The four levels of warehouse AI: where your operation sits, and which jump actually hurts
Almost no operation decides to “adopt AI”. It adopts AI in pieces, without announcing it: a slotting engine here, a replenishment suggestion there, an alert that now arrives with the fix already proposed. This month Gartner sorted that mess into four distinct levels — and lining them up surfaces the question almost nobody asked: how many decisions is your system taking today without asking you?
The question is no longer whether you use AI
For three years the conversation was binary: does the operation use artificial intelligence or not? That question no longer separates anyone from anyone. Any modern WMS ships with optimization algorithms inside, and any planner runs models that five years ago would have been called AI without argument.
The question that does separate is a different one: how much decision-making have you delegated. Because between “the system recommends where to place this SKU” and “the system diverted the order to another node and told me afterwards” there is an operational chasm — even though both are sold under the same label.
The four levels Gartner describes are not a maturity ladder you have to climb end to end. They are four different kinds of relationship between your team and the system, and each one demands a different kind of control.
Most operations sit at level 1 with loose pieces of level 2, and arrive at level 3 without having decided to: they switch on a feature that shipped enabled.
Level 1 — Optimization: the AI you already have and don’t call AI
Algorithms that process real-time data to solve narrow, well-defined problems: demand forecasting, shift labor planning, route optimization, inventory placement. They produce a recommendation, show why, and a person confirms or discards it.
It is the least glamorous level and the one that pays back first. Slotting recalculated against the actual rotation of the last eight weeks cuts metres walked per line picked from day one, with no debate about governance or the future of work.
Transparency is the defining feature of this level: the system can explain its recommendation in terms a warehouse manager understands and can argue with. If it can’t, it isn’t level 1 — it’s a black box that hasn’t caused you trouble yet.
If your operation hasn’t closed level 1, jumping to agents is building the second floor before the first. Gartner’s recommendation is explicit: start with the proven cases — labor forecasting, slotting — before opening the door to agents.
Level 2 — Operational generative AI: reading what nobody structured
Every logistics operation generates a mountain of information no system processes, because it isn’t shaped like a table: the carrier’s email flagging a delay, the incident a picker scrawled on the delivery note, the photo of the damaged pallet, the footnote on a purchase order with a special condition.
Level 2 is models that read all of that and return something actionable: a work instruction written in plain language, a shift summary, a classification that routes the incident into the right circuit. They decide nothing. They translate.
It’s the level the team notices fastest in daily life, because it attacks the work nobody enjoys and which therefore gets done late, badly or never. And it’s the cheapest to try: it doesn’t touch execution, so an error costs a confusing moment, not a misshipped order.
Practical rule for level 2: if the worst possible error is “somebody read something confusing and asked”, you’re on safe ground to experiment.
Level 3 — Agents that suggest and execute: here’s the jump
Semiautonomous agents read the situation, build a multistep plan and execute part of it on their own. They reassign a picking wave when a node is saturating. They divert an order to another distribution centre. They trigger an early replenishment. They reassign transport when the projected SLA doesn’t close. The human is left with the exception and the oversight.
This is the jump that changes the nature of the problem. At levels 1 and 2 the management question is “does the model work well?”. At level 3 it becomes “who answers for it when it’s wrong?” — and that isn’t a technical question. It’s an organizational one.
The uncomfortable part is that many operations have already crossed that threshold without registering it. A feature that came enabled by default, an automated flow somebody configured eight months ago and nobody revisited, an integration that executes without asking for confirmation because that’s how it appeared in the documentation example.
If nobody wrote down what the system may execute alone, the line still exists. Your vendor’s default drew it, not your operational judgment.
Level 4 — Physical AI: the most visible, almost never the first
Robotics combining AI, sensors and movement to handle picking, packing, sorting and material handling. It’s the level that ends up in the photos, and the one everyone pictures when “smart warehouse” comes up.
It’s also the one with the largest investment, the longest implementation and the greatest dependence on the layout you already have. And except in very high-volume operations with a very stable mix, it is rarely the one that returns the investment fastest.
The order Gartner lists them in is not a ranking of value. It’s a ranking of increasing autonomy. For most operations in the region, the value is still concentrated in the first two.
The pattern that repeats
Watching how operations move between levels, three things show up again and again.
Levels get skipped before the previous one is closed. Agents are switched on over dirty master data and stale locations. The agent executes fast and well on wrong information, which is the worst possible combination.
Autonomy gets confused with quality. A system deciding on its own isn’t better — it’s faster. If the decision was bad when a person made it, it will now be bad more times per hour.
Nobody measures the system’s decisions. Operational KPIs get measured, but not how many times the agent decided, how many it got right and how many had to be undone. Without that record there’s no way to know whether to widen its scope or narrow it.
All three share a root cause: autonomy was enabled in practice rather than by decision. And what isn’t decided explicitly never gets reviewed either.
How to decide what to delegate: impact times reversibility
The criterion that works best is neither technical nor consultant-shaped. Take every decision the system could make alone and place it on two axes: what happens if it’s wrong (impact) and what it costs to undo (reversibility).
Reordering a picking wave is low impact and fully reversible: if the agent gets it wrong, you reorder again and twenty minutes are lost. Let it execute alone, no asking. Cancelling a supplier purchase order is high impact and hard to reverse: it proposes and waits for confirmation. Always.
The interesting quadrant is the third one: high impact but reversible. Diverting a large order to another node lives there. The right answer is neither yes nor no — it’s execute, but flag it in the moment and keep the undo one click away. That quadrant is where most operations can gain real speed without taking on real risk.
The full exercise takes an hour with the warehouse lead and the planning lead at the same table. No budget, no vendor, no project required. And it frames the discussion for the next two years.
What changes for people
Gartner surveyed 509 supply chain leaders: 55% expect agentic AI to reduce entry-level hiring needs and 51% believe it will drive overall workforce reductions. But the most useful figure is the third: 86% agree adoption will demand new ways of developing talent.
Translated into the operation: the roles that today consist of watching a screen and confirming what the system proposes are the first to change. What appears in their place is a different role — whoever defines the system’s limits, reviews the exceptions, and decides when to widen or narrow an agent’s scope.
That role doesn’t exist on any org chart yet, and it’s the one that will separate an operation that uses agents from one that suffers them.
How Kovix solves it
Kovix’s WMS covers levels 1 and 2 on the operation you already run: it optimizes slotting, waves and staffing against actual rotation, and turns unstructured input — incidents, delivery notes, carrier emails — into concrete work instructions. That’s the floor, and it’s where the fastest return lives.
Our AI agents open level 3, with one condition we set rather than the system’s default: every agent starts with its scope written down. What it may execute alone, what it proposes and holds for confirmation, and what stays on the record so it can be audited afterwards. The impact-by-reversibility map isn’t an annex to the project — it’s the first deliverable.
The approach is the usual one: without replacing what works, on top of the systems you already have, with support from setup through the first 90 days.
Next steps
Before evaluating any tool, take the inventory: write down every decision your systems already take without asking you. You’ll find more than you expected, and almost certainly one sitting in the high-impact, irreversible quadrant that nobody chose to put there.
That inventory tells you which level you’re actually on — which is almost never the level you thought you were on. Then we can talk about what’s worth delegating and what’s worth taking back.
Shall we map your level of autonomy?
30 minutes. No commitment. You leave with your operation placed across the four levels and the two decisions worth delegating now.
