Why Warehouses Are Slow to Adopt AI (and What Actually Works)

Every warehouse technology vendor in 2026 has an AI pitch. Self-driving forklifts, AI-powered voice picking, computer vision for compliance checking. The demos look impressive, but what is keeping warehouses from adopting AI and automation? According to Justin Griffith, CTO of StayLinked and a 22-year veteran of warehouse technology research, most of these technologies are stuck in a gap between “can it do it?” and “is it practical to deploy?” StayLinked’s terminal emulation software runs on millions of frontline devices across warehouses worldwide. Griffith’s research arm, StayLinked Labs, has spent over a decade studying why promising technology fails to get adopted.

Watch the full episode for the complete discussion.

What AI Use Cases Actually Work on Frontline Devices Today?

AI running on a frontline device (Edge AI) has to earn its place by proving it does not degrade the workflows it is supposed to improve. Battery life, processing constraints, and latency all limit what is practical today. Three categories show the most real-world traction.

1. Computer vision for compliance and identity

Pointing a device camera at a shelf to verify product placement or compliance is relatively straightforward and runs well on current hardware. More advanced use cases, such as verifying that the identity of the person holding the device matches the user signed in to the WMS, are emerging. Griffith highlights the value of tying device identity to system identity: “The WMS knows the username. But the device up to now hasn’t really had any indication of who that actual person is.” Closing that gap prevents scenarios like unauthorized workers operating under someone else’s credentials during a shift.

2. On-device chatbots for procedure lookup

Loading a training manual or procedure guide into a local AI model lets workers ask questions without calling a trainer or radioing a manager. This is one of the first practical use cases Griffith sees working on current rugged devices. It does not require constant inference and does not compete with the primary picking or scanning workflow for resources.

3. Voice-directed work (with caveats)

Voice picking has run warehouses for decades through systems like Honeywell’s Vocollect. AI-powered voice promises better natural language understanding, but Griffith warns that voice quality must be near-perfect. “You can’t dabble in voice,” he says. “If you have to repeat yourself, you’re annoyed. If you have to repeat yourself twice, you’re ready to throw whatever you’re holding across the room.” Any AI inference latency added to a voice workflow subtracts directly from the speed gains voice is supposed to deliver. For warehouses where half a second per transaction multiplied across 10,000 daily transactions adds up to hundreds of thousands of dollars annually, that drag matters.

Prefer to listen? Catch this episode on your favorite podcast app.

Why Is AI Adoption in Warehouses Slower Than Vendors Promise?

The answer is not technological capability but rather integration risk. The warehouse management system is the most expensive, highest-risk piece of software in most enterprises. Deploying AI-driven automation (self-driving forklifts, robotic picking, drone-based inventory counts) requires integrating it with the warehouse management system (WMS). And that connection is where projects stall or fail.

StayLinked Labs ran back-to-back technology adoption studies in 2022 and 2024. In the first study, companies expressed strong interest in robotics and automation but lacked the software infrastructure to even test them. Many were on legacy WMS versions with no API endpoints for integration. By the second study, more companies had upgraded their WMS, expecting easier deployment. They were right that upgrades made integration easier. But Griffith draws a critical distinction: “There is a huge difference between something being easier and something being easy.”

New WMS versions shaved perhaps 25% off the integration timeline, taking a two-year project down to a year and a half. But a year and a half is not overnight. The result: more companies could trial AI and automation solutions, but there was no measurable increase in actual deployments.

The risk conversation compounds the problem. Migrating from one enterprise resource planning (ERP) system to another is a three- to five-year project if managed well. Many companies have tried two, three, or four times before pulling back to their existing system. “The juice was not worth the squeeze,” Griffith says. Any new AI tool a vendor pitches has to clear what he calls the “barrier of viability.” It must do at least as good a job as whatever exists today. The risk mitigation conversation kills the deal before the pilot even starts.

How Should You Justify AI Spending When Hardware Needs Are Concrete?

For IT leaders presenting AI investments to a CFO, the challenge is competing against tangible alternatives. If your company makes lumber, the CFO can see the return on a new saw. If your company makes bolts for bridges, the CFO understands a new production line. AI’s benefits are harder to quantify in advance, especially when the deployment timeline stretches into years, and the success rate for complex integrations is uncertain.

Griffith’s advice: treat AI like any other technology adoption. Start with a use case where the before-and-after results are shown in units the business already tracks. Think boxes out the door, pick accuracy, or transaction speed. Avoid the trap of buying a technology and then looking for problems to solve with it. “It would be like buying a screwdriver and walking around your house for everything that the screwdriver fits,” Griffith says, echoing a pattern he sees repeatedly across the industry.

Connectivity interruptions offer one concrete example. StayLinked Labs found that in a warehouse with 50 workers, Wi-Fi connectivity disruptions alone can cost $400,000 per year in lost productivity. That is the cost of a WMS upgrade, a device refresh, or a complete network overhaul, every single year. Solving a problem with known, measurable cost is far easier to justify. Deploying AI for unproven benefits in your specific environment is a harder sell.

Frequently Asked Questions