In 2026, nearly every warehouse software vendor is talking about AI agents. In the demo videos, the agent behaves like a tireless dispatcher: orders arrive and it schedules waves; stock runs low and it raises replenishment orders; a driver runs late and it reschedules the dock appointment. Watch a demo and it is easy to conclude that the era of autonomous warehouse execution has already arrived. But between that impression and reality sit several gaps that have not yet been crossed. This article tries to do something simple: take the public data, open up those gaps, and count how many steps remain between ai agents in the warehouse that answer questions and ones that act on their own.

Start With the Data: Real Excitement, Uncertain Landing

According to Gartner's June 2025 forecast, over 40% of agentic AI projects will be canceled by the end of 2027, with escalating costs, unclear business value, and inadequate risk controls as the main causes. This is a prediction, not an accomplished fact.

The same forecast contains another set of figures worth placing alongside it. Gartner expects that by 2028, at least 15% of day-to-day work decisions will be made autonomously through agentic AI, up from close to zero today; and that by 2030, 60% of enterprises using supply chain management software will have adopted agentic AI features, up from 5% in 2025. Industry analysis citing Gartner adds that by the end of 2026, 40% of enterprise applications will integrate task-specific AI agents, up from less than 5% in 2025.

Adoption data from the warehouse side shows the same gap. According to a 2025 study by MIT and Mecalux covering more than 2,000 supply chain professionals, over 60% of warehouses have implemented some form of AI or machine learning, and nearly 90% operate at automation levels beyond basic. But "using AI" and "letting an agent execute autonomously" are two different propositions. The first mostly means prediction, recognition, and recommendation. The second requires a system to rewrite real work instructions without anyone approving each step. The distance between them is exactly the few steps this article is about.

Real deployments today cluster at the coordination layer, not at the execution core. The partnership between DHL Supply Chain and HappyRobot is a representative case: AI agents handle appointment scheduling, driver follow-up calls, and high-priority warehouse coordination by phone and email — repetitive, rule-bound coordination work. UPS's ORION system analyzes more than one billion data points daily to orchestrate parcel delivery routes, again a case of computing an optimal answer and executing a pre-defined action. Amazon integrates AI agents directly into fulfillment center operations, managing inventory positions, shelf space optimization, order picking, and robotics coordination. These cases prove agents can already work where rules are explicit and goals are measurable. They do not prove an agent can decide alone where rules are ambiguous and stakes are high.

A Maturity Ladder for AI Agents in the Warehouse: L1 to L4

To judge "how many steps remain," autonomy first has to be broken into degrees. Drawing on commonly used grading approaches in the industry, here is a four-level framework:

Level Form What the agent can do Where the human sits Warehouse example
L1 Question answering Answer questions, look up data, generate reports Human asks, human decides "What was last week's pick error rate?"
L2 Advisor Analyze data, present ranked recommendations Human chooses, human executes Three wave plans with pros and cons
L3 Execution with confirmation Draft actions, execute only after human approval Human is the approval gate Replenishment order generated, supervisor confirms before release
L4 Autonomous execution with safeguards Execute within rules, escalate exceptions automatically Human supervises, handles exceptions only Night waves scheduled automatically, over-threshold orders routed to staff

L1 to L2 is a capability step: the quality of recommendations depends on the quality of the data. L2 to L3 is a trust step: whether people will hand over machine-drafted actions depends on how accurate past recommendations were. L3 to L4 is a responsibility step: autonomous execution means someone bears the cost of an error, and that someone can only be a person today. Most warehouses sit between L1 and L2 now, a few pilots have touched L3, and L4 exists only in narrow scenarios where rules are extremely clear and risks are very low.

Four Gaps Before AI Agents in the Warehouse Can Act Alone

The first gap is technical maturity. For an agent to execute autonomously, it must do three things at once: understand the goal, call the tools, and verify the result. The goal is understood through the model, the tools are called through system interfaces, and the result is verified through read-back and checks. The weakest link of the three is interfaces: many warehouses run WMS, TMS, and equipment control systems with no unified machine-callable interface between them, so an agent that wants to change a wave must cross three or four data silos first. Opendock's MCP server, live since June 2026, lets any MCP-compatible AI agent schedule, modify, cancel, and query dock appointments directly — a directional breakthrough, but one that covers dock appointments, not the execution core inside the warehouse.

The second gap is permissions and safeguards. An agent executing autonomously is spending money and rewriting plans on someone's behalf: one wrong automatic replenishment can mean a truckload of dead stock; one wrong wave reschedule can throw a whole night shift's picking rhythm off. Safeguards must be built before autonomy is granted: amount caps, quantity caps, time windows, blacklisted SKUs, automatic rollback. Autonomy without safeguards is not innovation; it is loss of control.

The third gap is the human supervisor's seat. L4 does not remove people from the warehouse; it moves them to supervision: watching dashboards, handling exceptions, periodically auditing the agent's decision logs. But supervision itself demands new capability: a supervisor must be able to see why the agent scheduled a wave this way, to stop it after three bad calls in a row, and to tell "the agent was wrong" apart from "the data was wrong." According to Gartner's survey of 782 infrastructure and operations leaders in late 2025, among those reporting setbacks, 38% cited poor data quality or limited data availability as a direct cause — the first thing a supervisor usually has to supervise is not the agent but the data fed to it.

The fourth gap is the cost ledger. An agent's cost is not just a one-time implementation fee; it includes ongoing model calls, interface maintenance, supervision labor, and safeguard reserves. Consider the math (illustrative example): if a mid-sized warehouse lets an agent take over night wave scheduling, monthly model and interface costs run about 8,000 yuan, while the night scheduling labor saved is roughly 1.5 hours a day; at 40 yuan an hour, that saves about 1,800 yuan a month — a loss on paper. Only when the same agent also handles replenishment advice, exception orders, and appointment coordination can the cost spread across enough tasks to break even. A single-point agent rarely pays for itself; a grouped agent might.

A Timeline Judgment: My View

Putting it together, my judgment runs as follows. In 2026 and 2027, L2 advisory ai agents in the warehouse will become standard in warehouse software, and L3 "execution with confirmation" will enter pilots in three scenarios — replenishment, appointments, and exception orders. Before the end of 2027, the 40% cancellation rate Gartner predicted will largely materialize, and the survivors will mostly be projects that narrowed their scope and built real safeguards. By 2028, the 15% autonomous-decision figure is more likely to be reached at the coordination layer than at the execution core — appointments, rescheduling, chasing drivers, exception notifications — rather than automatic rewriting of pick plans. Genuine L4 autonomous execution, I believe, will only become routine in narrow, rule-clear, risk-controlled scenarios around 2030, which roughly matches Gartner's prediction that 60% of enterprises using supply chain software will have adopted agentic AI features by then.

This judgment leaves one open question: if machine-callable interfaces between WMS and equipment systems become broadly standardized before 2027, the first gap will collapse faster than expected and the whole timeline shifts earlier; conversely, if data quality problems keep dragging, even the spread of L3 will be delayed. The story of AI agents in the warehouse is ultimately not about how clever the models are, but about whether the warehouse's foundation — data, interfaces, safeguard rules — is ready. If the foundation is not ready, then no matter how clever the agent is, it can only stand at the question-answering post and watch the warehouse work.

By Dr. Chen, Supply Chain Researcher