Gartner: warehouse AI shifts from optimisation to Physical AI on the floor

The trend Gartner calls enhanced optimisation is the least sexy but probably the safest place to put a dollar. It is demand forecasting and slotting that adapt in real time instead of being recalculated once a quarter by someone in a back office. Think about what that means on a floor: the system watches order patterns and physically reshuffles where fast-movers sit so pickers walk less, and it learns from every order that goes out the door rather than waiting for the next planning cycle. Most DCs I know run slotting reviews maybe twice a year, and they are always a little stale by month three, so pickers are quietly burning steps on a layout that made sense in February but not in October when the catalogue changed. Real-time adaptation closes that gap, and what I like about it is it does not require a single robot to be installed — you are getting paid on software you mostly already own if your WMS is any good. For an importer riding promotional spikes, this is the bit that quietly pays rent: the system pre-positions stock before the campaign lands instead of after the complaints roll in.

Operational generative AI is the one that caught my attention, because it is pure back-office leverage that pays back fast and does not touch the floor's physics at all. Gartner says these tools write SOPs, work instructions, and exception playbooks. If you have ever watched a warehouse manager spend a weekend rewriting a picking procedure after a process change, you know how much dead time that is, and how often it just does not get done and the team runs on tribal memory instead. A genAI tool drafts that in an afternoon and the supervisor edits and signs. The catch, and it is a real one, is the output still has to be checked by someone who actually knows the floor, because a confident-sounding SOP that is wrong will get someone hurt or will ship the wrong carton to a customer who will not forgive it. Treat the tool as a fast junior writer, not as the authority, and you get the speed without the risk. I would also use it to translate a good procedure across three sites that currently each do the same task three different ways, because consistency is where a lot of avoidable errors actually come from.

Then there is the human-in-the-loop suggestive agent, and this is the piece I would tell a nervous operations director to look at before anything else, because it does not take people off the floor — it makes the people you already have smarter. The agent recommends a next step or partly runs a multi-step task, but a human stays in oversight and can override at any moment. For a shipper weighing whether any of this is real, that is the gentlest entry point: you are not betting the whole labour model, you are giving your team a co-pilot that suggests while you keep the authority, and if it is wrong the human catches it before it costs you. I have seen pilots of this kind where pickers actually liked the prompts because it took the guesswork out of a messy exception, and that is the unlock — adoption follows usefulness, not mandates from head office. Picture an exception where a shipment is short: the agent pulls the order history, proposes a substitute location, and the supervisor confirms in two taps instead of a ten-minute hunt.

The robots come in what Gartner labels Physical AI, and it is the part that scares the cheque-signers for good reason. This pairs robotics with sensing to actually handle picking, packing, sorting, and material handling, the work that has always been human hands. That is where capex gets serious and where vendor demos look magical and real deployment looks like a six-month argument with your IT department about network coverage and edge compute. I am not saying do not do it — I am saying do not start here, because if your data underneath is shaky the robot will just move the wrong boxes faster, and now you have automated the mistake. The robots worth talking about are autonomous mobile units that ferry totes, and goods-to-person stations that bring the shelf to the picker; both are mature, but both assume the address and the quantity in the system are true, which brings us back to the data problem every single time.

Now let me put some numbers on the table, because AI will help is not a business case and I refuse to hand you a recommendation I cannot cost. Picture a fairly ordinary distribution centre: ten thousand square metres, a few hundred staff on shift, and let us say each picker does about two hundred picks an hour. That is a reasonable mid-size operation for an importer or a 3PL moving consumer or industrial goods. If you layer AI-driven slotting on top and it lifts throughput by ten percent — and Gartner's enhanced optimisation claim is exactly that kind of gain if your data is clean — you are now getting the equivalent of twenty extra picks per worker per hour without adding a body. Across a hundred pickers on a single shift, that is two thousand extra picks an hour, roughly sixteen thousand in an eight-hour day. At a loaded cost of, say, four dollars a pick including labour and overhead, that ten percent is worth somewhere around sixty-four thousand dollars a day in avoided incremental labour, or call it north of fifteen million dollars a year if you run multiple shifts. I am not promising you will bank all of it — some of it just absorbs volume growth you would otherwise have staffed for — but the magnitude is real and it is the kind of number that gets a CFO to return your call. If the gain comes in at only five percent, you are still at about thirty-two thousand dollars a day, which is a pilot that pays for itself before the paint dries.

The generative AI piece is easier to quantify and faster to feel, which is why I would actually start there. Writing a proper SOP with work instructions and an exception playbook for a new process used to eat two days of a competent supervisor's time — research, drafting, review, sign-off. Gartner's framing says genAI collapses that to roughly two hours of drafting plus a review pass. Take an operation that authors, say, ten SOPs a month. That is eighteen days of supervisor time freed up per month, better than three weeks of a skilled person you are already paying. Even if you only recover half of that to actual value, you have bought back the better part of a salary without touching headcount, and you have made the docs get written at all instead of living in someone's head. New hires also get productive a week sooner when the onboarding procedure is actually on paper and in their language, and that is a number nobody puts in the business case but every DC manager feels in Q4.

