsheet.

4.3 Line balancing — when the work content won't fit the takt

Given a takt, the line design problem is to assign elementary tasks — with their precedence constraints — to stations so every station sits under takt. The first check is the theoretical minimum station count:

and the quality of any given assignment is the line balance rate:

Work a small example. Six tasks: A = 25 s, B = 18 s, C = 21 s, D = 26 s, E = 14 s, F = 14 s; total work content \sum t_i = 118 s; takt 30 s. The bound gives N_{min} = \lceil 118/30 \rceil = 4 stations — but now try to actually pack the set into four 30-second stations. Each station can hold at most one of the 20-second-plus tasks (A 25, C 21, D 26) plus, at best, its complement: 26 wants ≤ 4 more; 25 wants ≤ 5; 21 wants ≤ 9. The only pair possible at all is E + F = 28. So the feasible packing is five stations: {D 26}, {A 25}, {C 21}, {B 18}, {E + F 28}. Then:

Twenty-one percent of every cycle is balance loss — operators waiting at the short stations. The lessons are three. N_{min} is a bound, not an answer: task lumpiness and precedence constraints routinely cost a station. Balance is cheapest to fix at the design stage — combining tasks into kitted sub-assemblies, or splitting the 26-second task (a fixture that lets two stations work in parallel) would recover most of the 32 seconds of waste per cycle, worth 0.32/30 = 1.1\,\% of the line's labor cost directly, plus the capacity. And re-balancing beats pushing: when demand forces takt down to 24 s, the same work content needs 118/24 = 4.92 \rightarrow 5 stations at 118/(5 \times 24) = 98.3\% utilisation — the arithmetic says the line is already nearly optimal at that pace, and the honest options are a sixth station, task re-composition, or overtime.

4.4 Heijunka — smoothing the demand side of the equation

Takt assumes level demand, and real demand is not level: month-end pushes, seasonal rushes, one customer's panic week. Heijunka (production smoothing) attacks the problem from the demand side — sequence and level the mix so the line produces every product every interval in the same rhythm, absorbing peaks with finished-goods buffer rather than with chaotic changeovers and overtime. The arithmetic that makes leveling feasible is EPEI (§5.4): a line that runs every part every day must change over constantly, and constant changeover is only economic when setups are minutes, not hours. Heijunka and SMED are two halves of one idea: you cannot level the schedule until setups stop being scary.

5. Batch Economics — EOQ, SMED, and the Square-Root Law

5.1 The setup cost nobody measures

Every batch decision in a factory is a trade between two costs: the cost of changing over (setup labour, lost machine time, startup scrap, first-piece inspection) and the cost of holding (capital tied in inventory, space, handling, obsolescence). Manufacturing's classical answer to that trade is the Economic Order Quantity:

with D the annual demand for the part, S the setup cost per batch in rupees, and H the holding cost per unit per year. The formula's most quoted property — take the derivative, set to zero — is less interesting than its most neglected input: S. In most Indian job shops S has never been measured. It is implicitly set to whatever the last ERP implementation guessed, or to zero ("changeovers are just part of the work"), and either error distorts every batch size on the floor.

The working definition: S = (setup labour + lost contribution margin of the protected machine time + material scrapped during ramp-up + inspection of first pieces), per changeover. On a VMC with a 90-minute true changeover quoted at ₹600/hour of contribution, plus ₹300 of startup scrap, the setup cost works out at S \approx 90/60 \times 600 + 300 — ₹1,200 — and often twice that once a busy bottleneck is the machine being changed. Measure it by stopwatch (or phone video, which is less confrontational and more accurate) — then the economics become visible:

5.2 The worked EOQ — and what SMED does to it

A turned part with annual demand D = 12{,}000 pieces, setup cost S = ₹2,400 (the CNC lathe's real changeover plus ramp-up scrap), holding cost H = ₹120/piece-year (material ₹600 at ~20 % carrying rate, §10). Then

about 17 working days of supply at 40 pieces/day. The optimal annual cost of setup-plus-holding is

split evenly — ₹41.6k of setups (17.3 changeovers × ₹2,400) and ₹41.6k of holding (346 pieces average × ₹120). That even split is a general property at the optimum, and it is exactly why setup reduction is the highest-leverage action in the factory: it shrinks both halves of the trade simultaneously. Now send the lathe through a real SMED project (§5.3) and cut S from ₹2,400 to ₹600 — a 20-minute changeover instead of 90, no other change:

Read what just happened, because it is the square-root law doing the work: quartering the setup halved the economic batch — and halved the total cost of the setup-versus-inventory trade. And the consequences compound through everything in this guide:

The square-root law also runs in reverse, and it explains a common regret: if a shop increases batch sizes to "save setups" — say 4× larger batches — it pays 2× the turnover of working capital in holding cost, buys only a 50 % cut in setups, and adds roughly \Delta Q/2 = ~1,000 pieces of standing inventory per part number, with lead times stretching proportionally. The ledger of §10 makes the same point at shop scale: batch-size inflation is how inventory silently replaces profit.

