More AI won't fix your warehouse. The right AI will.

Warehouse automation has spent three decades getting very good at one thing: execution. Goods move faster and more reliably than any manual operation could manage, and they're stored far more densely. The pressure to keep improving isn't letting up. Consumers want delivery sooner, labor is harder to find and keep, space costs more every year, and SKU complexity keeps growing. 

But execution has a ceiling. Even the most advanced automated system does exactly what it was designed to do; it doesn't ask whether that's still the right thing to do. When an unusual pattern shows up, it waits for a human to notice it, interpret it, and decide.

The obvious next step is to add intelligence, so the system can spot those patterns and act on them itself. Most assume that means reaching for the latest general-purpose AI models. I'd argue the opposite: the systems that actually make the difference won't be the broadest. They'll be narrow agents that understand one domain deeply and are grounded in the physical reality of how a warehouse really moves. 

Automation assumes stability. Autonomy expects disruption. 

It helps to be precise about two words that get blurred together. Automation executes predefined tasks reliably and at scale: It follows the rules it was given, every time. Autonomy goes further: It senses when conditions have changed and adjusts in real time, with people setting the guardrails. That difference matters more than it used to. Warehouses now run in far less predictable conditions, and that should change what we ask of our systems. 

Gartner draws the same line. It urges organizations to move from an automation mindset, breaking work into parts and optimizing each one, to an autonomous mindset built on interconnected, real-time decisions under human oversight. As Gartner VP Analyst Alan O'Keeffe put it at the Supply Chain Symposium/Xpo: "The automation mindset works when the world is stable. The autonomous mindset works because it isn't." Gartner finds leaders feel the gap too: 80% expect autonomous business to be the dominant model by 2030, and 77% already admit their current systems and strategies aren't ready for it.

In a volatile environment, the obvious failures are rarely what cost the most. The expensive problems start as small signals that are easy to miss, like a recurring bottleneck, a shift in order patterns, a robot charging oddly under load or a component wearing faster than it should. The operations that catch those in time are the ones that stay ahead. 

The real bottleneck is expertise 

Hardware and software are no longer limiting factors. The bottleneck is expertise, and the trusted data it runs on. 

Autonomous systems break down the moment decisions still have to be reconciled by hand. Yet Zero100's 2026 Data Readiness Survey found that 44% of organizations still resolve data conflicts in meetings rather than in systems (and some problems don't get resolved at all). If conflicting numbers have to be debated, escalated and cross-checked before anyone acts, you're still running at human decision speed. That's workable when conditions are stable, but not when they change by the hour.

There's a common belief that AI will sort this out on its own: feed it enough data and it will find patterns no person could. In my experience, it doesn't work like that. A model can only work with what it's given. If the data is badly structured, or nobody has defined how one signal relates to another, the model has no way of knowing. It won't tell you something is missing. It will still give you an answer, and that answer will sound just as sure as a correct one.

So "better data" was always the wrong phrase. Everyone in supply chain says it, and it has stopped meaning much. What a model actually needs is data that's reliable, structured consistently, and clear about what each signal represents, before the model ever touches it. In an autonomous operation, that data drives action across the system, so mistakes don't stay on a dashboard. If an agent is going to flag an issue, explain it and recommend a fix, operators have to trust it's reading the system correctly. Otherwise it's just one more opinion to verify. 

Autonomy starts with observability

General AI answers questions. Specialized AI delivers expertise, but only if it's grounded in what's physically happening on the floor. A warehouse isn't a clean digital environment. Software decisions move real robots, inventory and people, and a physical system is far less forgiving than a digital one. You can roll back a bad software update. You can't roll back a congested grid or a truck that left without the orders it was waiting for. 

So autonomy starts with observability. That means continuously reading how the system is behaving, so you can see where performance is drifting, where constraints are building up and which anomalies actually matter.

Take charging. On paper, a fleet of grid robots tops up in idle moments, and you never think about it, because normally there are idle moments to spare. But under sustained peak, say a promotional spike or a compressed cut-off window, the idle time disappears. Robots keep moving instead of charging, and their batteries drift lower shift after shift. A general dashboard still shows availability green. A narrow charging agent sees the pattern underneath: batteries bottoming out lower each shift, robots that aren't ready when they're needed, and more of them breaking off mid-job to charge. Left unchecked, output drops in exactly the week you need it most. 

On the surface it looks like a battery problem. Really, it's about coordination. Charging has to be planned as part of the flow, and you can't plan it if you don't know how intense the peak will be. If you only react once batteries are already low, you're stuck choosing between two bad options. You either pull robots offline to charge and lose capacity at peak, or keep them running and drain the batteries deeper than they should go, which shortens their life. A narrow agent sees this coming and plans charging ahead, so the fleet keeps moving without wearing the batteries out. 

This is why grounding matters more than scale. An agent that doesn't understand how the warehouse physically moves, where it gets congested, when the charging and service windows fall, won't produce sharper intelligence. It'll produce the wrong intelligence faster.

It's also why a bigger model isn't the answer, even though that's the first instinct. Point one general model at the whole operation, from inventory to throughput to robot health, and it learns each part more slowly, because every domain signal is just noise to the others. It also has a harder time spotting the connections that matter most, how one part of the operation affects another, because they get lost in everything else it's tracking. 

A narrow agent sees only what matters to its job, so it learns fast, runs cheap, and can be corrected without disturbing everything around it. 

From experiments to operational value

The mood around AI spend has flipped. A year ago, it was "build and spend". Now leaders are saying no and asking what the business value actually is.

Fair question, because the gap between ambition and impact is wide. Deloitte found that 74% of organizations want AI to drive revenue growth, but only 20% have got there. MIT's research points to the cause: most systems don't retain feedback, adapt to context, or plug into how work actually gets done.

That's exactly where narrow beats broad. A broad AI initiative is hard to measure, because it touches so many things that it's difficult to say what it actually improved. A specialized agent is easy to pin down. An uptime agent is measured against availability, a charging agent against robot readiness and energy efficiency. You know within a quarter whether it earned its place. Within a quarter, you know whether it's paying off. 

Narrow agents don't try to transform the whole operation at once. Each one owns a single problem and keeps getting better at it, and those gains compound into a system that improves itself over time. When disruption is the norm and every AI dollar has to prove itself, the intelligence that matters most is the one closest to the operation.