In short: A demand plan arrives in units and a warehouse consumes hours, and the conversion runs on lines rather than units, because travel dominates picking time. The same 100,000 units at eight units per line and at two and a half units per line is 12,500 lines against 40,000, or 125 hours against 400 at a hundred lines an hour. Warehouse labour planning has to be laid against a day of week and hour of day profile, since an hour cannot be banked from Tuesday and spent on Thursday. The flexible layer ranks one way by cost and another by notice period, and the notice period is what binds, so the forecast accuracy that matters for labour is accuracy three or four weeks out.
The demand plan for next week says 118,000 units, up nine percent on this week. The labour plan says forty two pickers on earlies and twenty eight on lates, which is what it said this week, and the week before.
Somebody did think about it. They looked at the volume, saw it was up, added two people to earlies and moved on. That is the entire conversion from a demand plan into a labour plan in a large number of distribution centres, and it is why the same site can be forty hours overstaffed on Monday and short on Thursday while the weekly total lands close to right.
The conversion itself is arithmetic on drivers you already capture. What makes it feel hard is that the demand plan arrives in units, the operation consumes hours, and nothing in the standard planning stack does the translation between them.
Units are the wrong driver and lines are only sometimes the right one
Different work in a warehouse is driven by different counts. Receiving scales with pallets and cartons. Put-away scales with pallets and with the distance to the location. Picking scales mostly with lines, because travel dominates and travel is a function of how many locations you visit. Packing scales with orders and then with items inside the order. Loading scales with pallets. Value added services scale with units and are usually the only activity where the unit count is the honest driver.
De Koster, Le-Duc and Roodbergen, reviewing order picking design and control in the European Journal of Operational Research in 2007, put travel at roughly half of picking time in a typical manual operation. Travel responds to the number of picks, so the number of lines carries more information about picking hours than the number of units does.
The consequence shows up as soon as basket composition moves. Take 100,000 units at eight units per line: 12,500 lines. Take the same 100,000 units at two and a half units per line: 40,000 lines. At a hundred lines an hour that is 125 hours against 400 hours, from an identical volume forecast. A shift of a few points in channel mix between retail replenishment and direct-to-consumer moves required hours by more than most sites' entire flexible layer, and the volume forecast never flickers.
Those ratios also move seasonally within a channel. Gift baskets, multibuys, seasonal ranges with deeper pack quantities and clearance activity all change units per line, and they change it in the weeks where you have the least room to react. Which means the conversion factors have to be estimated by channel and by period from your own order history rather than set once as a site average.
The practical version is a small table: for each channel and each period, units per line, lines per order, cases per pallet received. Apply those to whatever the demand plan produces and you have a driver forecast. That table is worth more to a warehouse manager than another decimal place on the volume forecast.
The intraday profile decides whether the plan is staffable
Weekly hours is close to useless for a shift-based operation, because you cannot bank an hour from Tuesday and spend it on Thursday, and labour arrives in blocks with a minimum length.
Two profiles matter. The day-of-week profile is driven by when orders land, which for retail replenishment is the customer's own ordering calendar and for direct-to-consumer is a Sunday evening and Monday concentration that most sites can describe from memory but few have in their plan. The intraday profile is driven by carrier cut-offs. Everything shipping today has to be picked, packed and on the trailer before the last collection, which puts a wall in the middle of the afternoon that the weekly total cannot see. Twelve hours of work in front of a four o'clock cut-off is a different operational problem from twelve hours of work with no cut-off at all.
Building the profile is straightforward. Take a year of transaction data, compute a day-of-week index and an hour-of-day index by channel, and apply them to the driver forecast. Then lay the shift pattern over the result hour by hour. What you get is a picture of which specific hours are short rather than a weekly variance number, and the two lead to completely different actions.
The profile is also partly under your control, which is the part that gets forgotten. Wave release timing, negotiated cut-offs, pre-picking known volume the afternoon before, and staggered shift starts all reshape the curve. Two hours of smoothing usually costs less than the flexible capacity it replaces, and it does not need a notice period.
The flexible layer and what each option costs to call on
Every site runs a fixed core plus something flexible on top. The design question is how large the core should be and what each layer above it costs, in money and in notice.
Overtime. The highest cost per hour and the shortest notice, sometimes same day. Reliable in small quantities and unreliable in large ones, because the same willing people get asked every week and the pool thins. Bounded by working time rules.
Agency labour. A moderate premium over own labour, notice usually measured in days for meaningful numbers, and a fill rate that behaves like a probability rather than a commitment. Comes with a productivity discount covered below.
Annualised hours. Contracted annual totals with a variable weekly distribution. The cheapest flexibility available because the hours are paid at standard rates, with two constraints: the notice period is contractual and often runs to weeks, and the bank of hours depletes across the year, so flexibility spent in September is not available in November.
