Industry

AI Is Rewiring the Supply Chain. Where Does the WMS Fit?

Every WMS vendor deck I've sat through this year has the word "agentic" somewhere around slide three. I'm not a neutral reader of those decks. I've built a warehouse management system from MVP to production scale, and at Apollo Supply Chain I'm now building VisionWare+, an analytics platform that sits on top of warehouse data and uses AI to make sense of it. So I get asked the same question from both directions: is the WMS about to be replaced by AI, or is AI just another module the vendors will bolt on? After a few months of reading and building, I think neither framing is quite right, and the gap between them is where the interesting decisions are.

What's Actually Happening on the Floor

Start with the part that's undeniable. Amazon announced in mid-2025 that it had deployed its one-millionth warehouse robot, alongside a foundation model called DeepFleet that coordinates the fleet and improves its travel time by 10%^[1]. Ten percent doesn't sound like much until you notice where it came from. Not better robots. Better decisions about where the existing ones go. That's a software gain, and it doesn't care whose logo is on the hardware.

The packaged-software vendors are moving in the same direction, from the other end. In January, Manhattan Associates made a set of AI agents commercially available across its Manhattan Active products, including a Wave Coordinator agent, a Labour agent and a Dock agent, plus an Agent Foundry that lets customers build their own in plain language^[2]. Look at which jobs those agents took. Wave planning, labor deployment, dock scheduling: that's the core of what a WMS supervisor screen has done for twenty years, only now there's something that can act instead of just display. Gartner's long-range view is that by 2030, half of cross-functional supply chain management solutions will use intelligent agents to execute decisions autonomously^[3].

The Part the Keynotes Leave Out

Gartner also said something last year that gets quoted a lot less. It expects more than 40% of agentic AI projects to be cancelled by the end of 2027, mostly over cost, unclear value and weak risk controls, and it reckons only around 130 of the thousands of vendors claiming agentic products are building anything that deserves the label^[4]. Its name for the rest is "agent washing," which is accurate and a little brutal. Read any agent announcement with that in mind, including the ones I just cited.

The less glamorous constraint is data. Impinj surveyed 1,000 US supply chain managers at the end of 2024: 91% said they were equipped to deliver accurate supply chain visibility, but only 33% said they consistently get accurate, real-time inventory data^[5]. Impinj sells RFID, so it has a reason to go looking for that gap, and I'd discount the exact figures accordingly. The shape of it matches what I see, though. An agent that acts on a wrong on-hand quantity doesn't make a smaller mistake than a person would. It makes the same mistake faster, at three in the morning, with nobody watching.

What the WMS Is Actually For

When people say AI will replace the WMS, I think they're picturing the screens. The screens were never the point. Say an agent decides to release wave 14. Something has to confirm the allocation still holds, that the lot isn't expired or on quality hold, that the pick locations really contain what the agent believes they do, and that the carrier cutoff is still reachable. It has to do all of that atomically, while three other things compete for the same inventory. That's a transactional referee, and it's the real job of a WMS.

I wrote last month about hexagonal architecture, with the business rules in the middle and everything else as adapters around them. That's how I've come to think about this too. RF scanners, robots, carrier APIs and now AI agents are all callers of the same rules, and the WMS is the thing that says no when a caller, human or not, asks for something the warehouse can't honor. I'd go further: that's what makes it safe to let an agent act at all. An agent should get the same permissions and limits a supervisor would, and the system enforcing those limits is the one that already owns the inventory. To be clear, that's how I'd design it. It isn't a description of what every vendor ships.

The WMS isn't competing with the agent. It's the thing that decides whether the agent's request is allowed to become a transaction.

Where the Traditional WMS Is Exposed

None of that means the incumbents are safe. A few things worry me, and they're the ones I keep tripping over when building around warehouse systems.

Wave and batch thinking is baked in deep. Most WMS designs assume work is planned on a schedule and released in chunks, which made sense when execution meant paper and RF guns. A fleet of robots reacting to every new order doesn't want a wave at 2 p.m. It wants a continuous stream and a system that re-plans as things change. The vendors know this. The architectures underneath are slower to move than the marketing.

Integration surface is the second problem. An agent, or a robot controller, needs fine-grained commands that are safe to retry and fast enough to sit inside a decision loop. A lot of WMS integration is still bulk file drops and polling, built for ERP handshakes. Configuration is the third. Vendors are now pitching natural-language setup, where you describe a receiving process and the system generates the configuration^[2]. That's useful, and I'd still want the generated configuration to go through the same review a pull request does, because a bad receiving rule is expensive in ways a typo in a README isn't.

The bigger structural risk sits above the WMS, not beside it. A layer of execution and orchestration software is growing between the WMS and the equipment, deciding how work gets prioritized across people, robots and exceptions. One analyst put it this way: the winners won't be the operations with the most robots, but the ones with the best decision layer. He also noted that most of the performance claims in the space are vendor-reported^[6]. If the decision layer owns real-time optimization, the WMS gets demoted to the ledger underneath. Ledgers matter, and nobody rips one out lightly. They're also lower margin and less strategic, and that demotion, not replacement, is what I think the incumbents should actually worry about.

What It Takes to Stay Relevant

If I were running product for a WMS, this is the list I'd work from.

Publish every state change as an event, not just a row someone has to poll for. Agents and analytics both want to react to inventory moving, not discover it later. Make commands safe to retry, and make the caller's identity part of every request, so the system can tell a supervisor from an agent and apply different limits to each. Keep an audit trail that ties a decision to the transaction it caused, with a correlation ID that survives from the agent's request all the way to the inventory movement. The first time an autonomous action goes wrong, someone will ask exactly why, and "the model decided" won't be an answer.

Then treat inventory accuracy as a product feature instead of a customer problem. Cycle counting that schedules itself, exception queues that actually get worked, discrepancy detection that doesn't wait for the annual audit. Everything an AI layer does downstream is only as good as that number. And stop presenting the wave planner as the product. It's a feature, and it's about to get a lot of competition.

What I Don't Know

I should be straight about the limits of all this. Most of the performance numbers in this space come from vendors, and I haven't seen independent measurement of agent-driven warehouse gains at scale. I don't know whether Gartner's 2030 number lands, and I suspect adoption will look very different in a 3PL running thin margins on shared infrastructure than in a manufacturer with one well-understood network. What I can say from our own work on VisionWare+ is that the model was the easier half. Getting clean, timely warehouse data into it was where most of the effort went.

None of that makes me less interested in where this goes. The work I'd want my own team to be good at in 2027 is the unglamorous middle: integration, data quality, and guardrails around autonomous actions. So the question I'd put to any WMS vendor now isn't whether the product has AI. It's what happens when something that isn't a person calls your API at three in the morning and is wrong.

← Back to all posts