The Now, Next & Future Of AI's Impact On Warehouse Operations With Dematic's John Mabe | Spotlight Series

Foreign.

Technology Spotlight series podcast is brought to you by the Omnitalk Retail Podcast Network, ranked In the top 10% of all podcasts globally and currently ranked in the top 100 of all business podcasts on Apple Podcasts.

The Omnitalk Retail Podcast Network is a network that we hope makes you feel a little smarter, but most importantly, a little happier each week, too.

Apple.

And this podcast is just one of the many great podcasts you can find from us here at Omnitalk Retail, alongside our Retail Daily Minute, which brings you a curated selection of the most important retail headlines every morning, and our signature podcast, the Retail Fast Five, that breaks down each week.

The top five headlines making waves in the world of omnichannel retailing.

And that comes your way every Wednesday afternoon.

Hello, everyone.

I am one of your co hosts for today's interview, Chris Walton.

And I'm Anne Mazenga.

And, you know, there's been a lot of talk around AI.

You know, we heard it in NRF Paris.

We've heard about it from Shop Fall, from some of our friends who were able to attend that.

And honestly, there's probably been a little bit too much, I think one could say.

And I'm guessing a lot of people are thinking that anyway.

But we're going to talk about it again, and that is because as an industry, I don't think we've done a good job of defining AI.

You know, we're talking about it too broadly and specifically which types of AI are most useful and in which settings.

And so to help us do that, we're going to examine warehouse operations.

We're going to look at AI through the lens of warehouse operations.

And to help us with that, we have invited Dematic's product manager, John Mabe, onto today's show.

John, thanks for joining us at omnitalk.

Hey, Chris.

Hey, Anne.

How are you guys doing?

We're doing great.

Great to have you.

Yeah, thanks for having me.

John, I'm, I'm curious, maybe you wouldn't mind starting with this.

Like Chris said, there's a lot of discussion around AI and its applications, but in your role as a product manager, how do you think of AI and its application, especially within warehousing, like, what framework would you recommend that we.

We use to start off?

You know, I think from a product perspective, first, it's probably best to break down AI into a couple, like, distinct categories because it's not just a single technology, it's more of a toolkit.

So the way we kind of look at that is there's one category is optimization AI.

So think of that as kind of the brains in a warehouse that makes decisions.

So what's the best, what's the next best order to work on?

For example then there's vision and perception AI.

So that's the eyes of your operation that's using cameras and sensors to understand the, the physical state of a warehouse in real time.

And then kind of a third category which I think most people are familiar with is kind of the chat GPT generative AI, you know, a new interface or conversational partner of how you interact with your systems.

Got it John.

So let's click into those a little bit then.

So like you know optimization AI, the first one you said like you know when I hear that like the first thing that comes to mind for me is like that's, that's not anything new.

Like I feel like I've been hearing about that for ever like since the history of doing Omnitok we've been doing the show for eight years so.

So is that true?

And then number one like if that is the case where you, you mentioned a place where it is applicable but where is it most applicable in the warehouse environment?

I mean, yeah, you're right, optimization is not new and even pre AI, right people have been using math to make decisions and.

Yes, exactly Matt.

So if you were to take, you know inventory optimization is pretty a big brinket umbrella where it can be applied really well.

So if you were to take slotting for example, I mean historically warehouses have used, you know, ABC classifications or velocities or order affinities, which items are typically ordered together to make these slotting decisions and you can get a lot of value out of that pretty quickly.

You know the main issue there is that that is kind of static and rules based and it doesn't really evolve over time.

So where I started to come in is adding better inputs and adaptivity.

You know with supervised learning you can really, really truly predict the velocity of skews.

Even you know, the really hard things to predict.

The fast movers are fairly easy but you know, maybe the medium and slow movers, you know, seasonality, different channels, promotions and you combine that with reinforcement learning we can go further where the system, the AI can learn and figure out how to best re slot the inventory to deal with, I don't know, an upcoming Flash sale or promotions or back to school Special.

