All postsJune 26, 2026

Heijunka for CPG Lines: Turning Lumpy Demand Into a Schedule the Floor Can Run

scheduling changeovers smed forecasting inventory planning cpg operations

You know the pattern. A big campaign of SKU A runs all week. Then a scramble of small urgent orders for C, D, and E lands, so you tear the line apart twice a shift. Overtime Thursday. Idle line Friday afternoon because the next big batch isn't released yet. Repeat next week with different SKUs in different sizes.

That sawtooth is what lean practitioners call mura — unevenness — and it breeds muri, the overburden that shows up as overtime, rushed changeovers, and tired operators making mistakes. Mura and muri together generate muda: waste. Most plants treat the sawtooth as a fact of life because "that's what the orders say." It isn't a fact of life. It's a scheduling choice.

Forecasting tells you how much to make. Heijunka — production leveling — tells you in what rhythm and sequence to make it. And rhythm is what actually protects your fill rate and your changeover budget. This piece is about what to do with the forecast once it hits the line.

What heijunka actually is

Production leveling — also called production smoothing or heijunka — is a technique for reducing the mura (unevenness) which in turn reduces muda (waste). It was central to the Toyota Production System and to lean manufacturing more broadly. The Lean Enterprise Institute defines it plainly: heijunka is leveling the type and quantity of production over a period of time, which enables production to efficiently meet customer demand while avoiding batching.

The core aim is stability. The goal is to produce intermediate goods at a constant rate so that further processing may also be carried out at a constant and predictable rate. When the upstream rate is steady, everything downstream — packaging, palletizing, warehouse staging, truck loading — can plan against a known beat instead of reacting to surges.

There are two flavors every planner has to separate in their head:

Leveling by volume — smoothing the total output (total cases, total units) across the interval so you're not running 12,000 cases Monday and 2,000 Tuesday.
Leveling by product type or mix — smoothing the sequence of SKUs so you cycle through your portfolio in a repeating pattern instead of monthly campaigns.

The two are closely related, but the second is the one CPG planners under-use. Toyota's signature move is mixed-model sequencing, not big batches: their final assembly line never assembles the same model in a batch. Instead they level production by assembling a mix of models in each batch, and they make the batches as small as possible.

Translate that to a beverage line: instead of "all of flavor A this week, all of B next week," you run a repeating short cycle — A, B, C, A, B, D — sized to actual demand proportions.

Why batching feels efficient but usually isn't

The instinct is to make to the order quantity: a big order for A shows up, so you run a big batch of A. Long runs feel efficient because they amortize the changeover over more units. That's the classic EOQ logic, and it's not wrong — given your current changeover cost.

That's the catch. Toyota's approach was to attack the changeover cost first. By reducing the time and cost of changeovers, smaller and smaller batches stopped being prohibitive. This meant demand for components could be leveled for the upstream sub-processes, and therefore lead time and total inventories were reduced along the entire value stream.

This is exactly why heijunka and SMED are joined at the hip. Leveling is only feasible if your changeovers are cheap enough to do them more often. If a flavor change costs you 90 minutes of CIP and line setup, frequent small runs are a non-starter. Drive that changeover down and the whole math flips. (We unpack the setup-reduction side in the hidden cost of changeovers.)

The philosophical shift is this: instead of making to each order's quantity, you make to long-term average demand and carry a buffer to absorb the variability. More on that buffer below — it's the trade-off nobody mentions up front.

The mechanics: building a level schedule

The concrete tool is the heijunka box — a visual scheduling tool used to achieve smoother production flow. Structurally, a typical heijunka box has horizontal rows for each product and vertical columns for identical time intervals of production. Production-control kanban are placed in the pigeon-holes in proportion to the number of items to be built of a given product type during a time interval.

Think of it as a grid:

Rows = your SKUs (or families).
Columns = your leveling intervals (a shift, a day, a two-hour pitch).
Cards in each cell = how much of that SKU to run in that slot, sized to demand share.

A worked CPG example

Say a beverage line runs five flavors with weekly demand of roughly 50% A, 20% B, 15% C, 10% D, 5% E. The campaign approach runs a giant block of A early in the week, then descending blocks.

The leveled approach divides the week into repeating cycles. If you can cycle the line several times per week, a single cycle might look like A, A, B, C, A, B, D with E folded in every other cycle — each run small, the mix matching the demand proportions, the sequence repeating. Downstream sees the same beat every day. Inventory of every SKU gets replenished frequently in small amounts instead of being feast-or-famine.

EPEI: the feasibility KPI

Whether you can actually cycle the line that often is captured by EPEI — Every Product Every Interval. EPEI answers: how often can this line cycle through its entire SKU set? An EPEI of one week means you touch every SKU once a week; an EPEI of one day means you touch every SKU daily.

