AI for Picking Optimization

If you want to find the biggest cost in a warehouse, follow the pickers. In most distribution operations, order picking accounts for somewhere between half and two thirds of total labour cost, and the majority of that cost is not the act of picking at all. It is walking. Pickers spend most of their shift travelling between locations, and every metre they walk that they did not need to walk is money spent on nothing. This is exactly why picking is where artificial intelligence has the clearest, most measurable payback in the modern warehouse, and it is why any serious conversation about warehouse automation eventually lands here. If you have not yet read the wider context, the complete guide to warehouse automation sets out where picking optimization sits among the other levers, and this article goes deep on the one that usually moves the numbers fastest.

The message up front: AI does not pick faster than a person. It makes the person walk less. Picking optimization is fundamentally a travel-reduction problem, and travel is a function of how you group orders, how you zone the building, how you route the walk, where you slot the stock, and how you wave the day. AI is good at solving all five of those together, which no spreadsheet and no human planner can do at scale. That is the whole value, and it is a large value.

1. Why picking is the prize

Start with the economics, because they explain everything that follows. A picker on a manual operation typically spends between fifty and seventy percent of working time simply moving between locations. The time spent actually reaching for an item, scanning it and placing it in a tote is a minority of the shift. That ratio is the single most important number in warehouse operations, and it is the reason picking dominates the labour budget. Everything else, receiving, put-away, packing, shipping, is real work, but none of it consumes labour at the rate that picking does, because none of the others involves a person walking a marathon a day across a building the size of several football pitches.

Because travel is the dominant cost, the optimization target is obvious: reduce total travel distance without reducing throughput or accuracy. That sounds simple and it is anything but, because the levers interact. Group two orders together to save a trip and you may create congestion in one aisle. Move a fast mover closer to dispatch and you may crowd the golden zone with items that conflict at pick time. Optimize a single picker's route perfectly and you may unbalance the workload so that half the team finishes early and the other half runs late. The picking problem is a system, and systems reward tools that can hold every variable in view at once. That is precisely what AI does well and humans do not.

There is also a demand-side reason picking has become the prize. Order profiles have shifted hard toward smaller, more frequent, more varied orders. The single-line, single-unit e-commerce order is now common where a decade ago the case-and-pallet profile dominated. Small diverse orders are the worst case for travel, because they scatter picks across the whole building with little natural batching. The harder the order profile, the more a good optimization engine is worth, and order profiles have only gotten harder.

2. How AI picking optimization works

At its heart, AI picking optimization takes the pool of open orders and decides three things together: which orders travel as a batch, in what sequence the locations in each batch are visited, and which picker or zone each batch is assigned to. It solves these jointly because solving them separately leaves value on the floor. A perfect batch with a poor route is slow. A perfect route on a poor batch is still slow. The engine searches across the combined space of grouping, sequencing and assignment to minimise total travel while balancing workload across the team and respecting the physical constraints of the building.

The diagram below shows the shape of it. A stream of individual orders arrives with picks scattered across the racking. The engine groups compatible orders into batches, sequences the stops within each batch into an efficient path, and distributes the batches across pickers so no one is idle and no one is drowning. The output is fewer total metres walked for the same set of orders.

The engines vary in sophistication. The simplest are rules and heuristics: cluster orders that share locations, then apply a serpentine or S-shape route through the aisles. The more advanced ones use combinatorial optimization and machine learning to search a far larger solution space, learn from historical pick times which routes are genuinely faster on your floor rather than in theory, and adapt to live conditions such as congestion and staffing. The word "AI" covers this whole range, and it is worth knowing where on the range a given product actually sits, because the marketing rarely tells you.

3. The optimization levers

Picking optimization is not one technique. It is a family of levers that each attack travel and productivity from a different direction, and a mature engine pulls several of them at once. Understanding the levers individually is the fastest way to judge whether a vendor is offering a genuine optimization capability or just a smarter pick list. The table below sets out the five core levers, what each one does, and what it primarily improves.

Read the table as a stack rather than a menu. Slotting reduces travel at the source by putting the right stock in the right place. Batching and zoning reduce how far a picker must range to fulfil the work. Routing shortens the walk that remains. Wave planning orchestrates all of it against the clock. Pull one lever in isolation and you get a modest gain. Pull them together with an engine that understands how they trade off, and the gains compound. That compounding is the reason AI-driven picking optimization outperforms the sum of the individual techniques run by hand.

4. Batching, zoning and wave planning

Batching is usually the first lever an operation reaches for, because it delivers the most obvious win. If two orders both need an item from aisle twelve, sending one picker to grab both in a single trip halves the travel for that location. Multiply that across hundreds of orders and thousands of lines and the saving is large. The catch is that naive batching, grouping by whatever is easiest, can backfire by creating batches that are large but incoherent, sending a picker on a long loop to serve orders that had nothing in common. Good batching is an optimization problem: form the batches that minimise total travel while keeping each batch small enough to carry and fast enough to hit its cut-off. This is exactly the kind of combinatorial problem where an AI engine beats a human planner, because the number of possible groupings is astronomical and the best answer is rarely the obvious one.