So it's really able to adapt to shifting demands over time and self learn and really discover non obvious patterns that like our traditional optimization couldn't do very well in the past.

Got it, got it, got it.

So, but the strict use case is like really around, like forecasting for inventory, labor planning.

I'm curious.

So John, like, if it has been around for so long, how many companies and particularly retailers that you in your experience are actually using it?

Like, is everyone using it or is a portion, like, what's your take there?

I'd say it's a pretty, you know, fairly small portion.

I think the largest, really the largest players are using it.

I mean, I think most companies by now are using some sort of optimization using math models.

Right.

But as far as true AI, it's still fairly limited in its use.

I mean, it's mostly the bigger players that are using it that have, you know, have access to really good data and have their data structured in a way that the AI can leverage it.

What keeps the smaller players from going all in on this?

It seems like it'd be even more advantageous to somebody like that who doesn't have the labor bank bandwidth.

I mean, part of it is the, you know, is the structuring of your data.

Like, if it's not structured properly, okay.

It's hard for these models to be effective, you know, and I think we're starting to see kind of a, you know, like a crawl, walk, run approach where people are starting to.

You.

Even the smaller players are starting to use it, but maybe they're using it more as a decision recommendation.

So it's not necessarily completely embedded into their system.

Jet.

It's more, hey, this is, this is the forecast that our model came up with.

Here's a report and you can take action on it.

But as far as like a fully automated, it's kind of the larger players that are kind of at that stage.

That makes sense.

How, how does growth play into it too, John?

Like if, say, say if I'm a smaller player and I'm not really sure my demand, I'm growing, like, is it as applicable to somebody that like, to your point, like, is more established and has very consistent patterns in what they're putting through their warehouse give every single day.

Is that fact factor into it at all or am I overthinking it?

I'm curious.

I mean, it should be, you know, a. I think as.

As AI gets more advanced, it's able to adapt better.

Should be able to do either one, right?

Yeah, it really should be able to do either one.

Yeah.

Right.

Well, let's talk a little bit too about vision and perception, John.

Can you explain kind of how that's coming into play?

Sure.

So vision is, you know, we view that as like foundational for getting to this.

You know, people talk about lights out warehouse, where it's a very autonomous run warehouse.

Right.

So optimization would be a big part of that.

Vision would also be a huge part of that.

And that's really about, you know, using cameras and sensors to process information and make decisions in real time.

So right now that's very limited in the scope of how that's being used.

So you've got like process integrity.

Say you have, you know, a tote on a conveyor.

Is it positioned appropriately on the conveyor?

That's not going to create an issue downstream.

So it's really about how can we minimize errors that take a lot of time to fix, you know, downstream from that process and then, you know, spotting, like if let's say there's a jam on a conveyor, the camera can see that and alert someone before it becomes, you know, like a major downline event.

And then probably the other area where it's applicable now in these kind of narrow use cases around kind of safety.

Are humans working safely with robots?

Are they, are humans working at a workstation safely, like in an ergonomic fashion so they're not getting fatigued like a camera and AI can pick up on that.

And so that's, you know, it's kind of being deployed now in these kind of somewhat narrow use cases, adding value but not quite there of lights out type warehouse, I guess.

So John, are these systems, these computer vision systems, for lack of a better way to put it, are they, are they fixed position?

Are they on robots?

We heard it nrf.

We were surprised, Anna and I both were surprised at nrf how many, you know, people were talking about robotics impacting both store in store operations, but also particularly the warehouse operations.

So like is what you're describing.

It sounds like you're talking more fixed position camera systems, but also potentially the application on robotics moving throughout the warehouse too.

Give us the lay of the land there.

What is, how do you think about all that?

I mean, I think it's a combination of both.

I mean that kind of example I gave you were more fixed like on a, at a workstation or on a conveyor to spot issues.

