How Is AI Used in Warehouses? Small-Warehouse Examples That Work

AI is used in warehouses to decide where stock lives, forecast how much work tomorrow brings, plan which locations to count and sort returns, and it drives robots in the largest sites. A small warehouse gets the most from the first four, using exports and a spreadsheet. Robots only start to pay when volume is high, steady and the layout is fixed.

Most pages on this topic describe automated sites with thousands of lines a day. A warehouse with six staff and a few hundred orders a day can still use AI productively, but the jobs are different: reading your own pick history, planning shifts and counts, and triaging returns. One illustrative small warehouse below works through four of those jobs, then faces a plain test for robots.

Follow me on Instagram@sagnikteaches

The warehouse and its starting numbers

Take an illustrative small operation that stores and ships stock for several online sellers: about 8,000 square feet, six people on the floor plus a supervisor, around 1,200 active products and roughly 150 orders a day, averaging 2.4 lines per order. That is about 360 pick lines daily. Pickers push carts, walk the aisles and pack at two stations. Stock is held in a basic inventory system that records each product's location.

Connect on LinkedInSagnik Bhattacharya

Four numbers describe how the place performs, and you should write yours down before changing anything: picks per labour hour, pick errors per thousand lines, the share of locations that are wrong at a count, and the average days a returned item waits before it is back on a shelf or written off. Say ours are 62 picks an hour, 7 errors per thousand lines, 9% of counted locations wrong, and 6 days for returns. Everything below tries to move those numbers.

Subscribe on YouTube@sagnikteaches

Job one: slotting by how often things are picked

Slotting means deciding where each product lives so that the most frequently picked items are the quickest to reach. In many warehouses a small share of products produces most of the picks, and those products often sit wherever there was space when they first arrived.

The data you need is three months of order lines (product code, quantity, date) and the current location list. A spreadsheet can do the ranking, and an AI assistant can set it up and check the logic if you describe the columns. With a business plan, upload the export with product codes only and no customer details, since customer names and addresses are irrelevant to the analysis.

I have a CSV of order lines: order_id, product_code, quantity, date.
And a second file of locations: product_code, aisle, bay, level.
1. Rank products by number of pick lines (not units) over 13 weeks.
2. Show the cumulative share of lines for the top 5%, 10%, 20% of products.
3. Mark Class A (top products giving 60% of lines), B (next 25%), C (rest).
4. List the Class A products that are NOT on level 1 or 2, or are more
   than three bays from the packing station.
5. Flag products that are frequently picked together (same order,
   more than 30 times) so they can sit next to each other.
Explain each step so I can repeat it in Excel.

Illustrative result for our warehouse: 120 products (10%) give 60% of lines, 300 products give the next 25%, and 780 give the last 15%. Of the 120 fast movers, 47 sit on the top or bottom shelf and 31 are more than 40 metres from the packing station. Moving those 78 products to waist-height locations near packing is a weekend's work for two people.

Estimate the saving with a sum, not a promise. Of the 360 daily lines, 216 come from Class A. If the average walk for those picks falls from 38 metres to 24 metres, each pick saves 14 metres, so 216 lines save about 3,000 metres a day. At a walking speed of roughly 1.2 metres a second, that is around 42 minutes of walking a day, about 0.7 of a picker-hour. At $18 an hour, that is roughly $12.60 a day, or about $265 over 21 working days. It is not life-changing, and it costs only the weekend of moving, but the effect on picks per hour is usually more noticeable than the sum suggests because shorter walks cause less fatigue and fewer wrong-bin errors.

Re-run the ranking every quarter. Seasonal products move in and out of Class A, and a slotting plan that is six months old puts this year's winter stock in last year's best spot. For the loading side of the same problem, planning pallets and trucks from orders uses the same kind of export.

Job two: forecasting labour and building the week's rota

Labour is the largest running cost in the building, and most small warehouses plan it by feel: fewer people on Monday because Monday is quiet, extra hands before a sale because there was a sale last year. AI helps by reading the last 12 to 26 weeks of orders by weekday and flagging what the pattern says about the coming one.

Give the assistant daily order and line counts, plus a list of known events: promotions your clients have told you about, public holidays, any carrier cut-off changes. Ask for a forecast range, not a single number, and ask it to show how it arrived at the range so you can disagree with it.

An illustrative output: "Based on 12 weeks, Mondays average 168 orders (range 140 to 195), Tuesdays 151, Wednesdays 142, Thursdays 148, Fridays 120. Next week includes a client promotion on Wednesday that last ran in week 6, when that week's Wednesday was 62% above its average. Plan for 230 orders on Wednesday as a likely high case." You then turn that into labour: at 62 picks an hour and 2.4 lines per order, 230 orders is 552 lines, or about 9 picking hours plus packing time. A rota that puts four people on picking on Wednesday instead of the usual three covers it, and costs one extra person-day, while a late-night scramble costs overtime and mistakes.

