Forecast warehouse labour in three steps: convert each order into minutes of work using time standards you measured yourself, forecast order volumes by week from last year's history plus planned promotions, then divide the hours by your core team's hours. Whatever is left over is the gap your agency staff or overtime must cover.
AI does the middle step well and the other two badly, so split the work that way. It reads a year of order history and projects it quickly, while you supply the standards and the promotion calendar. The example warehouse here is a small fulfilment site with a core team of three, and every figure in it is an illustration to replace with your own.
Turn one order into minutes of labour
A forecast of orders is no use to a shift planner until it becomes hours. The conversion factor is the time one average order takes across every step it passes through. You get it by measuring each task, then adding them.
Say a typical order has 2.4 lines (different products). Your measured standards might be 85 lines picked per hour, 24 orders packed per hour and 120 orders labelled and staged per hour. Convert each to hours per order and add them:
| Task | Standard | Hours per order | Minutes per order |
|---|---|---|---|
| Pick | 85 lines an hour, 2.4 lines per order | 0.0282 | 1.69 |
| Pack | 24 orders an hour | 0.0417 | 2.50 |
| Label and stage | 120 orders an hour | 0.0083 | 0.50 |
| Total direct work | 0.0782 | 4.69 |
That is 4.69 minutes of hands-on work per order. Nobody works 100% of a paid hour, though. Breaks, walking back from the despatch bay, waiting for a pallet and answering a question all take time. If your people spend about 85% of a paid hour on direct work, divide by 0.85 and you get roughly 5.5 paid minutes per order. That number drives every sum below.
Measuring your own standards
Do not copy another warehouse's rates. Pick a normal Tuesday, choose three people of average speed, and time ten orders each for each task with a phone stopwatch. Record the line counts so you can work out lines per hour. Average the nine results and round. Then repeat in a busy week, because people speed up when the dock is full, and decide which figure you want to plan against.
Planning against your fastest people leaves the schedule short every time. Planning against your slowest wastes hours. The middle is usually right, and you can fine-tune it after a month of comparing forecast hours with actual hours.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowA normal week for three people
Take a baseline week of 1,150 orders, spread unevenly across five days. Multiply each day's orders by 5.5 paid minutes and divide by 60 to get hours, then by eight to get person-shifts.
| Day | Orders | Hours needed | Person-shifts | Against a core of 3 |
|---|---|---|---|---|
| Monday | 280 | 25.8 | 3.2 | 1.8 hours short |
| Tuesday | 240 | 22.1 | 2.8 | 1.9 hours spare |
| Wednesday | 210 | 19.3 | 2.4 | 4.7 hours spare |
| Thursday | 200 | 18.4 | 2.3 | 5.6 hours spare |
| Friday | 220 | 20.2 | 2.5 | 3.8 hours spare |
| Week | 1,150 | 105.8 | 13.2 | 14.2 hours spare of 120 |
Three people cover the week with room to spare, but not the Monday. That is a staffing pattern, not a shortage: Monday's weekend orders pile up, and the sensible answer is a staggered start, not a fourth person. One of the three could start two hours earlier on Monday and finish early on Wednesday, and the week balances. Planning by day rather than by week is what makes that visible.
What to export and how to prepare it
The forecast needs one table: the date of each order and nothing sensitive. Most storefronts and warehouse systems can export order lists as a spreadsheet. Pull at least a year, then reduce it to the columns below before you give anything to an AI tool.
- Order date: the day the order was placed, or the day it was released to the warehouse if those differ. Pick one and keep it.
- Lines and units: the number of different products and total items, if you can export them.
- Channel: the shop, marketplace or wholesale account, because they peak at different times.
- Promotion flag: a yes or no column you add by hand for sale days.
Delete customer names, addresses and emails. The assistant never needs them for a volume forecast. Then summarise to weekly totals in a pivot table, so you hand over about 104 rows and not 60,000.