But there are cameras and robotics and AMRs to make sure that they're navigating safely through the warehouse, just like they're, you know, cameras on a Tesla for self driving.

So a similar concept, you know, and I think the we are at the early stages.

I mean, you know, humanoid robots are all kind of you know, really in the media now.

And that is something that is being worked towards the technology still very early, but that's quite interesting.

What needs to happen, do you think, John, before we get to that point?

I mean, is it.

Is it mostly perception?

Is it level of comfort?

Is it the technology needing to get to a certain state?

Like, where are we in that, in that.

And what needs to happen before we.

We kind of get to that next level?

Yeah, so I guess I probably address that from a vision perspective.

Like, you know, so from a vision perspective, you know, it's not good enough for a robot just to see a product and understand what it is.

It really needs to.

Let's say you want to have a robot that does picking.

You want to have a humanoid robot that is picking inventory from a inventory tote and placing it into an order shipping box.

It has to have a deep understanding of how to interact with that product.

So it needs to be able to answer pretty complex questions in milliseconds.

These are types of questions humans just do innately, but needs to understand, you know, based on the product.

How should I grasp the product?

How should I.

What is the orientation when I place it in the shipping box?

Can I place it on top of a bag of potato chips or is it going to crush it?

So there's a lot of, like, yeah.

Yeah, there's a lot of questions it needs to run through.

It needs to run through those very, very quickly.

Whereas humans, we just know, right?

We just pick this up and be like, I can't put this on top of potato chips.

It will crush it.

So there's a lot of buzz there.

But the vision is quite important to unlock that technology.

That's interesting, John.

And I remember Amazon Vulcan seeing some videos on that.

In terms of how that works is kind of talk.

It kind of gets at what you're saying, like a robot that can kind of get those use cases, edge cases.

But I'm curious because like I said before, I was surprised at NRF how many people were talking about huge humanoid robots.

And part of the pun of me doing this for eight years says, okay, that's because robots are always the sexy thing.

And are we just saying that because they're the sexy thing?

If you were to predict how far out are we from a humanoid robot functioning at scale in a warehouse environment, are we talking 10 years, 15 years, 5 years?

What's your take?

Well, let me start that off by saying I do feel like humanoids will be a big part in warehouses.

And one of the major Challenges is they fit right in.

They, you know, they have the, they have arms and legs or maybe they're on wheels.

But they can fit right into a warehouse environment without any costly infrastructure changes.

And there's a lot of really major players that are pouring a lot of money into it.

So my optimistic take is that that will gain traction and there will be manufacturing scale at some point.

Now, you know, I view where we are today as we're past the, the sci fi demo stage.

You know, when I was 10, like I watched the Jetsons all the time, I wanted Rosie to clean my room and I wanted a fly car and you know, we're not there yet, but we are past that sci fi demoist type stage.

Right.

So there are real warehousing.

You mentioned Amazon, Vulcan, there's our real warehouse and manufacturing pilots happening today.

I expect those to continue.

You know, those use cases happening today are fairly simple.

You know, they're not super complex use cases.

But the complexity of the capabilities they'll be able to, will grow over time.

And just the fact that a humanoid could do is more is multi purpose.

You know, it can, it could do picking and packing and receiving and maintenance.

You know, people that are old enough that, you know, pre smartphone, you know, you might have had a GPS, an iPod, calculator and a telephone.

That's all in one device now.

So general purpose usually wins in technology.

So I feel like we'll get there.

It will take time.

Envision is a big part of it, but I think you also have kind of the boring stuff, you know, like you got to get the cost down, battery has to have enough life, needs to be able to charge fast.

So there's a lot of other elements there.

But you know, I, I feel like it will come together, you know, as far as a time frame, it's really hard to say at scale.

I feel like over the next five years you'll see a lot more increasingly complex use cases being done and then eventually it will kind of just come together kind of like a smartphone to over time.

Got it, got it.

So I take away two things from what you said there.