5.3 SMED — from Shingo's four hours to three minutes

Single-Minute Exchange of Die is Shigeo Shingo's method, developed at Toyota and Mazda from the 1950s through the early 1970s, for cutting changeover time into single-digit minutes — "single minute" meaning under ten, not literally one [10][11]. The legend is a bolt-forging press changeover driven from 4 hours to 3 minutes, and the mechanism was not heroics but a single structural insight: most of what happens during a changeover does not require the machine to be stopped. Shingo split setup work into:

In an unimproved changeover, 30–50 % of the stopped time is external work that was simply postponed into the stop [10][12]. The SMED sequence, in order of return:

Stage · Move · Typical gain

1. Separate · Classify every element IED/OED (video the changeover; timestamp it) · 30–50 % immediately

2. Convert · Redesign so internal work becomes external: presetters, duplicate fixture plates, pre-heated dies · further 10–25 %

3. Streamline internal · Quick-release clamps, zero-point (ball-lock) fixture systems, standardised bolt heights, one-turn fasteners, no re-indicating · further 10–30 %

4. Streamline external · 5S the changeover cart, kitting, shadow boards, tool pre-staging · operator effort & variance

5. Parallelise · Two people, choreographed like a pit stop (F1 stops went from minutes to ~2 seconds on exactly this logic) · 30–50 % of the remainder

6. Standardise · Written standard work, audit the actual vs standard, sustain · keeps all of the above

Total typical reduction: 50–90 %, mostly without capital [10]. The right target is not dogma: cutting a 90-minute CNC changeover to 25 minutes captures most of the economic benefit — the batch-size math is a square root, so the first quarters of setup time pay the most [11][12]. And note the variance effect from §2: changeover times that swing "20 minutes, sometimes 3 hours" contribute c_e^2 to every queue behind them. A standardised 40-minute changeover beats an average 35-minute one, because the queue equation punishes variance, not the mean.

5.4 EPEI — every part every interval

The manufacturing expression of a shop's batch behaviour is EPEI: the interval in which the shop can cycle through all its part numbers. Its arithmetic is transparent:

A shop with 40 active part numbers, average setup 20 minutes, and 2,400 available minutes per week: a full rotation costs 40 \times 20 = 800 minutes of setup — a third of the week's machine time. That single ratio explains most scheduling pathology: with setups eating a third of capacity, the shop must run long campaigns, which means inventory, which means lead times, which means the expediting that eats the rest of the week's attention. Now put the same 40 parts through SMED to 5 minutes each: a full rotation costs 200 minutes — a twelfth of the week — and the rotation interval collapses. Where the shop ran a 3-week campaign cycle, it can now run daily or every-other-day small lots, with lead times following Little's Law down and inventory following batch sizes down. EPEI is the single number that summarises a shop's flow capability, and moving it from 15 days to 3 is worth more than any machine upgrade in the catalogue — because it is the one that changes the shape of the curve in §2 instead of sliding along it.

5.5 What a batch size really buys — and the one-piece ideal

Consolidate the mechanics into one statement: inventory is time. A batch of Q sitting between two operations embodies roughly Q \times t_{downstream}/2 of average waiting; halve the batches and you halve the shop's embodied lead time, its WIP capital, its concealed defect buffers and its change-friendliness — the last of which is strategic. Smaller batches also shrink the transfer quantity decision: even where a process needs a full heat-treat load or a plating barrel's worth of parts (a process batch), the transfer batch moving between the preceding and following operations can be split, and the queueing benefit is realized immediately. One-piece flow — the Toyota ideal — is the limit of this process, and it is a direction of travel, not a ceremony: every step toward it (dedicated cells, single-minute setups, right-sized equipment, sequenced sub-assemblies) pays at the same square-root exchange rate. The shop that halves batches twice in three years turns inventory over four times for the same demand, without buying anything.

6. Buffering Variability — Safety Stock, Capability, and the Cost of Poor Quality

6.1 Safety stock, for the real world (where lead times vary too)

§2 established that variability must be buffered by inventory, capacity, or time. The inventory buffer is safety stock, and the textbook formula most people carry around — SS = z\,\sigma_d\sqrt{L} — quietly assumes the supply is perfect: lead time fixed, supplier never late. In the real world — GST portals, monsoon roads, a supplier's furnace reline — lead time varies, and the correct formula carries both terms:

where d and \sigma_d are mean and standard deviation of demand per day, L and \sigma_L mean and standard deviation of replenishment lead time in days, and z is the service-level factor from the normal table:

Service level · 90 % · 95 % · 97.5 % · 99 % · 99.9 %

z · 1.282 · 1.645 · 1.960 · 2.326 · 3.090

Work it. A fabrication shop consumes a standard fastener at d = 120/day with \sigma_d = 30; the supplier's lead time averages L = 9 days with \sigma_L = 2 days; the shop runs a 95 % service level. Then:

Now read the two numbers that matter. First, 87.7 % of the variance sits in the lead-time term (57{,}600 / 65{,}700): the shop's stock problem is a supplier problem, not a demand problem — and the naive formula would have hidden that, prescribing only 1.645 \times 30 \times 3 = 148 pieces, less than a third of the real requirement. Suppliers must be managed by \sigma_L as an engineering quantity, measured like any other. Second, the cost of service is convex: going from 95 % to 99 % service multiplies safety stock by 2.326/1.645 = 1.41 — +41 % more capital for the last four points of service. The correct service level is set by comparing stockout cost (lost contribution, expediting, goodwill) against carrying cost — not by an inherited "we keep two weeks" habit. At ₹800/piece, the 422-piece buffer is ₹3.38 lakh of capital at ~22 % carrying (§10) ≈ ₹74k/year just to buffer this one item, and halving supplier lead-time variance from 2 days to 1 day would cut the requirement to 247 pieces — ₹1.98 lakh and ₹43k/year returned, from a supplier-management project rather than a spreadsheet cell.

6.2 Cpk, the sigma ladder, and what "capable" means in rupees

Quality variability is the last major input to c_e^2 — and its measurement instrument is process capability. Two indices, one story. Cp measures the process spread against the tolerance band:

Cpk measures spread and centring — the distance from the mean to the nearer spec limit:

Work a real number on a CNC grinding operation: target \varnothing 25.000 mm, tolerance \pm 0.05 mm, machine running at μ = 25.012 mm, σ = 0.012 mm:

The machine's spread is fine — but it is running 12 μm off centre, and the shop pays for that in the Cpk. The nearest limit is now 3.17 standard deviations away, exposing roughly 760 defective parts per million — at a 1,000-piece/month volume, that is ~9 units a year scrapped from centring alone. The fix is worth real money and usually costs a probe cycle: shift the mean to target (\mu = 25.000) and the two limbs of the minimum equalise at 0.050/0.036 = 1.389, moving the exposed defect rate from ~760 ppm to ~30 ppm — the same machine, the same σ, ₹0 of grinding improvements, one tool-offset discipline. This is the single most common quality-dollar in a machine shop, and it is invisible on any chart that only plots scrap counts.