The limit is that AI only knows what you tell it. The client who launches a promotion and doesn't say so makes any forecast wrong. So the best use of time is a one-line weekly message to clients: "tell us by Thursday of any promotions or big sends next week." That habit does more for the forecast than any model. Compare notes on this with AI tools for a small warehouse or 3PL.

Job three: cycle counts that target the stock that matters

A cycle count means counting a few locations every day instead of closing for a full stocktake. The question is which ones. A fixed list in location order is fair but wasteful. Counting by class is smarter: A products monthly, B quarterly, C twice a year, plus any location where a pick error or a "short pick" was reported that week.

With our classes, 120 A products need counting every 21 working days, which is about six a day. The 300 B products over 63 days is five a day, and the 780 C products over 126 days is six a day. That is 17 locations a day, plus the exceptions, say 25 in total, roughly 30 to 40 minutes for one person. Ask the assistant to produce each morning's list from a stock export and last week's error log, so the same list can be printed or sent to a handheld.

The payoff is in what the counts reveal. If 9% of counted locations are wrong, you can ask the AI to group the mismatches by cause: mis-slotted product (item in a neighbouring bay), unit-of-measure mistakes (a case counted as an each), unreceived goods, and damaged stock never written off. An illustrative month produces 40 mismatches: 18 neighbouring-bay errors, 9 unit-of-measure, 8 receiving errors and 5 unexplained. That tells you to fix the labelling in the aisles and to add a check at receiving, rather than counting harder. The cycle count template with ABC classes has the full method, and tracing pick errors back to their cause covers what to do with the error log.

Job four: returns triage

Returns are the least glamorous warehouse job and one of the costliest, because each item needs a decision: back on the shelf, repair, discount, or bin. In our warehouse, an item waits an average of 6 days for that decision, which ties up stock and delays refunds.

The AI job is classification, not judgement. Staff photograph each returned item at the returns bench and tick the reason code. The assistant then reads the client's returns policy and the notes, and proposes a disposition with its reasoning. A rule set, written once by the client and the supervisor, decides what is allowed to happen without a person.

The first row does the heavy lifting: sealed items back on the shelf the same day, which cuts the 6-day wait for the largest group. For the rest, AI drafts the photo note to the client in a consistent format. It also helps with the pattern side: if one product comes back "wrong size" four times in a month, the problem is the listing, not the warehouse. See spotting patterns in complaints, returns and defects for that analysis.

Two smaller uses worth a trial: receiving checks and shift handovers

Two jobs outside the main four cost almost nothing to test. The first is receiving. Inbound deliveries arrive with a packing list that rarely matches the purchase order perfectly. Ask the assistant to compare the two lists (product codes and quantities, no prices needed) and return only the differences: short, over, wrong product, and unknown codes. An illustrative delivery of 38 lines produces three differences, a short of 12 units, an over of 6 and one code the system has never seen. The receiver resolves those at the dock, with the driver still there, rather than three days later when a picker finds an empty carton.

The second is the shift handover note. At the end of each shift, the supervisor dictates or types five lines: what is behind, what broke, what is promised to a client tomorrow. The assistant turns those into a short, consistent note for the next shift, with open items listed first. It sounds trivial, but it removes the common warehouse problem of the afternoon team not knowing what the morning team had already agreed with a client. Keep the note factual, give each open item an owner, and clear the list daily so the note never grows into a diary.

Both uses share a property that makes them good first trials: a mistake is visible within hours. A wrong receiving comparison gets caught when the pallet is counted, and a vague handover gets caught when the next shift asks a question. Jobs with fast feedback teach you how far to trust the tool before you rely on it for something slower, like a quarterly slotting plan.

The quiet failure: slotting on bad location data

An illustrative mistake shows why counting comes before cleverness. The supervisor runs the slotting analysis, finds 78 products to move, and spends a weekend moving them. On Monday, the pick rate drops from 62 to 54 an hour. The reason: 11 of the moved products had been in different locations from the system record for months, so the "move" list sent staff to take stock from places where it wasn't, and the system's new locations were then copied from the old, wrong ones. Pickers lost time hunting for items and a few orders shipped short.

The lesson is a sequence. Count the Class A locations first (a day's work, 120 products), correct the records, then slot. And after slotting, run a count on the moved products the next week. Add "locations verified" to the plan as a step with a name against it. AI helps because it can read the count results and the move list together and flag any product that appears on both lists with a mismatch, which would take a person an hour with two spreadsheets.

