The constraint that stops an automotive line is usually several tiers below the part number you buy, and a tier one scorecard that is green on delivery, quality and financial health cannot see it. Sequenced delivery is a promise about order rather than about quantity, so a supplier holding the right total volume in the wrong build sequence still stops the line. The order bank changes what is being planned, since a confirmed customer sequence replaces the forecast across part of the horizon and the two halves need different treatment. Capacity was committed when the tooling was bought, which makes the volume forecast written three years before start of production the decision that actually sets your flexibility.
Your seat supplier confirms the full week on Monday. On Wednesday afternoon they call to say that the plant making the foam had an outage a fortnight ago, they have three days of material left, and after that they can build cushions in one trim only. You have four hours of seats on the ground, because four hours is what sequenced delivery means.
Nothing on your scorecards predicted it. The seat supplier is green on delivery, green on quality, financially sound and dual sourced at your tier. The exposure sat two levels below the part number in your bill of materials, in a chemical plant you have no commercial relationship with, and your MRP has never looked past the part you buy.
Planning here has a particular shape. A deep bill of materials, a line that consumes it in sequence, and an economics where a component worth a few cents can stop a plant for a shift.
The constraint is usually several tiers down
MRP explodes the bill of materials you own and stops at the boundary of what you purchase. Below that line every bought part is a leaf node with a lead time and a supplier name, and the structure underneath it is somebody else's problem until it becomes yours.
The parts that actually stop lines live below that boundary. A resin grade, a connector housing, a magnet, a coating, a specific microcontroller. What makes them dangerous is convergence: several of your tier ones, each of them properly dual sourced from your point of view, buy the same sub-component from the same plant. When that plant goes down, four unrelated suppliers fail in the same week and the failures look uncorrelated on every report you have.
Two structural effects make it worse. Order variability amplifies as it moves up the chain, which Forrester described in 1958 and Lee, Padmanabhan and Whang formalised in 1997, so a modest schedule change at your plant arrives at tier three as a large swing. And the capacity decisions at that tier were made against those swings, years ago, with no visibility of the end programme.
The planning move available to you is to extend the where-used graph past your own boundary for a defined list of components, using source declarations your tier ones will provide when asked part by part, and to run the explosion over the extended structure. Once even a partial sub-tier graph exists, the question a shortage triggers changes from whether you have enough of part X to how many end items depend on one node. Building and maintaining that map as a risk discipline, with scoring and monitoring, is a separate exercise; the planning use of it is narrower and starts working with a fraction of the coverage.
Sequenced delivery is a promise about order
In sequenced supply, a broadcast goes out when a body reaches a defined point, the supplier builds or picks modules in build order, and they arrive at a line-side location inside a short window. The plant holds hours of inventory by design.
What that asks of a supplier is stability of mix, which is a different requirement from stability of volume. A supplier sized on average daily units will fail on the day the sequence runs long on the variant with the longest build time, even though the daily total was exactly as forecast. Their capacity model has to be built on the distribution of mix inside the sequence window, and their buffer has to be expressed in units of sequence variability rather than in days of average demand.
The complication is that the sequence the supplier receives is not the sequence anyone planned. Paint rework, quality holds and repairs pull bodies out of order between body shop and final assembly, and the frozen period is frozen in name. Every metre of resequencing tolerance taken upstream is paid for downstream as supplier buffer, and it comes back in piece price at the next negotiation.
The assessment language for this already exists. The materials management operations guideline published jointly by AIAG and Odette gives a structured way to evaluate a supplier's own planning and materials capability, and it works well as a diagnostic on the suppliers whose sequence stability you depend on. Running it on your top ten sequenced suppliers tells you more about your line-stop risk than another round of delivery scorecards.
The order bank changes what you are planning
In a build-to-order market the order bank is real demand with configurations attached. The near horizon is known, the planning problem is slot allocation and customer lead time, and the forecast only starts to matter beyond the bank.
In a build-to-stock market the forecast drives the build, field and dealer stock absorb the error, and the planning problem is mix and stock ageing. Most manufacturers run both, split by region, and pushing them through one process produces a plan that is wrong in one of them.
The diagnostic number is order bank coverage against the frozen window. If the bank covers six weeks and the frozen period is eight, then two weeks of what you are building is a forecast wearing an order number, and the mix in it deserves the scrutiny a forecast gets rather than the confidence an order gets.
Mix is where this usually breaks. Component demand is driven by options and by the combinations customers choose, so the volume can be right to within a percent while the demand for a specific harness variant is out by a third. Take rates move with pricing, with package bundling, and with what was actually available to order last quarter. That last one is the trap: an option that was constrained shows a low take rate in history, and a model fitted to that history plans the constraint forward. Correcting for it is the same censored demand problem that shows up in retail sales history, applied to configuration data.
