# Engineer-to-order and series in one building

> A plant that builds special machines and also runs series production is running two different mathematics in the same ERP. One is a project with a critical path; the other is a flow with a queue. Planning both with one method is why the machine ships late and the series order ships short, and the arithmetic says which lever to pull.

Page: https://ganiindustrial.com/insights/two-planning-regimes-in-one-building · 2026-10-14 · Language: en

Our own plant has the problem this note is about, which is the only reason we are qualified to write about it. One part of the building designs and builds machines that have never existed before. Another part produces control panels and harnesses in repeating quantities to a frozen design. The two share a building, a quality system, a purchasing department and, most awkwardly, several of the same people. They do not share a planning method, and the attempts we have watched to make them share one have all failed in the same way: the project absorbs the shared resources whenever a milestone approaches, the series orders wait, and then the series customer escalates and the project slips instead.

The underlying reason is mathematical rather than managerial. Series production is a flow problem, governed by queueing behaviour and lot sizing. An engineer-to-order machine is a project, governed by a network of dependent activities with no repetition to learn from. The two have different objective functions, different units of measure and different failure modes, and the formulas that describe each one say plainly that a policy which is right for one is wrong for the other.

## What classic MRP assumes, and what it does not

Material requirements planning does one thing extremely well. It takes a demand for a finished item, explodes it through the bill of materials level by level, nets each requirement against stock on hand and orders already placed, and offsets the result backwards by the planned lead time to say when each order must be released. For a repeating product with a stable structure it is close to unimprovable, and spreadsheets that replace it are almost always worse.

What it does not do is consider capacity. Classic MRP plans material against infinite capacity and leaves the question of whether the work centres can actually do the work to a separate capacity requirements planning step afterwards, which reports an overload rather than resolving it. That division is defensible in a flow shop where the routings are similar and the load averages out. It is indefensible when one of the orders in the plan is a machine build whose activities are unique, whose precedence network is deep and whose engineering content changes while the plan is running.

## The series side: three formulas worth knowing

Lot sizing is where most of the inventory in a plant is decided. The textbook starting point is the economic order quantity, which balances the cost of ordering against the cost of holding. For an item with annual demand of 12,000 pieces, an order or setup cost of 120 euros, a unit cost of 8 euros and a holding charge of 20 per cent a year, the holding cost is 1.60 euros per piece-year and the quantity is the square root of two times demand times order cost divided by holding cost: the square root of 1,800,000, about 1,340 pieces, or nine orders a year.

The number is useful and its assumptions are worth stating, because they are routinely violated: constant demand, instant replenishment, no quantity discounts, no capacity limit. In the lumpy demand a machine builder actually sees, the honest tools are the dynamic ones — the Silver-Meal heuristic, which adds periods to a lot while the cost per period keeps falling, and the Wagner-Whitin algorithm, which solves the deterministic single-item case optimally by dynamic programming. Both need a demand schedule rather than an annual total, which is precisely the information an MRP run already has.

Safety stock is the second formula, and the version most plants use is the wrong one. With a reliable lead time and variable demand, the cover needed is the service factor multiplied by the standard deviation of demand multiplied by the square root of the lead time. At a 95 per cent service level the factor is 1.645; with a weekly demand standard deviation of 60 pieces and a lead time of 4 weeks, that is 1.645 times 60 times 2, about 197 pieces.

Now allow the lead time itself to vary, which is what actually happens with a supplier. The cover becomes the service factor multiplied by the square root of the lead time times the demand variance plus the mean demand squared times the lead-time variance. With a mean weekly demand of 500 pieces and a lead-time standard deviation of one week, the term inside the root is 14,400 plus 250,000, so the safety stock is about 846 pieces — more than four times the figure that demand variability alone produced. Read that twice before the next forecasting project is approved. Stabilising a supplier lead time is worth several times what improving the forecast is worth, because the mean demand enters the lead-time term squared while the demand variability does not.

The third is the simplest and the most ignored. Little law states that work in progress equals throughput multiplied by flow time. A cell producing 48 good parts an hour with a six-hour flow time holds 288 parts somewhere between its first and last operation. The relation is an identity, not a model, so it cuts both ways: at constant throughput, halving the work in progress halves the flow time, and the only way to shorten a lead time without adding capacity is to put less into the system. Every plant that responds to a late order by releasing more work has the identity working against it.