The sigma ladder that quality programs talk about is Cpk wearing different units:

Level · Defects per million (with the classic 1.5σ drift) · Cpk equivalent · What it reads like on a shop floor

±3σ · 66,807 · ~0.5–0.67 · Inspect-in quality; sorters at the end of the line

±4σ · 6,210 · ~1.0 · "Good vendor" by old standards

±5σ · 233 · ~1.33 · Modern minimum for critical dimensions

6σ · 3.4 · ~2.0 · The brand name — and a system, not a certificate

Two disciplines support the ladder: SPC (statistical process control — \bar{X}-R charts, control limits at ±3σ of the statistic, Western Electric run rules) to detect shifts before they become defects; and the distinction between first-pass yield and rolled throughput yield: five sequential operations each at 95 % FPY deliver 0.95^5 = 77.4\% — the arithmetic that kills "each step is fine" arguments and explains why in-process inspection can never inspect its way to good economics.

6.3 The cost of poor quality — the biggest line item nobody books

Quality's costs are traditionally split into prevention, appraisal, and failure — and their empirical ratio is brutally one-sided: the 1–10–100 rule, where a defect caught at design costs ₹1, at the machine ₹10, and at the customer ₹100 (returns, freight, rework, reputation; the SME literature's customer-loyalty "1-5-12" variant prices re-acquiring a customer at 12× keeping one [13]). Industry benchmarking puts manufacturing's scrap and rework at up to 2.2 % of annual revenue for the average performer — and 0.6 % for top performers; the full cost of poor quality, including inspection, warranty and lost goodwill, is commonly estimated at 5–30 % of revenue [14][15].

Run the §3.2 shift through this ledger. Eleven scrapped + four reworked pieces per shift is 15/309 = 4.9 % of output — but defects cost fractionally more than nothing and less than the sale: if the part sells at ₹400 and carries ~₹180 of unrecoverable material, labour and margin in a scrap event, the shift burns ₹2,220 (11 \times 180 + 4 \times 60); annualised over 250 shifts: ≈₹5.5 lakh. Against the line's ~₹3.1 crore of output, that lands right where the studies say: ~1.8 % of revenue in visible scrap and rework alone — before the customer caught what the gauge didn't. And the improvement economics follow directly: the re-centring exercise above, the SMED standardisation, and a first-piece check discipline together might cost ₹1.5 lakh and recover two-thirds of that ₹5.5 lakh — payback inside five months, on the line's least glamorous numbers.

7. The Theory of Constraints — Manage the Bottleneck, Ignore the Noise

7.1 Goldratt's law of the weakest link

Eliyahu Goldratt's Theory of Constraints (1984, The Goal) states the fact that §1-§2 have been circling: in any chain, throughput is set by the single weakest link, and improving anything else is expensive theatre [19]. The shop arithmetic version: if a five-station routing has weekly capacities of 200, 160, 145, 180 and 190 jobs against a market demand of 180 jobs, the shop ships 145 — and the other four stations are not "efficiently utilised", they are partly wasted in a way no utilisation report will ever admit, because station efficiency is measured against itself, not against the system's output. Station 2's manager at 160/160 = 100 % utilisation feels excellent; the company ships 145.

The five focusing steps are the operating discipline: identify the constraint (here, station 3 — the one whose load sits highest against its capacity in effective hours); exploit it (never starve it, never feed it scrap: protective capacity upstream, quality gates before it, no experimental jobs on it, work through breaks); subordinate everything else to it (the release of work is paced by the constraint, not by the ERP's wish list); elevate it (overtime, a second shift, a duplicate machine, debottlenecking projects); and repeat — because the constraint moves once the old one is broken.

7.2 Worked — eleven hours of overtime versus everyone-works-Saturday

Our shop at 145 jobs/week: the gap to demand is 180 - 145 = 35 jobs/week. The bottleneck's effective rate is 145/45 = 3.22 jobs/hour against a 45-hour week. To reach 180 it must run

At ₹600/hour, that is ₹6,600/week — under ₹3.5 lakh a year of bottleneck time buying 35 extra jobs a week, which at even ₹1,500 per job of contribution is ₹52,500 a week and over ₹25 lakh a year of throughput. Now price the two alternatives the folklore suggests. "Run Saturday across the shop": five stations × 8 hours = 40 hours of overtime to feed station 3 exactly the same 11 extra hours of real output — 29 of the 40 hours convert to WIP, not revenue, and the shop gets slower by the same §2 arithmetic (u rises at four non-bottlenecks, queues grow, lead times stretch). "Buy a second machine for station 2" (the 160-capacity station — the one that looks almost-constrained): zero throughput gain, ₹18 lakh spent, and a new queue of half-finished work looking for somewhere to sit. Every hour the whole plant spends around the bottleneck must be judged by what it does at the bottleneck — an hour lost there is an hour lost by the entire system, an hour saved anywhere else is a rounding error dressed as heroism.