I think one, I.

One, it's closer probably than I think.

You know, based on how I set up that question, it's probably closer than I think it is.

And then two, like I said at the outset, even the word humanoid is kind of a disservice because a humanoid can come in many shapes, sizes and forms too, if I'm hearing you right John.

And so like, yeah, that can, that's going to change and, and, and shape as the future plays out here too.

All right, well, let's go to the Last one then.

LLMs.

How are warehouse operations deploying that side of AI?

I mean, you know, this is fairly new.

You know, everyone uses ChatGPT.

This is about, you know, connecting that into your warehouse data.

A chat GPT like, you know, LLM and you know, kind of the value we see is that, you know, today to really get to information to understand your operation.

I mean there's, you know, people have dashboards for KPIs and you can see that, but it's not really, you can't get to the deep understanding or kind of root cause.

If there's problems, you might have to go to multiple sources, export some data to a spreadsheet, I don't know, create a pivot, pivot table.

So it takes a long time to get to insights and our feelings.

This kind of LLM models, you'll be able to ask questions like where is the bottleneck today?

And the LLM will be able to go off and triangulate data from multiple different sources and then source a more intelligent answer and display that back to the user within seconds and even display it as like a chart with trends or whatever format is appropriate.

So, so it's really about kind of dramatically reducing the time from question or you have a problem to solve to actually getting insights out of it and just being able to solve problems faster.

I have so many questions for you, John, here especially like, you know, you use that example of like the LLM model being able to ask like where, where's the friction point here?

How far out are we then from like that question being posed to like an AI employee and then the AI employee answers that and is really running the whole thing themselves.

I mean, I think that's kind of the ultimate vision where it's, you know, many different agents that have a, you know, very specific task and they're, that's, you know, I am the SKU demand forecasting agent and I'm the slotting agent and they collaborate together to make decisions and ultimately, you know, create the task and have the task executed automatically.

You know, I don't see that happening overnight.

This is like kind of like the crawl, crawl, walk, run path to that future.

So the crawl side is more, more around decision support where a human is in the loop.

Maybe if we go back to the, you know, a forecast, you know, maybe it surfaces a risk, hey, these SKUs are, are gonna, we think these SKUs will be out of stock between 2 and 4 today.

Here's, here's what we recommend that you do.

And the human looks at it and says, yep, I'm good with that.

And then they, they get those tasks executed.

And so the humans are still making the decisions there, and they're not just turning the keys over to AI.

And we need to build trust and validate the technology.

And then I think from there you go into a more decision intelligence where the AI is proposing actions and maybe they're proposing a confidence in those actions, and then maybe there are some more systematic rules that are configuring.

Say, okay, if the confidence is above 90%, automatically create those tasks.

If it's lower, service them to me.

I'll review them and either reject them or move forward with them.

And so that would be kind of a feedback mechanism that will also help the AI train and then also keep, you know, kind of keep the people still in charge.

So certainly just like, there's still a human in the loop and there's still somebody that's analyzing the data that's coming together, but it's fewer people that need to be part of that that are making those decisions.

So it can speed up manufacturing or speed up the process of whatever those robotics are trying to do.

Got it.

That's right.

And I think, you know, longer term, I think the agents are more running the building, you could say, but.

But there's always human oversight and the strategic direction.

You know, guardrails need to be in, be in place for safety and your labor rules in the market, depending on the market you're in and your KPI targets.

So there's still human involvement there and overall strategic leadership.

But yeah, ultimately I think you will see AIs making more decisions independently and collectively, I guess.

So, John, it sounds like.

So it sounds like at the start, it's more of a management tool and a productivity tool.

And I'm curious, you know, you know, how far out is that?

Like, are people starting this?

Like, are you seeing people try to do this?

I mean, the decision support is happening now where we will recommend.

So we'll recommend, hey, we have this AI algorithm.

We've forecasted the demand for the next hour, four hours, day, whatever.

