AI in Warehouse Operations: Slotting, Picking and Labor Planning

By Tara Minh · Updated Oct 11, 2026

AI in warehouse operations is the use of machine learning and optimization to decide where stock lives, in what order it's picked, how many people are needed each shift and how robots and humans share the floor. It doesn't replace the warehouse management system (WMS). It sits on top of it, reads the same orders, inventory and movement history, and recommends better answers to four recurring questions: where should this item go, which route should this picker walk, how many hours do we need tomorrow, and which task should the next worker or robot take?

This article is for operations leaders, warehouse and fulfillment managers and the finance partners who sign off on automation projects. It covers each of those four decisions, what changes when AI is involved, a step-by-step rollout and the traps that stall projects. It's a spoke of the logistics group in AI for Operations, so for the wider network picture start with AI in logistics, and for what happens to goods after they ship see AI in reverse logistics.

Key Facts

• Order picking is the most labor-intensive and costly activity in almost every warehouse. A widely cited academic review puts its cost at as much as 55% of total warehouse operating expense (de Koster, Le-Duc and Roodbergen, European Journal of Operational Research, 2007). That figure is the authors' estimate, not a measured industry average.
• The same review calls travel "often the dominant component" of picking time, which is why slotting and routing are the first places to look for savings (de Koster et al., 2007).
• Class-based storage groups products by popularity. In the classic Pareto description, the fastest class holds about 15% of products but drives about 85% of turnover (de Koster et al., 2007).
• Warehousing and storage recorded 4.8 total recordable injury and illness cases per 100 full-time workers in 2024, against 2.3 across private industry (U.S. Bureau of Labor Statistics). Safer routing and pacing is a real benefit, not only a speed gain.
• Amazon says its DeepFleet model, which coordinates robot movement across its fulfillment network, is improving robot fleet travel time by 10% (Amazon). That's the company's own claim about its own network.
• The practical rule: AI earns its keep where a warehouse makes thousands of small repeated decisions that people currently make by habit.

Where AI Fits in a Warehouse

A warehouse runs on a loop. Goods arrive, get put away into locations, wait, get picked against orders, get packed and leave. Every step creates data: scans, timestamps, location IDs, order lines, labor hours. Most of that data sits unused after the shift ends.

AI uses it in the ACE Framework sequence. It ingests orders, inventory and scan history, analyzes how items and people actually move, predicts tomorrow's order volume and mix, and, where you allow it, executes by releasing waves, assigning tasks or pushing a re-slot list to the floor. The first three steps only produce recommendations. The last one changes what people do, so it needs a clear checkpoint.

The four decision areas below are ordered by how quickly they usually pay back and how little they disrupt the floor.

Slotting: Putting the Right Item in the Right Place

Slotting is the assignment of products to storage locations. A good slot plan puts fast movers near the dock and the pack stations, puts items often ordered together near each other, and puts heavy goods at waist height.

The academic baseline is class-based storage. Products are ranked by demand and split into a few classes (A, B, C), each assigned a dedicated zone, with random placement inside the zone. The de Koster review notes that simulation work found full-turnover storage beats class-based on travel distance, but the researchers suggest class-based storage with two to four classes in practice because it's easier to run (de Koster et al., 2007). The review also warns that class rankings need reliable demand statistics, and that they're hard to build when the assortment changes too fast.

That last caveat is where AI helps.

Forecast-driven velocity. A static ABC list is a rear-view mirror. A model that uses seasonality, promotions and trends can tell you which items will be hot in three weeks, so you move them before the surge, not after. The inputs come from the same forecasts described in AI demand forecasting.

Affinity. Models can mine order history for items that ship together and slot them in neighboring locations, which cuts the number of aisles a picker enters per order.

Physical rules. Weight, size, hazard class and temperature are constraints, not preferences. A good optimizer treats them as hard limits.

Cost of moving. Re-slotting uses labor. A sensible model recommends a move only when the expected travel saving beats the cost of the move, and it batches moves into slow periods.

An illustrative example: a distributor with 12,000 active SKUs re-slots monthly by hand-built ABC lists. A model flags 400 items whose forecast velocity will change class within the month and proposes 150 moves ranked by expected saving. The planner approves the top 60 for the weekend crew. The point isn't the numbers, which are made up. It's that the model narrows thousands of possibilities to a short, ranked list that a person can review in minutes.

Picking: Shorter Routes, Smarter Batches

If picking dominates warehouse cost and travel dominates picking time, then walking less is the lever. The classical literature covers four tools, and AI improves how each is tuned.

Routing. Simple rules, such as the S-shape method (walk every aisle that has a pick, end to end) and the largest-gap method, are easy for pickers to follow. Optimal algorithms find shorter paths but can be hard for people to follow and ignore congestion, which is one reason the review notes heuristics are still used in practice (de Koster et al., 2007). Modern route engines compute paths in real time, account for blocked aisles and live congestion, and show the next pick on a handheld, so the picker never has to memorize a rule.

Batching. Grouping several orders into one trip saves walking. The same review reports that even simple batching methods improve on first-come first-served (de Koster et al., 2007). AI-assisted batching goes further by forming batches that account for order deadlines, carrier cutoffs, item locations and cart capacity together.

Zoning. Splitting the floor into zones means each picker walks a smaller area and aisle congestion falls. The cost is that orders are split and must be consolidated before shipping (de Koster et al., 2007). A model can size zones to the expected order mix each shift instead of using one fixed layout.

