Physical AI is becoming one of the biggest stories in logistics. But companies can become so focused on the most advanced layer that they overlook where AI can create value today.

This week, Agility Robotics unveiled Digit 5, its next-generation humanoid designed to work closer to people and across broader industrial workflows.

Tutor Intelligence introduced new versions of Cassie and Sonny, robots powered by foundation models and designed for tasks ranging from case picking to e-commerce fulfillment.

And Gartner published a useful framework describing four levels of AI emerging inside warehouses, from traditional optimization all the way to physical AI agents.

The direction is clear.

AI is moving from analyzing the warehouse to making decisions inside the warehouse and eventually to acting physically inside it.

But there is a trap here.

Companies can become so focused on the most advanced layer that they overlook where AI can create value today.

The AI Stack Inside the Warehouse

Gartner divides warehouse AI into four broad categories:

1. Optimization-oriented AI — forecasting, labor planning, routing, inventory optimization.
2. Operational generative AI — creating work instructions, SOPs, exception protocols and decision support.
3. Suggestive and semiautonomous agents — systems that recommend or execute multistep decisions while keeping humans involved.
4. Physical AI agents — robots that perceive, decide and physically execute warehouse tasks.

I like the progression because it separates something we often talk about as if it were one technology.

It isn’t.

Each layer solves a different problem.

And more importantly, each layer requires a different level of operational maturity.

You can deploy a sophisticated robot into a warehouse with poor inventory accuracy, weak slotting, badly configured replenishment logic and uncontrolled exceptions.

You will still have those problems.

Now you just have a sophisticated robot operating inside them.

Intelligence Before Autonomy

In my experience, some of the largest opportunities in fulfillment are not necessarily tasks that require a robot.

They are decisions.

Which SKUs should move?

Where should they be slotted?

When should replenishment happen?

How should orders be released?

How should labor move between processes?

What should happen when an order or inventory condition falls outside the standard flow?

Warehouses make thousands of these decisions every day.

Many are still driven by static rules, spreadsheets, tribal knowledge or configurations that were established years ago and rarely challenged.

That is why the lower levels of the AI stack may create disproportionate value.

Take slotting.

The physical act of moving inventory is straightforward. The intelligence problem behind it is much harder.

A good decision needs to consider SKU dimensions, velocity, demand variability, order affinity, storage media, pick path, replenishment frequency, inventory levels, ergonomics and building constraints.

AI does not need arms or legs to improve that decision.

The same applies to exception management.

A large amount of operational waste happens outside the happy path: inventory discrepancies, shorts, blocked locations, failed picks, equipment faults, orders requiring intervention.

Today, people often identify the exception, investigate it, determine the appropriate response and then initiate another workflow.

That is exactly where the transition from decision support to agentic execution becomes interesting.

The system doesn’t just tell the operator what happened.

It determines the next best action.

Eventually, it executes it.

Then Physical AI Changes the Equation

None of this diminishes what is happening in robotics.

Quite the opposite.

Agility Robotics says Digit 5 incorporates lessons from more than 65,000 hours of operation and commercial deployments. The new generation is being designed to work in closer proximity to people without the physical barriers traditionally associated with industrial automation.

Tutor Intelligence is taking another approach. Its next-generation Cassie is a mobile single-arm platform for material handling, while Sonny uses two arms and greater dexterity for fulfillment and other manipulation tasks. Tutor says the models powering these robots are trained on millions of real-world robot interactions.

This matters because traditional warehouse automation has generally worked by doing one of two things:

Constrain the environment around the machine, or constrain the task the machine performs.

Fixed automation does this extremely well.

Physical AI is attempting something fundamentally harder: allowing the machine to deal with more of the variability itself.

If that works reliably and economically, the addressable automation opportunity expands dramatically.

Brownfield facilities become more interesting.

Low-volume tasks become more interesting.

Processes that change frequently become more interesting.

And automation becomes less dependent on engineering every possible condition in advance.

But this is also why I would not start an AI strategy with the question:

“Where can we deploy humanoids?”

I would start with:

“Where are we losing the most value because our operation cannot sense, decide or act fast enough?”

Then determine what level of intelligence the problem actually requires.

Sometimes the answer will be better optimization.

Sometimes an AI copilot.

Sometimes an autonomous software agent.

And increasingly, sometimes it will be a robot.

Build the Ladder

There is another reason to think about these technologies as a progression.

Each layer creates capabilities the next one can use.

Better operational data improves optimization.

Better optimization creates better recommendations.

Better recommendations create the foundation for autonomous decisions.

And better decision-making makes physical automation more effective.

This week, GXO and Exotec announced the successful deployment of a Skypod system supporting Guess in the Netherlands: 127 robots, 60,000 rack locations, eight goods-to-person stations and reported peak capacity of up to 2,200 order lines per hour.

That is a reminder that the future warehouse will probably not be defined by one form of automation.

It will be an orchestration of conventional automation, AMRs, goods-to-person systems, intelligent software, AI agents and increasingly capable general-purpose robots.

The competitive advantage will not come from owning the most advanced robot.

It will come from knowing which level of intelligence to apply to which operational problem, and how to make all of those layers work as one system.

3 Signals Worth Watching

Tutor Intelligence pushes toward general-purpose warehouse robots

Its second-generation Cassie and Sonny platforms target multiple workflows rather than a single fixed application. More important than the form factor may be Tutor’s model: train robots from large volumes of real-world production interactions and continuously expand what the same hardware can do.

Scott Technology attacks the brownfield constraint

Its new NexPAL 200HL palletizer is designed to deliver more than 35 cases per minute while occupying a claimed 25% smaller footprint than comparable systems. The bigger signal: automation vendors are increasingly designing around existing facilities rather than assuming greenfield conditions.

Comau is building a broader intralogistics technology stack

Comau says it plans to strengthen the integration of Automha and Invent into a single technology hub capable of addressing larger automation projects across the warehouse and production value chain. Consolidation and integration across technology categories are worth watching as customers demand more complete solutions.

The question for automation leaders is changing.

Not:

Where can we use AI?

But:

What decisions and physical actions constrain this operation today, and what is the minimum level of intelligence required to remove them?

That is where I would start.