Fixed-term seasonal contracts. Committed months ahead, cheapest per hour for sustained peak volume, and no flexibility at all once signed.
Ranking these by cost per incremental hour gives one order. Ranking them by notice period gives a different one, and the second ranking is the one that binds. A flexible layer is committed against a forecast made at the notice horizon, so the forecast accuracy that matters for labour is accuracy three or four weeks out rather than accuracy next week. Most sites measure the second and plan against the first.
Write the call-off rules down before peak: which layer gets used at which level of forecast volume, who authorises each, and what the deadline is for each decision. A rule that says agency headcount for week 47 is fixed on the Monday of week 43 is worth having even when the forecast is still moving, because it converts an argument into a date.
Temporary labour does not arrive at standard
A new picker is slower than an experienced one, and the plan usually assumes otherwise. Wright described the underlying pattern in the Journal of the Aeronautical Sciences in 1936: unit time falls by a roughly constant percentage with each doubling of cumulative output. In a warehouse the practical shape is a new starter somewhere around two thirds of standard rate in the first days, converging over several weeks, with the convergence speed depending on how much of the task is location knowledge rather than motion.
Forty agency heads planned at standard rate is not forty heads of output. If first-week productivity is sixty five percent, the plan needs the effective figure, and peak is exactly the period where the temporary share is highest and the ramp time is shortest.
Two costs travel with it. Experienced staff supervising new starters lose their own output, so buddying one known-good picker to two new ones costs you part of the picker as well. And some proportion of peak hires do not complete, so the hire number has to be sized against the headcount you need on the floor in week six rather than the headcount you need on day one.
Your own learning curve is recoverable from data you already hold. Every operator has a first shift in the WMS. Plot rate against shifts worked, average across a few hundred starters, and you have a curve specific to your site, your system and your task mix. Use it to convert planned heads into effective hours, and use it to justify starting temporary staff two or three weeks before the volume needs them, which is almost always cheaper than hiring the same people later at full rate.
Measure against a standard rather than against last year
Comparing this November against last November confounds volume, mix, layout changes, system changes and staffing. It tells you the operation used more hours without telling you whether it should have.
Two routes to a standard are worth knowing. Predetermined motion time systems decompose a task into elemental motions with published times, an approach set out by Maynard, Stegemerten and Schwab in Methods-Time Measurement in 1948 and still the basis of engineered standards work. They are precise and expensive to build and maintain, and they earn their cost where the task set is stable.
The other route is regression on your own transaction history. Take shift-level paid hours as the dependent variable and shift-level activity counts as the independent variables: lines picked, units, cartons received, pallets put away, orders packed. The coefficients are your minutes per driver. This costs a few days, uses data already sitting in the WMS and the payroll extract, and it refreshes itself as the operation changes.
Regression is the better starting point for most sites, partly because it is quick and partly because the residuals are informative on their own. A shift that consistently burns more hours than its drivers explain contains something the model has not been told about, and that something usually turns out to be worth finding: a slotting problem, a recurring system outage, a supervisor difference, or a product group whose work content nobody had measured.
Once the coefficients exist, the labour plan becomes mechanical. Forecast drivers, multiply by coefficients, add the fixed hours that do not scale, divide by effective productivity, spread across the day and hour profile, compare against shift capacity.
The limit
Everything above is bounded by whether the people exist. A site on an estate with three other distribution centres competing for the same labour pool has a hard ceiling on weekly additions, and that ceiling has nothing to do with what the plan calculated. Agency fill on a peak week is frequently below what was ordered, and the shortfall correlates across the estate because everyone peaks in the same fortnight.
A labour plan that is optimal on paper and unfillable is worse than a conservative one that can actually be staffed, because commitments get made downstream of it. Carrier bookings, customer cut-off promises and inbound appointment slots all assume the labour will be there, and when it is not, the failure surfaces three steps away from its cause and gets attributed to the wrong thing.
So the plan has to be constrained by realistic supply rather than by cost alone. Track your own agency fill rate by week of year and treat it as a probability: if you need forty and you historically receive thirty three in week 47, order against thirty three and close the gap with a layer that has a longer notice period and a higher certainty. Where the notice period and the forecast horizon do not fit together, the answer is structural, meaning start earlier, contract earlier, or reshape the intraday profile, rather than a better weekly plan.
And where the forecast genuinely is not reliable at the notice horizon, the honest position is to commit a buffer of hours and accept that some will be idle. That is a real cost, it is smaller than the cost of missing a peak, and it belongs in the plan as a stated decision instead of arriving as a surprise.
Pull one quarter of shift-level data from the WMS and the payroll extract, regress paid hours on lines, units, cartons and pallets, and look at both the coefficients and the residuals; you will have a working labour standard by the end of the week and a shortlist of the shifts your operation has never been able to explain.