Wave and task release. Rather than releasing work on a fixed schedule, a system can release it by predicting where congestion will form and holding back work that would clash.

Notice what AI isn't doing here. It's not inventing a new idea. It's tuning decisions that every WMS already makes, using more current data and a wider set of constraints.

Labor Planning: Matching Hours to Volume

Labor is the largest controllable cost in most warehouses, and the hardest to match to demand. Order volume spikes with promotions, weather and carrier cutoffs, while headcount is set days or weeks ahead.

AI helps in three ways.

Volume forecasting at the right grain. Planners need order lines per hour, not units per month. Hourly and daily forecasts by work type (pick, pack, receive, put away, returns) turn into labor hours using engineered standards.

Scenario scheduling. A scheduler can build shift plans that respect contracts, skills, breaks and overtime limits, then show the cost of covering the 90th-percentile day versus the average day. That makes the decision about temporary labor and flex shifts explicit.

Real-time rebalancing. During the shift, a model watches queue depths and throughput and suggests moving people between zones. This works only if supervisors trust it, so show the reasoning ("pack is 40 minutes behind, put-away is idle").

Two cautions. First, labor data is personal data, and productivity scoring can damage morale if workers see it as surveillance. Plan measurement at the team and process level, tell people what's tracked and why, and check local labor and privacy rules. Second, speed isn't the only target. Warehousing and storage had 4.8 recordable injury and illness cases per 100 full-time workers in 2024, more than double the private industry rate of 2.3 (U.S. Bureau of Labor Statistics). Pacing rules that push people past safe limits will show up in injury, turnover and absence figures. Build rest, rotation and ergonomic limits in as constraints from the start.

For the person who owns this work day to day, see the warehouse manager job description.

Robotics Coordination

Many warehouses now run mixed fleets: autonomous mobile robots (AMRs) carrying totes, robotic arms picking items, and people working beside both. The hard problem is coordination, not any single robot.

A fleet manager decides which robot takes which task, which path it uses and who yields at an intersection. Amazon describes its DeepFleet model as a generative AI foundation model that coordinates robot movement across its fulfillment network and says it's improving fleet travel time by 10% (Amazon). Treat that as a vendor's claim about a very large, purpose-built network. A mid-sized site shouldn't expect the same figure, but the direction is useful: congestion between machines is a software problem, and software can learn it.

For most operators the practical questions are smaller:

• Does the robot fleet software talk to the WMS, or do you re-key tasks?
• Who decides priority when a rush order arrives?
• What happens when a robot fails mid-task, and how do people take over?
• Can the same slotting and labor models see robot throughput, so the plan includes both?

If the answer to the last question is no, you'll plan labor and robots in separate spreadsheets, and the gains from both will overlap and cancel.

Summary Table: Decision, AI Role, First Metric

How to Implement It, Step by Step

1. Pick one problem. Choose the one with the clearest cost, usually travel in a single pick area or overtime on peak days. Don't start with all four decisions at once.
2. Measure the baseline. Record lines per hour, travel time or distance per order, overtime hours, order cycle time and injury or near-miss counts. If the WMS can't give you these, fix the reporting first.
3. Check the data. You need accurate location records, item dimensions and weights, timestamped scans and order history that's long enough to cover a seasonal peak. Wrong item dimensions quietly ruin slotting. The checklist is in operational data for AI.
4. Align the inventory side. Slotting and replenishment are linked. If reorder points push too much stock into the forward pick area, no slot plan will save you. See AI inventory optimization.
5. Run in shadow mode. Let the model make recommendations for a few weeks while the current process continues. Compare the recommended routes, batches or slots against what happened.
6. Involve the floor. Walk pickers and supervisors through the recommendations. They'll spot the aisle that's always blocked on Tuesdays or the shelf that can't hold the weight.
7. Pilot on one zone or shift. Keep a control group on the old method so you can tell what changed.
8. Set the autonomy line. Decide what the system can do alone (re-sequence a pick list) and what needs a person (approve a mass re-slot, change a shift plan). Monitoring and alerting can be automated first, as in the AI inventory alert agent blueprint.
9. Expand and review. Widen the rollout only when productivity and safety both hold. Assign an owner for model monitoring and for pulling a change back.

For a control-room view that ties warehouse signals to the rest of the network, see AI operations control tower.

Pitfalls to Avoid

Automating a messy layout. A model can't fix aisles that are too narrow or a dock that sits at the wrong end. Fix physical constraints that a layout change would resolve before you buy software.

Dirty master data. Wrong dimensions, missing weights and stale location records produce confident, wrong recommendations.

Re-slotting too often. Every move costs labor and temporarily confuses pickers who know where things are. Set a minimum benefit for a move and batch moves into quiet periods.

Optimizing speed alone. Faster pick paths that increase repetitive strain are a bad trade. Put injury, rest and rotation rules in as constraints.

Ignoring the people. Workers who feel measured but not heard will find ways around the system. Explain what's tracked, share the results and act on their feedback.

Skipping the baseline. Without before-and-after numbers from a comparable period, you can't tell whether the model helped or the season did.

Planning labor and robots separately. Treat the combined system as one capacity plan.

What to Measure

Track productivity (lines per labor hour, travel time per order), service (on-time shipment, order accuracy, dock-to-stock time), cost (labor hours per unit, overtime share, re-slot labor), safety (recordable injuries, near misses) and model health (how often supervisors override recommendations, and why).