A first forecast prompt
Below are weekly order totals for the last 104 weeks, with a column marking weeks that had a sale. Forecast the next 12 weeks. Use the same week last year as the base, adjusted for the growth between the last 13 weeks and the same 13 weeks a year earlier. Show the working in a table: week, last year, growth factor, forecast. Say which weeks you are least sure of and why. Do not invent promotions; I will tell you the calendar next.Asking for the working in a table makes the result checkable. A reply for the six peak weeks might read (illustrative): week 47, 1,500 last year, factor 1.12, forecast 1,680; week 48, 2,900, 1.12, 3,248; and so on. You can verify each multiplication in the sheet in seconds, and that is the point, because a forecast you cannot check is a guess with a nice layout.
The same technique is covered in more detail in the seasonal forecast tutorial, which is worth a read if your history has gaps or a recent step change.
From forecast orders to a staffing gap
Now apply the 5.5 minutes. Multiply forecast orders by 5.5 and divide by 60, then compare with the core team's weekly hours. With three people on 40 hours, that is 120 hours. The gap, in hours, is what the agency or overtime covers. Divide by the productive hours one extra person gives you to get heads.
Agency starters are slower for their first week while they learn the site, so apply a haircut. Here we assume an agency person delivers 75% of a core person's output in a week. The sum for the peak looks like this:
| Week | Last year | Forecast orders | Hours needed | Gap over 120 | Agency heads at 75% |
|---|---|---|---|---|---|
| 47 | 1,500 | 1,680 | 154.6 | 34.6 | 1 |
| 48 (big sale) | 2,900 | 3,248 | 299.0 | 179.0 | 6 |
| 49 | 2,300 | 2,576 | 237.1 | 117.1 | 4 |
| 50 | 2,100 | 2,352 | 216.5 | 96.5 | 3 |
| 51 | 1,900 | 2,128 | 195.9 | 75.9 | 3 |
| 52 | 1,200 | 1,344 | 123.7 | 3.7 | 0 |
Each row is the forecast times 5.5, divided by 60, minus 120, then divided by 30 (40 hours at 75%) and rounded to the nearest whole person, with core overtime covering the small remainders, such as the 3.7 hours in week 52. Notice the shape. A three-person core runs the year, and the peak calls for six temporary people for one week and three or four for the next three. Booking six people on a Wednesday for Monday is not realistic, which is why the trigger rules below fire weeks ahead.
Writing agency triggers before the pressure arrives
Labour decided on the day usually means short-notice rates for whoever happens to be free. Write the triggers down in advance, in numbers, so the decision is a lookup. Below is a set a small warehouse could use, with the reasoning for each.
- Forward forecast trigger. If the forecast for any week four weeks out exceeds core hours by more than 10%, ask the agency for availability. Lead time is the constraint, so the trigger uses the forecast rather than the backlog.
- Book trigger. If the same week still exceeds core hours by more than 10% two weeks out, confirm the booking. Cancelling a booking usually costs less than lacking people in the week.
- Daily spike trigger. If a single day's forecast exceeds core capacity by more than 30%, add overtime for named staff and one agency shift. Overtime has limits per person, so check them.
- Backlog trigger. If unpicked orders at 14:00 exceed half a normal day, call the agency for next-day cover and brief the team on priorities.
- Stand-down trigger. If actual orders run 15% under forecast for three days, release the next block of agency hours at the earliest allowed time.
The percentages are starting points for you to tune after a season. What matters is that someone owns them and that they are written where the supervisor can see them. When a trigger fires, the agency call is a task, not a debate.
A quick cost comparison for the shortfall
Say the agency charges $26 an hour per person and you pay core staff $18, with overtime at 1.5 times. An agency person at 75% output costs $26 divided by 0.75, about $34.70 per productive hour. Overtime for a core person costs $27 per hour, but each person can give perhaps eight extra hours in a week before tiredness and rules bite. For a gap of 117 hours in week 49, three core people could cover about 24 hours of it in overtime for roughly $650, and the remaining 93 hours would need agency staff at about $3,230. Replace every rate with yours, but the pattern holds: overtime first, up to its limit, then agency.