7.3 Drum-buffer-ropes and protective capacity

The scheduling system that implements this is drum-buffer-rope: the drum is the bottleneck's schedule (it sets the beat, §4's takt applied to the constraint); the buffer is time protection before the bottleneck — work released early enough that the drum never idles, typically 24–72 hours of queued work sized by the upstream variability (§2, again: the buffer absorbs c_a^2 so the constraint sees smooth flow); the rope ties raw-material release to the drum's consumption — release a new job only as the bottleneck finishes one. Note how the rope converts §1's WIP cap into a schedule: bottleneck throughput is 3.22 jobs/hour, so release at 3.22 jobs/hour and the shop's WIP settles wherever the lead time policy allows, instead of swelling to fill every slack hour.

And the counterintuitive piece — protective capacity: deliberately underloading the non-bottlenecks (80–85 % at §2's reliability line, not 95 %+) is not inefficiency; it is the capacity buffer that keeps upstream hiccups from ever reaching the drum. A non-bottleneck at 70 % "utilisation" with a short queue behind it and a never-idle bottleneck is a better system than one at 95 % with the drum starving weekly. The management insight worth framing: buffers are insurance, insurance costs money, and the only question is which policy buys the most uptime per rupee. For most shops the cheapest insurance is protective capacity at non-constraints and a time buffer at the constraint — both cost accounting air, not concrete.

8. Value Streams and the Lead-Time Equation

8.1 Process cycle efficiency — the ratio that embarrasses

Map the flow of one product family from order to despatch — material steps and information steps — and you have a value-stream map: for each operation, its cycle time, setup, uptime, batch, inventory before it, and the queue it feeds. Sum the value-adding touch time; compare with the total lead time through the stream. The ratio is process cycle efficiency (PCE, also called flow efficiency):

A typical job-shop value stream: ~12 operations, total touch time 186 minutes across turning, milling, deburr, heat treat, grind, plate, inspect; lead time 3.8 working days, i.e. 3.8 \times 450 = 1{,}710 minutes. Then

Nearly nine-tenths of the product's life is waiting — and crucially, the waiting is not evenly spread: inventory piles and queues concentrate in three or four places (typically before the bottleneck, before outsourced operations, and in the pre-inspection corner). The same reduction logic as every other section applies, but now you can price it: halve the three biggest queues — batch sizing (§5) plus a WIP cap (§1) plus supplier discipline (§6) — and lead time falls to 1.6 days, PCE = 186/720 = \mathbf{25.8\%}, with zero change to the touch operations. The delivery quote that was 4 weeks is now credibly below 2, and the shop did it by reducing work in progress, not by buying a single thing.

8.2 The lead-time equation, assembled

Assemble the guide's building blocks into the one formula shops actually need — average cycle time for a job in a shop:

Every term has a section number attached. Touch time (§5.5, machine guides) is the smallest term in most shops — 10.9 % of the total in our worked stream. Queueing (§2) is the biggest, and it responds to load discipline, variance reduction and WIP caps. Batch drag (§5) is arithmetic once batch sizes and rates are known. And the tails are real: because the distribution of cycle times in a variable shop has a long right tail, quoting the average lead time guarantees that half the jobs are late — the honest quoting policy uses the 90th-95th percentile of the shop's own history and adjusts it as WIP and utilisation move. A shop that cannot state its P95 lead time from data is quoting from imagination, and losing orders it doesn't know it lost.

8.3 CONWIP — the one-page flow control

The simplest working implementation of all of the above is CONWIP (constant work in process): define a card count N; a job enters the shop only when a finished job leaves. Throughput is set by the bottleneck; WIP is pinned at N; and by Little's Law the cycle time is CT = N/TH — deterministic, explainable, and directly controllable. Worked: shop bottleneck 3 jobs/hour. With N = 60 cards, CT = 60/3 = \mathbf{20\ h}; pull the count down to N = 36 and CT = \mathbf{12\ h} at the same throughput, as long as N \ge W_0 (§1.2: the critical WIP). If the shop starts starving the bottleneck, N was below W_0 — raise it slightly. That is the entire tuning procedure: find the smallest card count that keeps the bottleneck busy; everything above it was lead time you were donating for free. CONWIP beats a fully decomposed kanban system for job shops because routings vary: one pool of cards governs the whole shop without per-part card arithmetic — the reason it spread from MTS automotive into high-mix environments.

9. Pull Systems and the Measurement Layer

9.1 Kanban sizing — the supermarket arithmetic

A kanban (card, signal) is a pull mechanism: consumption at a downstream station triggers replenishment at the upstream one. The number of containers/cards required for an item with usage D per period, replenishment lead time L periods, a policy safety factor \alpha, and container size C is

Worked: usage 120 pieces/day, replenishment lead time 5 days, safety factor 20 % (the policy knob standing in for §6's statistical buffer), container 40 pieces:

Eighteen containers circulating is the entire policy — visible, self-correcting, and impossible to bullwhip: when demand doubles, the cards empty faster and pull replenishment faster, with no forecast in the loop. That last property is the point. Push systems (MRP) plan against forecasts, and forecast error compounds as it travels upstream — the bullwhip effect: end-demand variance of ±10 % becomes order variance of ±30 % at tier 2 and ±50 % at tier 3, each tier buffering against the tier before it. Pull systems replace forecast-driven batch release with consumption-driven replenishment, and their bill comes due only in the prerequisites: *pull needs small, frequent, reliable replenishment, which needs SMED (short, standard setups — §5.3), stable quality (no rework loops in the pipeline — §6), and level demand (heijunka — §4.4).* Implementing kanban before setup reduction is the classic sequence error — the cards simply get bigger and hide the failure.