When robots enter the picture

Robots in warehouses are real and useful at scale: mobile robots that carry totes, goods-to-person stations that bring shelves to a picker, automated sorters. For a small warehouse, the useful question is not "do robots work?" but "does my volume use one enough?". A practical test, using your own figures:

1. Is the walking the problem? If pickers spend over 40% of their time walking and your layout is stable, mobile robots that move carts could cut that. If packing is the bottleneck, they won't help.
2. Is volume steady? Equipment bought for a peak sits idle for most of the year. If your peak is three times your average, a flexible lease or extra temporary staff may be cheaper.
3. Is the data clean? Robots follow the system's locations. They make bad location data a daily crisis rather than an occasional nuisance.
4. What is the payback in your numbers? Take the quote, add installation, software and support, divide by the labour hours saved a year at your hourly cost, and ask for the vendor's reference in a site of your size. If the answer is over three years, treat it as a long bet.

Take our warehouse: 360 lines a day at 62 picks an hour is under six labour-hours of picking. If walking is half of that and robots honestly cut it by a third, the saving is about an hour a day, roughly 240 hours or $4,350 a year at $18 an hour with on-costs ignored. Even a site with six people picking all day would save about one person's time on the same assumptions, around $37,000 a year. Whether a system costs less than that over its life, with support and downtime included, is a question for real quotes, and at this warehouse's volume the earlier jobs are the better first spend. AI robots for warehouses: do they pay off for a small operator? works through the sums in more depth.

Where to start in a six-week plan

If you want one path through all this, sequence the jobs by dependence, not by excitement.

1. Weeks 1 and 2: count the A locations and fix records. Set the baseline numbers: picks per hour, errors per thousand, wrong locations, return wait.
2. Week 3: run the slotting analysis, move the first 30 products, and recheck the pick rate the following week.
3. Week 4: start daily cycle-count lists by class and log mismatch causes.
4. Week 5: write the returns rule table with each client, and begin photo-and-code returns.
5. Week 6: build the weekly labour forecast and begin the Thursday message to clients about promotions.

If you work in spreadsheets, a general assistant can write and explain the formulas for each of these analyses; ask it to show every step so you can repeat the job next quarter without it. For anything involving a stock system's own features, check the vendor's help pages, because built-in AI varies a lot between products and changes often. If you want help deciding which of these jobs fits your own building, that is what my AI implementation consultation covers.

Reading the results after a quarter

At the end of the first quarter, compare the four starting numbers. Picks per hour should have risen from slotting and fewer wrong-bin hunts. Errors per thousand lines should have fallen as counts caught bad locations. The wrong-location share should be well below its starting 9%, and the returns wait should be down towards a day or two for the sealed group. If one number hasn't moved, look at the step that was meant to move it. A quarter is also the right time to rerun the slotting analysis and to decide whether any remaining pain is about walking, counting or packing, which tells you what to spend on next, whether that is another process change, a handheld scanner, or, much later, a robot.

Questions small warehouse owners ask about AI

Do I need a warehouse management system to use AI?

Not for the first jobs. Slotting analysis, labour forecasts, cycle-count lists and returns triage all work from exports you can pull from a stock system or spreadsheet. A WMS makes the results easier to act on and keeps locations accurate, but poor location data in a WMS will mislead AI just as it misleads people.

How accurate does my stock data need to be?

Accurate enough that a picker finds the item where the system says. If locations are wrong on more than a few per cent of lines, fix that first with a cycle-count programme. AI amplifies the quality of the data it is given, so slotting a warehouse on bad location data simply moves the confusion around.

Will AI replace my pickers?

For a small warehouse the realistic effect is different. AI plans the work, spots errors and reduces walking, so the same team handles more orders or the same orders with fewer mistakes. Robots, where they make sense, change the job rather than remove it, and they need people for exceptions, loading and maintenance.

What is the cheapest first step?

Take three months of order lines and rank products by how often they are picked. Move the fastest 10% to the easiest-to-reach locations, count them first and see whether picks per hour improve. It costs staff time and a spreadsheet, and tells you how much the other ideas are worth.

Further reads

• Forklift Safety Checklist: AI-Built Daily Inspection Forms — Daily forklift inspection forms drafted with AI.
• How to Automate Returns and Refunds With Clear AI Rules — Clear rules for automating returns and refunds.
• AI for Delivery Managers: A Daily Exceptions Report From Route Exports — A daily exceptions report built from route exports.
• Can an AI Chatbot Handle Order Tracking and Returns? — Whether a chatbot can handle tracking and returns.
• Dock Scheduling for a Small Warehouse: Slot Rules and AI Reminders — Slot rules and reminders for inbound deliveries.
• Cost per Drop: Work Out Delivery Cost From Route and Fuel Data — Costing outbound deliveries per drop.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: general warehouse operations practice; vendor documentation for stock and warehouse systems; illustrative figures throughout.