## Why the shared work centre is the whole problem

The formula that explains most delivery failures in a mixed plant is the queueing approximation usually attributed to Kingman. Waiting time at a work centre is roughly the average of the arrival and service variability terms, multiplied by the utilisation term, multiplied by the effective process time. The utilisation term is the utilisation divided by one minus the utilisation, and it is the reason intuition fails.

At 80 per cent loading that term is 4. At 90 per cent it is 9. At 95 per cent it is 19. Pushing a shared resource from 80 to 95 per cent loading does not make it 19 per cent busier in any way a customer notices; it multiplies the queue in front of it by nearly five. The variability term matters just as much: a work centre fed by a mixture of 200-piece series batches and one-off machine parts has high arrival variability by construction, so its queue is long at utilisations where a dedicated cell would be comfortable.

That single relation is the strongest argument for not treating a machine shop, a wiring bench or an engineering team as a pool to be filled. It also explains why plants that measure utilisation deliver badly: a manager rewarded for keeping a work centre at 95 per cent is being rewarded for building a queue.

## The project side obeys different rules

A special machine is not a lot. It is a network of activities with precedence constraints and shared, limited resources, which in the general case is a resource-constrained project scheduling problem — provably hard to solve optimally, and more importantly, not something an MRP run is attempting at all. Its critical path is a chain of engineering decisions, long-lead components and integration steps, and the thing that determines the delivery date is usually not capacity but a decision somebody has not made yet.

Three consequences follow. First, buffers belong at the end of the chain rather than on every task, which is the useful core of critical-chain thinking: padding each activity individually guarantees that every activity consumes its padding, while a single project buffer absorbs the variation that actually occurs. Second, the long-lead items have to be ordered against an engineering release that is deliberately staged, so that a drive, a robot or a bought-in index table can be committed while the guarding is still being drawn. Third, engineering changes on a live project are not disturbances to be minimised away; they are the normal condition of batch size one, and the plan has to have somewhere for them to go.

## Nervousness, and the fence that stops it

An MRP regenerated every week on a rolling horizon will reschedule orders that did not need rescheduling. A small change in a forecast several levels up, a lot size that rounds differently, a single late receipt, and the run produces a fresh set of release dates for parts whose real situation has not changed at all. The production-planning literature named this system nervousness in the 1970s and the cure has not changed since: a planning time fence inside which the system is not allowed to replan, planned orders firmed by a planner rather than by the algorithm, and lot sizes that are not recalculated every cycle.

In a mixed plant the nervousness has a second source, which is the project. Every engineering change on a machine build enters the material plan as a new or cancelled requirement, and without a fence those changes churn the release dates of the series work that shares the same components and the same people. The practical remedy is unglamorous: plan the project outside the MRP horizon and let it consume reserved capacity, rather than letting it compete inside the same netting run.

## How we actually split it

What works for us is a deliberate separation with exactly one shared object. The series side is planned by MRP with a frozen window of two weeks, firm planned orders beyond it, lot sizes reviewed quarterly rather than recalculated weekly, and routing times taken from measurement instead of from standards. The project side is planned as a project, with a work breakdown structure, a critical chain and a single project buffer, and its long-lead material released against staged engineering approvals.

The one thing planned in both places is people. A named capacity reservation — so many engineering hours and so many wiring hours per week committed to the project, visible to the series planner as unavailable — is the only mechanism we have found that stops the two regimes from silently robbing each other. It is an uncomfortable number to set, because it forces somebody to decide in advance how much of the week belongs to the machine and how much to the panels, and that decision is exactly the one that otherwise gets made every morning by whoever shouts loudest.

## What to measure

Utilisation is the wrong headline measure for a mixed plant, for the reason the queueing term gives. The pair we watch instead is schedule adherence on the series side — the percentage of orders completed in the period they were planned for, counted without credit for early — and buffer consumption against progress on the project side, which tells you whether a machine is in trouble long before a milestone is missed. Both are harder to make look good than utilisation, which is the point of them.

Underneath both sits the same requirement as always: routing times that were measured, bills of material that match what is built, and a per-part record from the floor that turns next quarter planning assumptions into this quarter measurements. Neither regime can be planned well on data that was entered once when the system went live and never checked against a shift report since.


---

GANI Industrial — engineering, manufacturing and lifecycle for European OEMs. Bursa, Türkiye.
Site index: https://ganiindustrial.com/llms.txt