9.2 The measurement layer — five KPIs a small shop can keep

Every mechanism in this guide runs on data, and the data stack for an Indian SME shop does not need to start with a ₹50-lakh MES. The ladder: shift logbooks (paper or tablet — captures losses only if the six big losses are pre-printed as checkboxes); barcode scans at operation start/finish (gives true queue times, not estimates); machine signals (hour meters, spindle load via OPC-UA/MTConnect, or even a current clamp) feeding automatic runtime capture; and only then a full MES/KPI platform aligned to ISO 22400-2 definitions [20]. Whatever the tooling, five numbers cover 90 % of the decisions:

  1. OEE at the constraint (§3) — and OEE at the next two worst machines, monthly;
  2. On-time in full (OTIF) — orders delivered complete, on the promised date, as a percentage — the number the market actually scores you on;
  3. WIP turns / WIP value — ₹ on the floor, and turns per year (§1): the numerator of Little's Law;
  4. Lead time, P50 and P95 — quoted from data, tracked against promise (§8.2);
  5. First-pass yield at each critical operation, with the Cpk of the two dimensions that generate the most customer complaints (§6.2).

A shop that maintains those five for two quarters can price differently: its quotes stop embedding invisible losses, its lead-time promises survive contact with reality, and its improvement projects choose themselves from the largest gap. The platform layer — and this is where job-tracking marketplaces like FabFlow sit in the stack — adds the demand side of the same data: every job accepted, quoted, delivered and rated through the system is a measurement the shop no longer has to reconstruct from memory. Shops that run the arithmetic in this guide and feed it real numbers outbid their neighbours on the same jobs with the same machines, because they are the only ones who know what the job actually costs and when it can actually ship.

10. The Ledger — Rupees per Hour, per Piece, and per Rupee of WIP

10.1 The machine-hour rate, built from parts

Everything in this guide gets its final test in one number: what does an effective machine-hour actually cost, and what price covers it? Build the rate for a mid-range Indian VMC (₹18 lakh installed, 5-year life, two shifts, ~4,000 effective spindle-hours/year), the way a costing engineer would [18]:

Cost element · ₹/hour · Basis

Depreciation · 180 · ₹18,00,000 ÷ (5 yr × 4,000 h) ≈ ₹90/h at full two-shift use; paced down for realistic ~2,000 h/yr yields ₹180

Power · 64 · ~8 kW average draw × ₹8/kWh

Operator (fully loaded) · 152 · ₹26,840/month (₹22k wage + ~22 % statutory) ÷ 176 h

Tooling & consumables · 120 · inserts, coolant, wear parts

Space, compressed air, overheads · 60 · floor area share at §rent ledger

Maintenance & AMC · 40 · annual contract ÷ hours

Machine cost · ≈₹616/h · before profit, before losses

Now apply the honesty layer from §3: at OEE 78.4 %, every effective (sold) hour consumes 1/0.784 = 1.28 machine-hours of cost — the true rate to recover is ₹786 per effective hour (616 \times 1.275), before profit. A shop pricing at its "machine rate" of ₹600 and running at 78 % OEE is selling below cost by ₹186 on every hour it sells; a shop at 85 % and the same cost base is selling with margin. This single multiplication — costing at effective hours, not at nameplate hours — is the most direct financial consequence of the entire OEE discipline. And it lands in a market that publishes its cruelty: 2026 market bands for fully-loaded rates run ₹500–900/h for 3-axis VMCs, ₹450–800 for turning centres, ₹600–1,200 for wire-cut EDM, and ₹1,200–2,500 for 5-axis precision work, with job-work billing commonly at ₹2.50–8.00 per minute — while small two-machine single-shift shops need ₹4.20–5.80 per billable minute just to cover fully loaded cost [16][18]. Below ₹4.50/minute is below survival for the smallest shops, which is why the Coimbatore squeeze — ₹150 against a needed ₹350 for CNC hours; ₹250 against ₹550 for VMC hours [17] — is not merely a margin story but a measurement story: the shops being squeezed hardest are, with few exceptions, the ones with no defensible cost model, no OEE data, and no lead-time credibility. You cannot defend a price you cannot explain, and you cannot explain a price built on folklore.

10.2 What one rupee of WIP costs per year

Inventory carrying cost — the annual cost of holding goods — is benchmarked at 20–25 % of inventory value per year for healthy operations, 15–30 % across industries, composed of capital cost (6–15 %), storage (2–5 %), service/insurance (1–4 %), and risk: obsolescence, shrinkage, damage (4–10 %) [21][22][23]. Work-in-process at a machine shop carries the full stack and then some — WIP is exposed to damage on pallets, engineering changes mid-flight, and the occasional rust weekend.

The shop from §1-§2, holding ₹80 lakh of WIP, pays

to do nothing — a ₹17.6 lakh line item that appears on no P&L, because it is smeared across rent, interest, handling wages and write-offs. Apply the guide's levers (SMED §5, WIP cap §8, batch discipline §5): WIP falls from ₹80 lakh to ₹40 lakh, and ₹8.8 lakh a year returns to the shop — alongside the lead time halving that wins the orders in the first place. This is the arithmetic that makes the introduction's Shop A versus Shop B real: Shop B's "full utilisation" is financing ₹8.8 lakh/year of nothing and selling it as commitment.

10.3 The improvement programme, priced end to end

Stack the projects of this guide into one conservative programme for a 6-machine shop, and count only recovered capacity and cash, not new sales:

Project · Spend · Annual return · Mechanism

SMED on 2 changeover-heavy machines (90 → 30 min; 4 setups/day total…) · ₹1.2 lakh (fixture plates, preset cart, quick clamps, training) · ~₹6.0 lakh capacity (4 h/day × 250 days = 1,000 h × ₹600) · §5.3

OEE programme: 78.4 → 85 % across 6 machines · ₹1.8 lakh (data capture, TPM circles, spares discipline) · ~2,000 effective hours recovered ≈ ₹12 lakh · §3.4

WIP cap + batch discipline (CONWIP, no capital) · ₹0.5 lakh (training, cards, re-sequencing) · ₹8.8 lakh carrying + lead-time wins · §1, §8

Supplier & quality discipline (σ_L, Cpk re-centring) · ₹0.7 lakh · ~₹3 lakh (defect + buffer reduction) · §6

Even after netting the overlaps (changeover minutes appear in both SMED and OEE) the programme clears well over ₹20 lakh a year against under ₹4.5 lakh of spend — and none of it requires a new machine, a new building, or a new phone number in Coimbatore. The returns are not bonuses paid to heroics; they were always there, distributed invisibly across queue time, carrying cost and unmeasured loss. Production systems engineering is the act of collecting them deliberately.