Engineering change lands on a plan that is already running
A design change arrives with an effectivity, either a date or a serial breakpoint, and with an interchangeability status that decides how much freedom you have. One-way interchangeable means you can run the old part out first. Not interchangeable means both parts have to be planned in parallel to a hard cut, and the exposure at the cut is doubled.
The planner's job is choosing the breakpoint. That decision needs pipeline stock at the plant, in transit, at the supplier's finished goods, and raw material the supplier has already committed with their own suppliers, because the obsolescence liability in most contracts reaches further up than people expect. It also needs the new part's approval status, since a changed part goes back through production part approval before it can be shipped in volume, and an approval that slips turns a planned breakpoint into an unplanned shortage.
Mandatory changes driven by safety, homologation or regulation have a date that does not move, and the run-out cost is whatever the pipeline says it is. Economic changes have a date you choose, and choosing it is an arithmetic problem: obsolescence cost at each candidate breakpoint against the per-unit benefit of starting the new design earlier. Run that as a scenario over the actual pipeline. The version done in a meeting from memory is wrong in a predictable direction, because the people in the room can see plant stock and cannot see the supplier's raw material commitment.
Capacity was committed when the tooling was bought
Sourcing decisions are made years ahead of start of production, and the tool that comes out of them is a physical object in one location with a cycle time and a cavity count that set a hard ceiling. Capacity is a step function. Getting more of it means a second tool or a second press, which is months of lead time, a capital approval, and usually a fresh approval process because a new tool is a new source of variation.
That has an uncomfortable implication for the planning cycle. The number that sized the tooling was the volume in the nomination, produced by a forecast made three or four years before the first vehicle was built. Programmes that call off well above it create a shortage no amount of scheduling can solve. Programmes that call off well below it create an amortisation argument and a supplier who is losing money on you, which shows up later as reluctance when you need flexibility.
Almost nobody scores that forecast. The comparison between the volume assumed at nomination and what a programme actually called off in its first two years sits in your own records for every programme you have run, and across ten of them it becomes a distribution rather than an anecdote. That distribution is what should inform how much tooling flexibility you buy on the next award, and whether the extra cavity is worth its capital.
The aftermarket is a different business planned by the same team
Parts demand keeps growing after a programme ends, because the installed base is still growing, and service commitments commonly run a decade or more past end of production. The same planning group usually owns it, on the same calendar, with the same tools.
The demand behaves nothing like assembly demand. Assembly consumption is a schedule you can read. Parts consumption is failure driven, spread across tens of thousands of part numbers and hundreds of stocking locations, most of them selling in ones and twos with long gaps between orders. Supersession makes it harder: a part number replaced by a successor has to be planned as one continuous demand stream, and history split across three part numbers reads as three dying items when it is one healthy one.
At end of production the whole thing compresses into a last-time-buy, a single order covering remaining lifetime demand with no second chance. The forecasting and stocking methods that suit intermittent, long-tail demand are a separate subject (service parts planning). The consequence that belongs here is a governance one: a single team running both businesses on one cycle will always prioritise the one that stops the line today, and the aftermarket runs on parameters nobody has revisited since the programme launched.
Where this stops
Line-stop economics dominate every buffer decision in this industry, and they dominate for a defensible reason. A stop is immediate, visible, attributable to a person, and costed within the hour. The carrying cost of the inventory that would have prevented it is diffuse, accrues over months, and often lands on a balance sheet the plant manager is not measured on. Given that asymmetry, holding extra buffer is the correct decision for the plant every single time.
The analysis of where buffer should actually sit is settled work. Clark and Scarf in 1960 for serial systems, Graves and Willems in 2000 for general networks, both giving you a placement rather than a uniform rule. The output will be ignored while the accountability stays where it is, which makes this a governance problem rather than an analytical one. The parts that have to change are narrow: who owns the buffer level, who answers for a stop that happens inside an agreed buffer, and whether anyone in the room can see the network carrying cost of the local decision.
A workable arrangement sets the buffer at network level against a stated risk appetite, holds the plant harmless for stops that occur within that appetite, and reports stops caused by an agreed buffer level separately from stops caused by supplier failure. Without that separation, the first stop after a buffer reduction ends the programme regardless of what the numbers said.
One more limit worth stating. All of this rests on bill of materials accuracy and on effectivity dates being right, and in a business running hundreds of engineering changes a month, that accuracy decays continuously. A network buffer model fed by a bill of materials with stale effectivity will produce confident placements for parts that are no longer used.
Score the programme volume forecasts you made three years before start of production against what those programmes actually called off, because that error distribution is what your next tooling decision is really resting on.
More on this in industry on the DemandLabs platform.