Zoning attacks travel from a different direction. Instead of one picker ranging across the entire building, you divide the floor into zones and assign pickers to them. Each picker works a smaller area and knows it well, and orders are either passed between zones (pick-and-pass) or picked in parallel across zones and consolidated later (pick-and-sort). Zoning slashes the distance any individual walks and reduces the chance of two pickers colliding in the same aisle. The design question, how many zones, where the boundaries fall, how work moves between them, is where AI helps, because the right zoning depends on the order profile and that profile shifts by season, promotion and day of week. A static zone map designed once and never revisited is almost always wrong by the time the demand pattern has moved on.

Wave planning is the orchestration layer. Rather than releasing every order the moment it lands, you release work in timed waves aligned to carrier cut-offs, staffing levels and the capacity of packing and shipping downstream. A well-planned wave means pickers are never starved of work and never buried, packing stations are fed at a steady rate, and orders that must make a particular truck are prioritised so they are done in time. The modern trend is toward waveless or continuous release for some operations, where the engine drips work out dynamically rather than in discrete waves, but the underlying job is the same: match the release of work to the capacity to process it, against the clock. AI shines here because it can forecast how long the current queue will take, watch the live picking rate, and adjust the release so the operation hits its cut-offs without a late-afternoon scramble. For the system that actually executes all of this on the floor, see what a WMS is and what it does.

5. Combining slotting, routing and batching

The real power appears when the levers are combined, because they influence one another. Slotting decides where the stock sits. Routing decides the path through that layout. Batching decides which orders share the path. Change the slotting and the best routes change. Change the batching and the value of a given slotting layout changes. These are not independent problems, and treating them as independent is the classic mistake that leaves most of the available gain on the floor.

Consider a concrete example. Slotting analysis shows that two products are very frequently ordered together. Move them close to each other and every batch that contains both now has a shorter route. But that only pays off if the batching engine is actually forming batches that exploit the proximity, and if the routing engine is sequencing the stops to take advantage of it. Optimize slotting in isolation and you get a layout that looks good on paper but that the batching and routing logic never fully use. Optimize all three together and the co-located products, the batches that contain them, and the routes through them all reinforce each other. This is why the strongest picking platforms treat slotting, routing and batching as one coupled problem rather than three separate modules bolted together.

Slotting deserves its own attention because it is the lever that works while everyone is asleep. Get the slotting right and every single pick that follows is a little shorter, a little more ergonomic, and a little less likely to need a replenishment mid-shift. Slotting is also where machine learning earns its keep most cleanly, because the optimal layout depends on demand patterns, product affinities and velocity that shift continuously and that a human cannot re-analyse across thousands of SKUs by hand. For the deep treatment of that lever, see the dedicated pillar on AI slotting optimization, and for the routing lever specifically, the pillar on AI route optimization. The two are the natural companions to this article, and together they cover the levers in the depth a single guide cannot.

The lever that changes the game: at high volume, picking optimization stops being about moving people around a fixed layout and starts being about bringing the stock to the person. Goods-to-person systems, where robots or shuttles deliver the required bins to a stationary picker, eliminate travel almost entirely. The optimization problem shifts from "shortest walk" to "smartest sequence of bin presentations", but it is still an AI-driven picking optimization problem at heart. See the pillar on goods-to-person systems for where the ceiling on manual optimization gets replaced by a different model entirely.

6. Balancing workload and congestion

Minimising total travel is the headline objective, but it is not the only one, and an engine that chases travel alone will produce a schedule that looks great on the total and feels terrible on the floor. Two constraints in particular have to sit alongside travel: workload balance and congestion.

Workload balance is the requirement that the work spreads evenly across the team. An optimizer that packs the most efficient batches for the first three pickers and leaves the rest with the awkward scraps will finish the efficient batches early and the awkward ones late, and the operation is only as fast as its slowest finisher. Balancing means the engine distributes not just travel but total effort, including pick counts, weights and the difficulty of each batch, so everyone finishes together. This matters enormously against a carrier cut-off, where a single late picker can hold up a whole truck. A good engine treats even completion as a first-class objective, not an afterthought.

Congestion is the constraint that the theoretical shortest route ignores. Send every picker down the same optimal aisle and you get a traffic jam, and two pickers waiting for each other to pass have both stopped being productive. As density rises, the shortest individual route stops being the fastest actual route, because the fastest route is the one that avoids the queue. Advanced engines model congestion explicitly, spreading pickers across the building so their paths interfere as little as possible, sometimes accepting a slightly longer individual route in exchange for a smoother collective flow. This is a genuinely hard problem and it is where the gap between a basic heuristic and a real optimization engine shows most clearly, because congestion only reveals itself at scale and under load, which is exactly when it matters.