So where should an operator actually start, and who owns it? My advice, drawn from watching what sticks and what gets shelved, is to pilot the boring stuff up front and let the proof earn the later spend. Run a ninety-day pilot on labour forecasting and slotting optimisation using the data you already have, scoped to a single zone so failure is cheap and success is visible. Name a single owner — usually the DC manager or the head of operations — and give them a target: prove a measurable throughput or labour-hour gain on that one zone before any capital gets committed to robotics. Build the data foundation in parallel: clean SKU masters, accurate on-hand counts, and a WMS that actually exports event logs someone can use. Most Physical AI failures I have seen were not robot problems, they were the system thought we had stock we didn't problems, and no amount of clever hardware fixes a lie in the database. Track three metrics only during the pilot — pick lines per labour hour, slotting-change cycle time, and SOP authoring hours — because a dashboard with forty KPIs is how a good idea goes to die in committee.

On timing I would set the pilot to land its readout at ninety days and a go/no-go on robotics spend at day one hundred twenty, and I would write both dates into the charter so they do not slide. The person to bring in early is not a robot vendor, it is your WMS provider and your industrial-engineer type, because the leverage is in the model before the metal, and those conversations cost you a plane ticket, not a capital request. If the pilot shows even a five percent gain on a single zone, that funds the next phase and gives you a real number to take to the board. If it shows nothing, you have spent almost nothing and learned your data is the real blocker, which is a far cheaper lesson than a cancelled automation line that already signed a lease.

Let me also be honest about the roads you do not take, because doing nothing and doing everything are both live options and I do not want you thinking the only choice is between magic and bankruptcy. The alternative to all of this is to keep running pure manual — and for a lot of smaller DCs that is still defensible, especially if your volumes are lumpy, your margins are thin, and you can flex labour up and down with the order book. The other extreme is full automation from day one, and I have watched that bankrupt the patience of otherwise healthy companies because the integration tax is brutal and the payback stretches past the attention span of the board that approved it. The Gartner framing is useful precisely because it gives you a ramp instead of a cliff: optimisation and genAI are cheap and quick, human-in-the-loop is a middle step, Physical AI is the heaviest move, and you can stop at any rung and still be better than you were.

The pitfalls are where I would spend your worry budget, because the technology is the easy part and the organisation is the hard part. Dirty data is the silent killer — every one of these trends assumes your WMS knows what is true, and most do not, so the opening project should be a data cleanse, not a robot. Over-promising vendors are the next risk; the demo always works, the reference site always has a PhD on staff, and your site always has neither, so insist on a paid pilot in your own four walls before you sign. Change management is the one after that: a suggestive agent only helps if pickers trust it and supervisors use it, and that takes training and a few early wins, not a memo from head office. And on Physical AI specifically, safety is non-negotiable — robots and humans sharing a floor is a real hazard that needs proper guarding, defined zones, and certified emergency stops, not a hopeful sticker on the cage and a prayer. I have seen a site where the union walked the floor the day a robot arrived and nothing moved for a week; bring the workforce into the design early and that week becomes an afternoon.

One thing I want to say directly to the importer or 3PL reading this, because your situation is a little different from a standalone retailer. Your distribution centre is a node in a chain that starts at a factory gate and ends at a customer's door, and every hour of delay or every mis-pick ricochets into a late container, a penalty, or a lost account. That is exactly why the unglamorous optimisation and genAI moves matter more to you than to anyone — they compress the slack without asking you to rebuild the building. When a promo lands or a ship comes in a week early, real-time slotting is what keeps the floor from melting, and a genAI-written exception playbook is what stops a green junior from freezing when a hold goes wrong at 2 a.m. You do not need robots to protect that margin; you need the model underneath to be honest, and that is the cheapest insurance on this whole list.

Let me also say a word about how to measure this honestly so you do not fool yourself the way I have seen boards fooled. Do not count the headline ten percent as pure saving on day one; count what you actually avoided spending — the temp agency you did not call, the Saturday overtime you cancelled, the extra shift you deferred. Those are the lines a CFO believes, and they are usually enough to fund the next step without a heroic business case. Keep the pilot's success definition in the charter before you start, not after the numbers are in, because a target written after the fact is just a story you tell yourself.

For the data foundation, concretely, the three things that break Physical AI are a SKU master with duplicate or mislabeled entries, an on-hand count that drifts because cycle counting is sporadic, and a WMS that stores events but cannot export them in a shape a model can eat. You can fix all three with a modest project and no capital equipment, and doing so is what turns the later robot business from a gamble into a line item. I tell clients to spend the opening thirty days of the pilot just getting on-hand accuracy above 98 percent in the chosen zone; everything downstream is easier once that number is honest, and the robotics conversation stops being a leap of faith.

I will land it like this. The smart play for an importer or 3PL in 2026 is not to chase the robot video that is circulating on LinkedIn. It is to clean your data, free your supervisors from two-day SOP marathons with genAI, and prove a slotting gain on one zone in ninety days. If that works, the robots become a reasoned next step instead of an expensive guess, and you will walk into the vendor meeting knowing exactly what problem you are solving. If it does not, you have learned the truth cheaply and nobody got hurt. Either way you are ahead of the operator who bought the metal up front and is now arguing with their WMS about inventory that does not exist, and that operator is more common than the vendors will ever admit.

• Launch a 90-day pilot on labour forecasting and slotting optimisation in one DC zone, owner: DC manager, readout by 2026-12-31.
• Cut SOP and work-instruction authoring from 2 days to under 2 hours using genAI on 10 SOPs per month, validated by ops supervisor by Q1 2027.
• Stand up a clean-data foundation: reconcile SKU master and on-hand accuracy to 99 percent before any robotics spend is committed.
• Set a hard go/no-go gate at day 120 — release robotics capital only if the pilot shows at least a 5 percent throughput or labour-hour gain.
• Run a Physical AI safety review covering guarding, zones and emergency stops with a certified integrator before any robot enters the floor.