10.4 The quoting consequence

Put the ledger on a real quote. Part: 22-minute effective VMC cycle, ₹640 of material, batch of 30, one-off job. At the §10.1 honest rate of ₹786/effective hour, with setup amortised over the batch (30 min / 30 = 1 min/piece) and an 18 % pre-tax contribution target:

A desk-quoted "market rate" of ₹950 wins the job and does not cover the work — and it will be won by somebody, which is precisely why the disciplined shop must know the ₹1,148 is real and defend it with delivery credibility, or redesign the job (cycles, batch, process — the machine guides) until the arithmetic accepts ₹950. Quoting is where every section of this guide files its report: touch time from the machining knowledge, load and queue from §1-§2, losses from §3, batch from §5, risk from §6, and the target contribution from the shop's own ledger. A quote built this way can be explained, revised and defended line by line. A quote built on folklore can only be lowered.

Pitfalls — Twenty Ways a Shop Optimises Itself to Death

  1. Utilisation targets above 90 % on lead-time-critical resources. §2's hockey stick: the last ten points of load cost 2–3× the queue, and the shop becomes a WIP warehouse with a machining problem.
  2. Improving OEE at the non-bottleneck. More efficiency away from the constraint converts capacity into inventory (§7); the money is only at the drum.
  3. Batch sizes set by habit — or by ERP defaults nobody has reviewed. If S has never been measured (§5.1), the batch size is a rumour, not a decision.
  4. Setup costed at zero. The accounting that says changeovers are "free" (machine already depreciated!) produces infinite batches, infinite inventory and a shop that cannot re-quote quickly.
  5. Managing by averages. Average OEE, average lead time, average setup — each hides the variance that §2 taxonomises; quote P95, report spreads.
  6. Expediting as a policy. Moving a job to the front moves every other job back by the same total (§1); the average, and the customer, remember.
  7. Letting MRP schedule a high-mix shop. Forecast-driven release against fixed planned lead times manufactures both the bullwhip and the queue (§9.1).
  8. Kanban before SMED. Pull systems need short, standard replenishment; installing cards over 90-minute changeovers just relabels the batches (§9.1).
  9. Safety stock from the demand-only formula. When \sigma_L rules (§6.1: 87.7 % of the variance in the worked example), z\sigma_d\sqrt{L} under-buffers by two-thirds.
  10. Chasing Cp when the problem is Cpk. A perfectly spread process off-centre grinds ₹ away (§6.2); re-centre before re-tooling.
  11. Inspection as the quality plan. Sorting detects defects at ₹10–100 per the 1-10-100 rule; prevention costs ₹1 (§6.3). Rolled yield at 95 % × 5 stations — 77 % — is not opinable arithmetic.
  12. "Idle machines are waste." At non-constraints, idle is the protective capacity that protects the drum (§7.3); a starved bottleneck is the only true idleness in the building.
  13. Never filming a changeover. The 30–50 % of stopped time that is postponed external work (§5.3) is invisible on any walk-by; it takes a video and a stopwatch.
  14. *Ignoring setup and cycle variance.* A standard 40-minute setup beats an average 35-minute one when c_e^2 feeds every queue behind it (§2.3, §5.3).
  15. Big batches to "save capacity". The ledger (§10.2) shows where the savings actually land: in WIP carrying, lead time, and the quote you lose because you deliver in five weeks.
  16. Capacity plans on nameplate hours. Net capacity = hours × OEE — 1,600 effective hours from 2,000 nameplate at 80 % OEE; plans built on nameplate run permanent overtime (§3).
  17. One-number management. OEE, utilisation, or scrap percentage alone — each can be improved by wrecking another (speed for quality, availability for inventory); manage the set (§9.2).
  18. No WIP cap. Without a ceiling, release pressure pushes lead time up forever, and the drift is invisible because every individual week looks normal (§1.2, §8.3).
  19. Quoting average lead times and average costs. Half the jobs late, and the margin spent before the invoice (§8.2, §10.4).
  20. Dashboards nobody acts on. Measurement without the measure→decide→change loop is decoration; the waterfall of §3.2 exists to nominate the next project, not to be admired.

Frequently Asked Questions

How do I calculate OEE for a machine shop?

\text{OEE} = A \times P \times Q against planned production time (§3). Worked example: an 8-hour shift with 30 min of planned breaks → 450 planned minutes; 38 min breakdown + 22 min changeover → A = 390/450 = 86.7\%; 309 pieces at a 1.2 min ideal cycle against 390 min run → P = 95.1\%; 294 good of 309 → Q = 95.1\%; OEE = 78.4 %. As a waterfall: 450 - 60 - 19.2 - 18 = 352.8 productive minutes. Benchmarks: 85 % world class (built from 90/95/99.9), ~60 % the industry median, 40–50 % common in unmeasured small shops.

Why does high machine utilisation increase lead time?

Because of the queueing term \frac{u}{1-u} in the Kingman/VUT equation: CT_q \approx \frac{c_a^2+c_e^2}{2} \cdot \frac{u}{1-u} \cdot t_e (§2). At u = 80\% the factor is 4; at 90 % it is 9 — so a CNC cell with a 40-minute effective process time carries 160 minutes of queue at 80 % load and 360 minutes at 90 %, with the same machines and the same jobs. Plan lead-time-critical resources at 75–85 % of effective capacity, and use the spare 15–25 % as the deliberate capacity buffer.

