If you’ve been following retail technology news lately, you’ve probably noticed that AI dominates nearly every conversation. From NRF to Shoptalk, it seems everyone is talking about artificial intelligence but are we talking about the right things?
The problem isn’t that we’re discussing AI too much. It’s that we’re discussing it too broadly, without distinguishing between the different types of AI and their specific applications. This lack of clarity makes it difficult for retail leaders to understand what’s actually worth investing in today versus what remains on the horizon.
That’s why we invited John Mabe, Product Manager at Dematic, to join us on Omni Talk for a deep dive into AI applications specifically within warehouse operations. What John revealed is that “AI” isn’t a single technology—it’s a toolkit with three distinct categories, each at different stages of maturity and deployment.
The Three Categories of Warehouse AI
1. Optimization AI: The Brain
This is the AI that makes decisions about what happens next in your operation. Think of it as the brains of your warehouse, determining the optimal order to fulfill next or how to slot inventory for maximum efficiency.
Here’s what might surprise you: optimization AI isn’t new. Warehouses have been using mathematical models for decision-making for years. What’s changed is the addition of machine learning capabilities that allow these systems to adapt over time. Through supervised learning, modern optimization AI can accurately predict SKU velocity, even for those tricky medium and slow movers. Combined with reinforcement learning, these systems can automatically adjust to handle flash sales, promotions, or seasonal shifts.
The catch? While larger retailers are leveraging optimization AI effectively, smaller players struggle with adoption. The barrier isn’t the technology itself. It’s data structure. These AI models need properly organized, high-quality data to function effectively. Many smaller operations haven’t yet invested in structuring their data in ways that AI can leverage.
2. Vision and Perception AI: The Eyes
Computer vision systems use cameras and sensors to understand the physical state of a warehouse in real time. Today, these systems are being deployed for narrow but valuable use cases: e.g. ensuring totes are properly positioned on conveyors, spotting jams before they cause major downtime, and monitoring worker safety and ergonomics.
But the vision for computer vision (pun intended) extends much further. This technology is foundational for the “lights out warehouse”—a fully autonomous operation. For humanoid robots to function at scale, they need incredibly sophisticated vision capabilities. It’s not enough for a robot to see a product and identify it. The robot needs to understand how to grasp it, what orientation to place it in a box, and whether it can safely stack it on top of other items—all in milliseconds.
As John pointed out, humans do this innately. We automatically know we can’t place a heavy item on top of potato chips. Teaching robots this level of contextual understanding is where the real challenge lies.
3. Large Language Models: The Interface
LLMs represent the newest category of warehouse AI, and they might scale faster than the others. Why? Because they’re pure software upgrades that don’t require new equipment installations.
Today, getting deep insights from warehouse operations requires pulling data from multiple dashboards, exporting to spreadsheets, and creating pivot tables. It’s time-consuming and limits how quickly operators can identify and solve problems. LLMs change this by allowing operators to ask natural language questions like “Where is the bottleneck today?” and receive intelligent answers that triangulate data from multiple sources within seconds.
This dramatically reduces the time from identifying a problem to solving it, an idea that is paramount when it comes to operational efficiency.
The Crawl-Walk-Run Path to AI Agents
Perhaps the most fascinating part of our conversation was John’s explanation of how these AI categories will eventually converge into autonomous agents running warehouse operations.
- Crawl Phase (Happening Now): AI provides decision support with humans in the loop. The system might flag that certain SKUs will be out of stock and recommend actions, but humans review and approve everything.
- Walk Phase (Coming Soon): AI proposes actions with confidence scores. If confidence exceeds a threshold (say, 90%), the system automatically executes. Lower confidence recommendations go to humans for review. This feedback mechanism helps train the AI while keeping humans in control.
- Run Phase (The Future): Multiple specialized AI agents—each focused on specific tasks like demand forecasting or slotting—collaborate to make decisions and execute tasks automatically. Humans provide strategic direction and guardrails, but day-to-day operations run autonomously.
The Timeline: Sooner Than You Think
When I (Chris) pressed John on timelines for humanoid robots at scale, John’s answer might surprise you: we’re past the sci-fi demo stage. Real pilots are happening in warehouses today with companies like Amazon testing systems like Vulcan.
Over the next five years, expect increasingly complex use cases. While true “lights out” warehouses remain further out, the building blocks are coming together faster than many anticipated. Battery life, cost reduction, and vision capabilities continue improving, and major players are investing heavily in the technology.
What Should You Do Now?
John’s advice follows a logical prioritization based on ROI and maturity:
- Start with Optimization AI: Proven models exist with measurable benefits. Keeping humans in the loop provides confidence while you validate results.
- Scale Computer Vision: Particularly valuable in highly automated facilities for process control, safety monitoring, and preventing downstream issues.
- Implement LLM Tools: Next year will likely see significant deployment of generative AI for unlocking data insights. Since it’s a software upgrade rather than equipment installation, the barrier to entry is lower.
Eventually, these distinct categories will converge into a single intelligent system that prevents issues, optimizes efficiency, and orchestrates all warehouse processes. But that convergence will follow a logical path—one that’s already visible if you know where to look.
Listen to the Full Episode
This blog post only scratches the surface of John’s insights. To hear the complete conversation, including detailed examples and John’s predictions for specific use cases, listen to the full episode above or via the platform of your choice below:
Apple Podcasts | Spotify | SoundCloud | Amazon Music
Be careful out there,
– Chris, Anne, and the Omni Talk team
P.S. See our past 8 years of wonderful Spotlight Series podcast guests, featuring roughly 200 movers and shakers in retail, by clicking here
#warehouseai #retailtech #supplychain #automation #dematic #omnitalk #robotics #computervision #LLM #warehouseautomation #retailpodcast
Music by hooksounds.com
*Sponsored Content*
Omni Talk® is the retail blog for retailers, written by retailers. Chris Walton and Anne Mezzenga founded Omni Talk® in 2017 and have quickly turned it into one of the fastest growing blogs in retail.