Orders per hour: the daily check
A weekly forecast tells you what to book. A live check tells you whether the day is going to plan. Orders per hour is the measure: the count of orders released to the floor each hour against the count the team shipped. Chart the two lines for the morning and you will see a backlog forming an hour before the supervisor feels it.
Suppose Monday's cut-off for same-day despatch is 15:00 and the day's 280 orders are 110 waiting at 08:00 and 170 arriving through the day. The team of three can ship about 33 orders an hour (180 paid minutes divided by 5.5), so a full eight-hour day clears roughly 262 orders and Monday is already short. Track the gap between released and shipped at 10:00, 12:00 and 14:00. If more than 60 orders are still waiting at noon, you know before lunch that the 15:00 cut-off is at risk and can borrow someone from goods-in.
Put the check on a screen near the pack benches, with a single number: unshipped orders due today. When that number moves in the wrong direction two hours running, a supervisor reacts. The slotting tutorial explains how to shorten the pick walk, which lowers the minutes per order in the first place.
Where the forecast went wrong in one pre-peak week
A story in illustration. A warehouse forecast 1,680 orders for the week before its big sale and booked one agency person. The marketing team had sent an email to the whole customer list on the Tuesday, which no one told operations about, and the week landed at 2,150 orders. The team ran four days behind, and two days of next-day delivery promises were broken.
Nothing was wrong with the sums. The input was incomplete. The fix was a standing item: marketing posts its calendar in a shared sheet every Friday, and the forecast prompt includes a column for each campaign with its list size. Ask the assistant to compare any week's forecast with the campaign calendar and flag any promotion that the history does not explain.
The same warning appears in the stock forecasting mistakes tutorial, where promotions and stock-outs distort the history the model learns from. A week with a stock-out shows low orders because customers could not buy, not because demand was low.
Shift patterns that match the day, not just the total
Total hours are only half of a plan. When the work arrives matters as much, and the smartest saving is often moving start times. Look at your hourly order release profile for a typical day and compare it with your shift pattern.
A warehouse whose orders land overnight and in a rush at 09:00 needs people at 07:00 on Monday, not at 09:00. A warehouse with a 15:00 courier collection needs packers working hardest between 13:00 and 15:00. Ask the assistant to take the hourly counts and propose a rota for the core team that puts the most hands where the work peaks, with a rule such as "no one works more than five days or more than 9.5 hours in a day". Check the proposal against the working-time rules that apply to your staff, because the assistant does not know them.
The same approach works in other trades: a restaurant matches its rota to covers, and that rota tutorial shows the pattern in a setting where the peaks are daily. The warehouse version is the same, only the unit is orders.
Handling returns, inbound and the work that is not orders
Orders are the biggest driver of hours, but not the only one. Goods-in, returns processing, stock counts and cleaning take time too, and they follow different curves. Returns lag sales by one to three weeks, so a sale week is followed by a returns wave. Inbound deliveries follow your purchase orders, not your customer orders.
Add a second small table to your sheet with an hours-per-unit standard for each. For example, say goods-in takes 0.35 hours per pallet and returns processing takes 4 minutes per item. If your forecast has 20 pallets arriving in week 46 and 140 returns expected in week 49, those are 7 hours and 9.3 hours to add to the totals. These numbers look small, yet they decide whether the Wednesday of a peak week is workable.
Fixed jobs, such as a weekly clean or a monthly count, should be placed into the quiet days on purpose. A short list in the sheet, with the week and the hours, stops them landing in the busiest week by default.
Checking the forecast against what happened
Every forecast is wrong by some amount, so the useful question is how wrong and in which direction. At the end of each week, add two columns: actual orders and actual hours worked. Then calculate the percentage difference between forecast and actual orders. A weekly gap under 10% is good for a small business, and a bias in one direction means an assumption needs adjusting.
- Orders higher than forecast, hours as expected: your standards are generous, so tighten them.
- Orders as forecast, hours higher: your standards are optimistic, or new starters are slower than 75%.