How do I calculate takt time, and what do I do with it?

t_{takt} = available production time ÷ customer demand. For 600 pieces/day on one 450-minute shift: 27{,}000 \div 600 = \mathbf{45\ s/piece} (§4.1). Then convert it into a capacity verdict through OEE: a machine with a 40 s ideal cycle at 78.4 % OEE runs effectively at 40/0.784 = 51 s — 13 % below takt — so the shop needs OEE ≥ 88.9 % (cheapest), more machines, or more hours, in that order [18].

What batch size should a machine shop run?

The economic batch is Q^ = \sqrt{2DS/H} — for D = 12{,}000/yr, S = ₹2,400 setup, H = ₹120/unit-yr, Q^ \approx 693 pieces and ₹83.1k/yr of setup-plus-holding (§5.2). The lever that matters is S: cutting the changeover from 90 to 20 minutes (SMED) quarters S, halves the economic batch to 346, halves annual cost to ₹41.6k, and halves lead time for that part** — the square-root law. Measure S with a stopwatch before believing any batch size you're running.

How much safety stock do I need if my supplier is unreliable?

Use the two-term formula: SS = z\sqrt{L\sigma_d^2 + d^2\sigma_L^2} (§6.1). Worked: d = 120/day, \sigma_d = 30, L = 9 days, \sigma_L = 2 days, 95 % service (z = 1.645): SS = 1.645 \times \sqrt{9(900) + 14{,}400(4)} \approx \mathbf{422} pieces, reorder point 1,502. Note 87.7 % of the required buffer came from supplier lead-time variance — cutting \sigma_L from 2 days to 1 shrinks the buffer to 247 pieces, worth more than any demand forecast improvement.

What is a good Cpk value?

1.33 is the modern floor for critical dimensions (≈63 ppm two-sided), 1.67 for safety-critical, ~2.0 (the "six sigma" branding, 3.4 ppm with the classic 1.5σ drift) for world-class criticals (§6.2). A worked case: ±0.05 mm grind at μ = 25.012, σ = 0.012 gives C_p = 1.39 but C_{pk} = \mathbf{1.06} — spread fine, centring poor, ~760 ppm exposed. Re-centring to target lifts Cpk to 1.39 and cuts exposure to ~30 ppm — with no new tooling.

What is SMED and how much setup reduction is realistic?

SMED — Single-Minute Exchange of Die — is Shingo's method for cutting changeovers into single-digit minutes by separating internal work (machine stopped) from external work (possible while running), converting and streamlining each (§5.3). Typical results: 50–90 % reduction, and 30–50 % of stopped time is usually just postponed external work — free to recover. A 90-minute CNC changeover cut to 25 minutes captures most of the economics (the batch-size benefit is a square root); single-digit targets come after the first two passes [10][11][12].

Which KPIs should a small machine shop track?

Five (§9.2): OEE at the constraint (improving), OTIF (on-time in full), WIP value and turns, lead time at P50 and P95 (quoted from data), and first-pass yield with Cpk on the two most-complained-about dimensions. Five numbers, reviewed weekly, outperform fifty numbers nobody reads — and they are the raw material of every quote the shop writes.

The Discipline in One Page

Production systems engineering is one idea worked in a dozen arithmetics: a factory is a network of queues, and the queues obey laws whether or not anyone has read them. Little's Law fixes lead time at WIP/TH — 45 jobs at 9 completions a day is five days, always (§1); the Kingman/VUT equation prices the load dial — the same cell carries 160 minutes of queue at 80 % utilisation and 360 at 90 %, so the 15 % capacity buffer is bought with the cheapest currency a shop has (§2); OEE decomposes a shift into its own waterfall — 450 minutes in, 352.8 out, 78.4 %, against a world-class 85 % and a median 60 %, with ten points worth a machine and a half of capacity on a twenty-machine floor (§3); takt turns demand into 45 seconds and, chained through OEE, turns seconds into a hiring decision (§4); the square-root law of batch size says quarter the setup, halve the batch, halve the cost — which is why Shingo's single-digit minutes are the foundation stone of everything downstream (§5); variability gets buffered deliberately — safety stock of 422 pieces when lead-time variance rules, Cpk 1.06 re-centred to 1.39, scrap and rework priced at last year's 2.2 %-of-revenue benchmarks (§6); the Theory of Constraints says eleven hours of bottleneck overtime beat forty hours of everyone-works-Saturday, every week of the year (§7); value-stream arithmetic says 10.9 % of a part's life is value-adding and the rest is a policy choice (§8); pull systems and five KPIs close the loop from cards to data (§9); and the ledger closes the argument — a VMC hour that costs ₹616 nameplate and ₹786 effective, job work crushed to ₹150 an hour in Coimbatore, ₹17.6 lakh a year evaporating out of ₹80 lakh of WIP, and an improvement programme returning over ₹20 lakh a year for ₹4.5 lakh of spend with zero new machines (§10).

That is the standard the FabFlow network exists to keep — machine shops, fabricators, and the buyers who need parts on a date, not a story. The shops that win the repeat orders are not the ones with the newest spindles; they are the ones who did the arithmetic in this guide, put numbers on their floor, and then let the numbers argue for them. Every job on the platform moves because somebody measured instead of guessed.

Previous guides in this series: Belt Conveyors & Bulk Material Handling · Cranes, Hoists & Rigging · Cleanroom Engineering & Contamination Control · Industrial Refrigeration & Cold Chain Systems · Industrial Electrical Power Distribution · Metrology & Dimensional Inspection · Industrial Furnaces, Kilns & Refractories · Structural Steel Design & Fabrication · Pressure Vessels & Storage Tanks · Process Piping & Pipe Fabrication · Industrial Steam Boilers & Steam Systems · Non-Destructive Testing.