EPEI is the number that tells you whether leveling is feasible on your current changeover times. The more time your changeovers eat, the longer your EPEI, and the bigger the batches you're forced into. Cutting changeover time shrinks EPEI — which is the same thing as saying it lets you level more finely. EPEI is the bridge metric between your SMED work and your leveling ambition.

The trade-off no one mentions: the leveling buffer

Here's the honest part. Leveling deliberately decouples your production rate from your order rate, and that decoupling costs inventory. When demand varies, Toyota manufactures at the long-term average demand and carries an inventory proportional to demand variability, process stability, and shipment frequency. The advantage of carrying that inventory is that it smooths production throughout the plant, reduces process inventories, and simplifies operations — which reduces total cost.

So you trade a modest, deliberate finished/intermediate buffer for plant-wide smoothness. That buffer scales with three things:

Demand variability — spikier orders need more buffer.
Process stability — flaky lines and unreliable changeovers need more buffer.
Shipment frequency — less frequent shipping needs more buffer.

This leveling buffer is not the same as safety stock. Safety stock protects against forecast error and supply disruption. The leveling buffer exists specifically to let you run a smooth, repeating schedule against lumpy orders. They sit in different parts of your plan and you should size them separately. If you blur them together, you'll either starve the line or drown in stock. (If your par levels and safety stock are still set by gut feel, start with stop firefighting stockouts.)

Preconditions and failure modes in food and beverage

Heijunka was born in automotive. Food and beverage adds constraints that can strain pure leveling, and pretending otherwise destroys operator trust. Be honest about these:

Allergen and CIP sequencing. You often can't run SKUs in arbitrary order. Allergen-containing products may need to run last before a full sanitation cycle. That constrains which leveled sequences are legal.
Sequence-dependent sanitation. Light-to-dark, low-to-high allergen, low-to-high acid — the cost of a changeover depends on the pair of products, not just the next one. A naive level schedule can accidentally maximize cleaning time.
Shelf life and perishability. A leveling buffer you can't sell before expiry isn't a buffer; it's shrink. Perishable SKUs cap how much smoothing inventory you can hold.

The overarching rule: don't level chaos. Heijunka assumes a baseline of stability and reliable changeovers. If your line goes down unpredictably and your setups vary wildly in length, fix that first. Leveling amplifies a stable process; it can't paper over an unstable one.

A practical rollout for a CPG plant

Segment your SKUs. Sort into runners (high, steady volume), repeaters (moderate, regular), and strangers (low or sporadic). Runners get tight leveling; strangers may stay campaign-scheduled or get folded in periodically.
Choose the leveling interval. Daily, shift, or pitch. Start coarse — you can tighten later.
Compute EPEI against takt. Add up your run times plus realistic changeover times across the full SKU set. If EPEI is far longer than your target interval, your changeovers — not your demand — are the bottleneck. Go do SMED.
Size the leveling buffer. Use demand variability, process stability, and shipment frequency. Keep it separate from safety stock on paper.
Build the heijunka box (or its digital equivalent). Rows = SKUs, columns = intervals, fill cells in proportion to demand. Respect allergen/CIP sequencing as hard constraints.
Measure adherence, then shrink the interval. Track how reliably the floor runs the pattern. As changeovers and stability improve, lower EPEI and level more finely.

Grounding the demand inputs that feed this is its own discipline — see forecasting that survives contact with the floor.

How an agentic scheduler operationalizes heijunka

A heijunka box on a whiteboard is brilliant when demand and changeover times are stable. They rarely stay stable in CPG. Every time the forecast shifts or a sequence-dependent changeover changes, the leveled pattern needs recomputing — and doing that by hand is exactly the work that gets skipped under pressure.

This is where an agentic, constraint-aware scheduler earns its keep: maintaining a level, repeating, sequencing-legal pattern automatically, and recomputing the mix as demand proportions and changeover times move. It's the digital heijunka box that never goes stale — preserving the rhythm the floor depends on while respecting allergen, CIP, and shelf-life constraints. (For the architecture behind that, see why headless and agentic.)

Takeaways

Mura is a choice, not a fact. The sawtooth schedule is fixable.
Forecast gives quantity; heijunka gives rhythm and sequence. You need both.
Level by volume and by mix. Mixed-model sequencing in small batches is the move.
EPEI is your feasibility gauge. It's also why SMED and leveling are inseparable.
Budget the leveling buffer deliberately — and keep it distinct from safety stock.
Don't level chaos. Stability and reliable changeovers come first.

Using Toyota's playbook, the payoff at scale was real: production leveling and broader lean techniques helped Toyota massively reduce production times and inventory levels through the 1980s. The same logic — smooth the rhythm, shrink the batches, level the mix — is what turns a lumpy CPG schedule into one the floor can actually run.