- Orders lower than forecast three weeks running: the growth factor is too high, so cut it.
Review this on a fixed day with the supervisors. They know why a day ran long, and their reasons become the next version of your assumptions.
Cross-training as a cheaper buffer than extra heads
Before booking agency staff, look at the slack inside the core team. In the baseline week above, the three people had 14 spare hours, but the spare hours sat on Wednesday to Friday while the shortage sat on Monday. People who can pick, pack and receive move that slack to where it is needed. List each person's skills in the sheet and mark who can cover which task. If only one person can run the label printer, a day off for that person becomes a hole in your forecast that no sum predicts.
A simple rule is that every task has at least two people who can do it, and every starter, agency or permanent, learns the two highest-volume tasks first. The picking and packing training cards tutorial shows how to put that training on a single card per task, so a new starter is productive on day one and the 75% assumption can rise.
What to keep for next season
At the end of the peak, save a copy of the sheet with the forecast, actuals and triggers that fired. Next year it becomes your best input, and it records which agency start dates worked. Add a line on what you would change, written the week it happened, not three months later. With that record, the second season starts from your own numbers rather than from memory.
Common questions about forecasting warehouse labour
How much order history does the forecast need?
Twelve months of weekly orders is the practical minimum, because it lets you see one full seasonal cycle. Two years is better, since one odd year can mislead. If you have less, say so in the prompt, use the recent weeks as the baseline, and add a manual allowance for known peaks.
What if I do not have time standards for each task?
Time yourself. Watch three people doing ten orders each at a normal pace, record the minutes for picking, packing and labelling, and average them. A rough standard checked on the floor beats a precise figure copied from a book that describes a different building.
Should the AI decide how many agency staff to book?
No. It calculates the gap between forecast hours and core hours, which is arithmetic. You decide whether to book, because you know the lead time, the budget and how quickly your site can train new starters. Keep the booking decision with a named manager.
How often should I rerun the forecast?
Weekly for the next four weeks, and daily during a peak. A weekly rerun takes about ten minutes once the sheet is set up. Always compare the last forecast against what actually happened and note the gap, because that gap is how you learn which assumption is wrong.
Further reads
- AI Tools for Warehousing: Which Suit a Small Warehouse or 3PL? — See which warehouse tools suit a small site or 3PL.
- Cycle Count Template: AI Plans Counts by ABC Class — Schedule counts in the quiet days your forecast reveals.
- Dock Scheduling for a Small Warehouse: Slot Rules and AI Reminders — Spread inbound loads so goods-in does not collide with picking.
- AI Robots for Warehouses: Do They Pay Off for a Small Operator? — Check whether automation beats more agency hours at your size.
- Best AI for Ecommerce Operations: Orders, Stock, Returns, Suppliers — Wider view of the order, stock and returns jobs behind the volumes.
- AI for Warehouse Loading: Plan Pallets and Trucks From Orders — Plan outbound loading, the other big call on dock labour.
- 3PL Pricing Calculator: AI Quotes Storage, Pick and Pack Rates — A 3PL operator builds its own rate card and quote calculator: cost-based storage, pick and pack rates, a worked quote and a margin check.
- 3PL Client KPI Reports: AI Builds Weekly Reports From WMS Exports — A weekly one-page report per 3PL client, built from WMS exports: three KPIs defined exactly, charts that do not mislead, and AI commentary a person signs off.
- AI in Warehouse Operations: Trace Pick Errors Back to Their Cause — Turn a month of pick errors into ranked root causes and a short list of fixes to test, using only your own error log and an AI assistant.
- Walmart OTIF Fines: Build a Dispute Pack From Delivery Records — Which Walmart OTIF fines are worth disputing, the five records to gather, and how AI assembles one evidence pack per PO.
- AI for a Small 3PL: Six Jobs Ranked by Payback — Six small-3PL jobs ranked by payback, with an effort-versus-payback table and a worked billing example.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: the author's own method; figures in the worked example are illustrative assumptions, not benchmarks.