This is what you should replenish.

And then the humans will make decision.

Like, yep, of these hundred that you recommended, I'll do the top 80 or something.

So that's happening today.

And the decision intelligence is probably the next phase where it's where we kind of have more rules embedded that will decide, like, okay, if the Confidence of the AI output is above a certain percentage, then we'll just automatically accept that and create those tasks.

Otherwise we want to be involved in the loop.

But that's all coming soon.

And John, are you seeing that firsthand?

Like, like, should I interpolate that you have.

You and Domatic have POCs on the ground with different resellers who are trying to experiment with this in the warehouse?

Yes, I should.

Okay, yeah, yeah, we do.

We have customers now that are piloting like, the, our forecasting algorithms and forecasting models.

Okay.

All right.

Wow.

All right.

Wow.

This is really fascinating and I love how John breaks it down.

It's really simple and really easy to follow too, and gives us a good idea of where things are in the timeline.

Because what's interesting to me is based on what John said, it sounds like the last part of this, the LLMs might actually take off faster than everything else, which has been in place for a while, which also makes sense because it's, you know, pretty much.

I mean, they're all software, but it's very easily attainable software too.

So.

So, John, with that said, what are the main, what are your main takeaways here for the audience?

Like, you know, and, and, and, and going back to that last point, in what order do you think all of this will happen?

Okay, so I guess maybe the first takeaway is AI is not just a single technology.

You know, it's a toolkit.

So we talked about the optimization is the brains of your operation.

The vision and perception are the eyes, and LLMs are the way you can interact with your software.

So that's probably the first one.

And the second one is this is, as we talked about, this is not overnight thing.

This is a journey that's going to follow a logical order based on ROI and maturity.

And it will take time.

Some areas will happen faster than others.

But I think starting with optimization, AI, I mean, there are proven models that work for that.

It's pretty measurable, pretty easy to measure that you're getting benefits.

And if you're keeping the human in the loop, that gives you the confidence that if we don't like the output, we don't have to use it.

And then I think the computer vision is kind of scaling now more in your really automated facilities that less so in manual warehouses, but really automated facilities.

And I think you'll see that continue to scale for specific kind of process control to make sure that we're operating as efficiently as possible and, you know, and safety and preventing downstream issues.

I think that will Continue to expand to more use cases in the warehouse.

And then I think, you know, next year you'll probably see a lot of this generative AI, like just unlocking the value of the data through AI that's in your systems today.

It's already there, so you don't have to, you know, that is not a huge initiative.

That's a software upgrade, right?

To get that type of functionality out there.

It's not like we have to go, you know, install a bunch of new equipment and then I think, you know, eventually all these pieces will come.

I mean they're now pretty distinct separate categories, but they'll all converge into a more single intelligent system that can like, you know, whatever, prevent issues and optimize efficiency and orchestrate the agents orchestrating all the processes in the warehouse.

So there, I think there will be a logical path for those to converge at some point.

Is that the, AKA the brain, John?

Is that what that is?

Is that what that refers to, like the, the proverbial AI brain?

Yeah, I think so.

I mean it's just kind of the, that concept of a lights out warehouse where it's like, yeah, you can.

Wow, John, this has been so insightful.

I especially love what I think this unlocked for some of the people we were talking about earlier here who are earlier in their stage of automation and especially the AI components of automation in their warehousing facilities.

I'm sure there's people that are going to want to continue this conversation, dive deeper into their own use cases with you.

What's the best way for them to do that, John?

So I'm happy people can reach out to me directly.

I'm John Mabematic.com or find me on LinkedIn.

Also on the Domatic website there is, there is a site, a place there to connect with sales.

Excellent.

Right, well that wraps us up.

Thanks to John Mabe of Dematic.

Thank you so much John.

It was great.

Thanks for educating us today and thanks, thanks to everyone out there for listening in and on behalf of all of us at omnitalk, as always, be careful out there.