[1] Press Information Bureau, Government of India — MSME sector releases: sector contributes ≈31.1 % of GDP, 35.4 % of manufacturing output, 48.58 % of exports (data as of January 2026); Udyam + Udyam Assist registrations crossed 8.7 crore as of June 2026; manufacturing-category registrations ≈1.6 crore; MSME Day 2025 materials (five crore-plus registrations, PM Vishwakarma). [2] SIDBI, "Understanding Indian MSME Sector" (May 2025) — 7.34 crore estimated enterprises (ASUSE 2023-24), ~6.2 crore Udyam registrations by March 2025, MSME export share 45.79 % (FY2024-25), MSME share in GVA 30.1 % (FY2022-23). [3] Seiichi Nakajima / Japan Institute of Plant Maintenance, Total Productive Maintenance framework — OEE = A × P × Q, the six big losses (breakdowns, setup/adjustment, minor stops, speed, startup defects, production defects); world-class construction 90/95/99.9 ≈ 85 %. [4] MESA / industry benchmark compilations — median OEE ≈60 % across discrete manufacturing; 60→85 % gap ≈41 % more output at the same assets. [5] Industry OEE benchmark summaries (2025–2026) — world-class 85 %+, component targets 90/95/99.9, gap framing. [6] Tractian, "What is World Class OEE" — component targets and sector OEE ranges (discrete 65–75 %, automotive 70–80 %, food 55–70 %, pharma 40–60 %, mining 50–70 %, packaging 60–75 %). [7] Wallace J. Hopp & Mark L. Spearman, "Factory Physics" (3rd ed.) — Little's Law applications, critical WIP (W_0 = r_b T_0), VUT/Kingman queueing model, law of variability buffering (inventory/capacity/time), variability propagation. [8] Western Michigan University, CS 6570 lecture notes on queueing (Kingman's equation / VUT) — CT_q \approx \frac{c_a^2+c_e^2}{2}\cdot\frac{u}{1-u}\cdot t_e; variability definitions (CV, SCV); "variability causes congestion"; propagation. [9] Factory Physics Inc. / Intel IIE presentation (2008) — static capacity modelling with the VUT equation in semiconductor fabs; "↓ variability = ↑ consistent output = ↓ cycle time". [10] Shigeo Shingo, "A Revolution in Manufacturing: The SMED System" — internal vs external setup (IED/OED); single-digit minute target; four-hour to three-minute press changeover; and SMED method guides (Symestic, 2026) — 30–50 % of stopped time is postponed external work; stage-wise reductions; OTED <100 s. [11] Symestic, "SMED (Single-Minute Exchange of Die)" — definition (<10 minutes, not <60 seconds), Toyota/Mazda development 1950s–1969, 90-minute → 9-minute economics example. [12] Netray SMED glossary — practical targets (90 → 25 min still delivers most benefit), setup time as dominant driver of economic lot size, ~30 % lot-size reduction per halving of setup (\sqrt{} law). [13] Society of Manufacturing Engineers, "Manufacturing Insights: Costs of Poor Quality" — quality cost categories (prevention/appraisal/failure), customer economics "1-5-12" ratio. [14] APQC via Ease.io — scrap and rework at up to 2.2 % of annual revenue (average), 0.6 % (top performers). [15] Industry COPQ compilations — total cost of poor quality estimated at 5–30 % of revenue. [16] CISUITE / India CNC cost guides (2026) — machine hourly rate bands: 3-axis VMC ₹500–900, turning ₹450–800, 5-axis ₹1,200–2,500, wire-cut EDM ₹600–1,200. [17] The Times of India, "Lower job order rates threaten small units' survival in Coimbatore" — CNC job-work rates paid ₹150/h against ₹350/h; VMC ₹250/h against ₹550/h. [18] KarobarUdhar, "CNC Machining Business Feasibility in India 2026" and CompetenceTechno machine-hour-rate guide — job-work billing ₹2.50–8.00/min; sub-₹4.50/min below fully loaded cost for small shops; fully loaded cost ₹4.20–5.80/min on two mid-range VMCs; operator wages ₹12–40k/month + 20–25 % statutory; industrial VMC ₹12–80 lakh. [19] Eliyahu M. Goldratt & Jeff Cox, "The Goal" (1984) and the Theory of Constraints body of work — five focusing steps, bottleneck law, drum-buffer-rope (developed in "The Race" / Goldratt's later works). [20] ISO 22400-2:2014 — standardized definitions of manufacturing operations KPIs including OEE. [21] WarehousingCosts.com, "Inventory Carrying Cost" (2026) — all-in carrying 15–30 %/yr; healthy 20–25 %; component table (capital 6–15 %, storage 2–5 %, service 1–4 %, risk 4–10 %); industrial/MRO 18–25 %. [22] NetSuite inventory carrying cost guide — 20–30 % typical range. [23] SupplyChainMath carrying-cost analysis — 20–30 % benchmark; worked component splits. [24] J. D. C. Little (1961), "A Proof for the Queuing Formula: L = λW", Operations Research 9(3) — the theorem behind WIP = TH \times CT. [25] Mike Rother & John Shook, "Learning to See" (LEI) — value-stream mapping practice, lead-time ladder, process cycle efficiency measurement.

Note on sources: benchmark figures (OEE medians, carrying-cost rates, COPQ percentages, market price bands) are drawn from the published surveys and guides cited above and are presented as ranges for engineering context, not as guarantees for any individual shop; every worked example is arithmetic performed on the stated inputs and should be re-run on a shop's own data before decisions. Standards references (ISO 22400-2, and the ISO/IS documents cited in earlier guides of this series) are summarised from publicly available material; where money is at stake, verify against the current edition of the governing standard and your own ledger.

More FabFlow blog posts