To find out how close a warehouse is to running more autonomously, count how often someone has to step in during an ordinary shift. The count is rarely zero, even in heavily automated operations.
An operator sees a station beginning to back up before a dashboard catches it. A product arrives in packaging the workflow was not designed to handle. Equipment begins showing signs of trouble earlier than expected.
Each time, someone notices. They make a judgment call and keep the operation moving. I think of these moments as an operation’s exception list. The exception list is every point where the system has to stop and ask a person what happens next.
That list says more about how close a warehouse is to autonomy than the number of robots on the floor or AI tools in the technology stack. Getting closer to autonomy means working that list down, one exception at a time.
Automation removes tasks, but autonomy removes exceptions
Automation has transformed fulfillment by taking predefined work and executing it reliably, repeatedly, and at scale. Autonomy asks for something harder: A system that can keep operating when reality stops following the plan.
Warehouse conditions change constantly. Products come in different sizes. Packaging changes. Full cases and individual items may need to move through the same operation. Equipment behaves differently under different loads.
Whenever that variability forces a handoff to a person, an exception stays on the list. For practical purposes, those exceptions fall into four categories.
1. Physical: The system cannot handle a particular item or task, so a person has to take over.
2. Connective: Two processes are automated separately but still require human intervention between them.
3. Informational: The data needed for a decision exists, but the system is not yet turning it into one.
4. Judgment: The situation genuinely calls for human experience and accountability.
The first three are where technology can systematically reduce intervention. The fourth is where human judgment continues to add value.
Sorting interventions this way turns “more autonomy” from a vague ambition into a practical list a team can work through.
The trouble often lives between systems
Plenty of warehouses with heavy automation still depend on people to keep the overall operation connected. One system stores inventory. Another handles picking. Another monitors equipment health. Each can perform well within its own boundaries while people spend time managing what happens between them.
So a more useful question is: Which exception can the system take on next? Sometimes the answer is physical, such as handling a wider range of products without introducing another manual process.
Sometimes it’s connective, such as bringing case handling and individual-item flows together more effectively. Sometimes it’s informational, such as acting on early signs of equipment wear rather than waiting for a fixed maintenance interval.
The point is to reduce the number of situations in which the operation has to stop and wait for someone to intervene.
Intelligence that shortens the list
The same test should apply to AI. Before a warehouse even opens, some exceptions are created—capacity, layout, and configuration—and all depend on assumptions about future demand. If those assumptions prove wrong, the result can be a workaround that follows the operation for years.
Scenario planning becomes a practical use of intelligence. Evaluating more possibilities before committing capital can reduce the number of problems that must be managed later.
After go-live, the opportunity changes. A warehouse produces a steady stream of signals about movement, throughput, equipment behavior, and inventory activity. Historically, much of that data has been used to explain what already happened.
The bigger opportunity is using it to help decide what happens next. If operating data shows that equipment needs attention earlier than expected, intelligence can turn a potential interruption into a planned action.
If throughput begins to drift, the useful question is not only what happened, but what the operation should do about it. That gives intelligence a clear test in physical operations: Does it help the operation make a better decision, remove an intervention, or simply move the work somewhere else?
The people closest to the operation know the list best
Warehouse workers often know the exceptions before anyone else, including when processes become fragile during peak season, which packaging changes create trouble, and which alarms signal a real problem rather than routine noise.
Their workarounds reveal where the system still falls short, and that knowledge should shape what gets improved next. As systems take on more routine decisions, people can spend more time supervising, improving the operation, and handling situations that genuinely require judgment.
Trust develops the same way. Nobody responsible for a critical operation hands a system another decision simply because the technology sounds impressive. The system earns each new responsibility by handling the previous one reliably. That is why autonomy advances one decision at a time.
Lights-out is a shrinking exception list
“Lights-out” can make autonomy sound binary, as though a warehouse either needs people or it does not.
Real operations will move in steps. One workflow becomes autonomous before another, and one exception category disappears. One maintenance decision becomes predictive, then a planning decision becomes intelligence-assisted.
The exception list gets shorter. The most useful question an operations leader can ask is a simple one: Where does our system still have to stop and ask a person what happens next?
Then ask the follow-up: What would it need to know, decide, and do for that intervention to disappear?
Ask those two questions often enough, and you start building an operation that can handle more variability, make more decisions on its own, and keep moving when reality stops following the plan.
Parth Joshi is chief product officer and president at AutoStore.