The practitioner's point is that picking optimization is multi-objective. Travel, balance and congestion trade off against one another, and the right weighting depends on the operation. A low-density operation can chase travel hard. A dense, high-volume operation has to weight congestion and balance more heavily or the travel savings evaporate in queues. An engine that lets you tune those weights to your floor is worth more than one that optimises a single hidden objective and hopes it fits.

7. The honest limits and change management

Now the part the vendor demonstrations skip. Picking optimization is real and the returns are real, but there are limits and traps that decide whether a deployment delivers or disappoints, and none of them are about the algorithm.

The first limit is data quality. Every optimization decision rests on knowing where stock actually is, how much of it there is, and how long picks really take on your floor. If your location accuracy is poor, if inventory records drift from reality, if item dimensions and weights are wrong, then the engine is optimising a fiction. It will confidently route a picker to a location that is empty, or batch by an affinity that no longer holds. The uncomfortable truth is that most of the value gap in picking optimization is a data gap, not an algorithm gap, and no amount of machine learning rescues bad location and inventory data. Fix the master data first.

The second limit is the model of the building. The engine's route is only as good as its map of the aisles, the one-way constraints, the pick faces, the congestion points and the physical reality of the floor. A layout model that is idealised or out of date produces routes that are shorter in theory and slower in practice. Keeping the building model honest, and updating it when the racking moves, is unglamorous work that quietly determines how much of the promised saving actually shows up.

The honest limitation: an optimization engine can compute a schedule that is provably better, and the floor can still reject it. If pickers do not trust the routes, if the batches feel wrong, if the system tells an experienced picker to walk a path that looks foolish even when it is faster, they will work around it, and the measured saving disappears into the gap between the plan and what actually happened. Picking optimization is at least as much a change-management project as a technology project, and the operations that treat it purely as software are the ones that end up with an impressive dashboard and unchanged productivity.

This is why the roll-out matters as much as the engine. The optimizations that stick are the ones introduced with the pickers, not imposed on them: explain why a counter-intuitive route is faster, show the numbers, let experienced staff flag where the model is wrong, and feed that feedback back into the system. Trust is earned when the engine is right often enough that the floor stops second-guessing it, and lost the first time it sends someone to an empty location and nobody fixes the underlying data. The technology is necessary but it is not sufficient. The sufficient part is the discipline of clean data, an honest building model, and a team that believes the routes are worth following.

A final honest note on expectations. Picking optimization improves an existing manual process, and the gains, while real and worth chasing, have a ceiling set by the fact that a human still walks and still reaches. When an operation approaches that ceiling and volume keeps rising, the next step is not a better walking route, it is a different physical model, automation that reduces or removes the walk entirely. Knowing when you have squeezed the manual optimization and it is time to change the model is a judgement the software will not make for you.

8. References

• Bartholdi, J. J. and Hackman, S. T. Warehouse & Distribution Science, Supply Chain and Logistics Institute, Georgia Institute of Technology. Foundational treatment of picking, routing and travel-time modelling.
• De Koster, R., Le-Duc, T. and Roodbergen, K. J. "Design and control of warehouse order picking: a literature review." European Journal of Operational Research. Survey of batching, zoning and routing methods.
• Gu, J., Goetschalckx, M. and McGinnis, L. F. "Research on warehouse operation: a comprehensive review." European Journal of Operational Research. Overview of slotting, wave planning and operational optimization.
• Warehousing Education and Research Council (WERC). Annual DC Measures benchmarking reports on picking productivity and order accuracy.
• MHI and industry practitioner guidance on goods-to-person systems, waveless order release and automated picking. Cross-reference the warehouse automation complete guide for the broader context.

Final thoughts

Picking is the prize because picking is where the labour goes, and the labour goes into walking. AI picking optimization is, at bottom, a travel-reduction machine that works five levers at once: it batches orders so shared trips are shared, zones the building so no one walks the whole floor, routes each trip down the shortest sensible path, slots the stock so the picks are short at the source, and waves the day so work flows against the clock. Pull those levers in isolation and you get modest, disappointing gains. Pull them together with an engine that understands how they trade off, and the gains compound into the single most reliable payback in warehouse operations.

The honest framing I give clients is that the algorithm is the easy part. The value shows up when the data is clean, the building model is true, the workload is balanced, congestion is respected, and the floor trusts the routes enough to follow them. Get those right and picking optimization delivers exactly what it promises. Skip them and you join the operations with a beautiful optimization dashboard and a productivity number that never moved. Start with the data and the people, layer the optimization on top, and know when you have reached the ceiling of manual picking and it is time to change the model rather than the route.

Related reading: Warehouse automation: the complete guide, AI slotting optimization, AI route optimization, Goods-to-person systems, What is a WMS?.