Security Cameras as AI Teammates: Spot AI on Agents for Factory Safety and Operations
Manufacturing Happy Hour
In this episode of Manufacturing Happy Hour, recorded just before a large Automate afterparty in Detroit, the host and co-host Jake Hall (The Manufacturing Millennial) talk with Dunchadhn Lyons of Spot AI, a company that describes itself as doing video AI for the physical world. The conversation centers on one question: what AI agents actually do on a plant floor, and how existing IP cameras can be turned into what Lyons calls "AI teammates." His position is that these agents, watching video around the clock, can make workers safer and operations more efficient without new infrastructure, and that their role will expand quickly as underlying models improve.
What an AI Agent Is
The host asked Lyons to explain an AI agent as if over a beer at a brewery. Lyons defined it as artificial intelligence that performs a task on your behalf: it acts autonomously, understands its environment, and does something for you. His brewery example was a video AI agent on a camera watching the taproom. If a line forms at the bar and no one is pouring, the agent could play a message over a speaker in the back saying that customers are waiting and someone needs to come out.
Jake Hall offered his own, more digital use. He posts on LinkedIn every day and has built an AI agent that knows his general format, flow, and tone, including his preference for facts and bolded text, which makes writing and content creation easier for him. He then asked how agents, beyond the buzzword, make life easier for manufacturers.
Two Focus Areas: Safety and Operational Efficiency
Lyons said Spot AI concentrates on safety and operational efficiency, working primarily with video from IP cameras. On the safety side, an agent can watch all camera footage in real time, 24/7, and recognize whether people are wearing required personal protective equipment such as hard hats, safety vests, and gloves. It can check whether forklifts are driving too fast, entering restricted zones, or having near misses with pedestrians.
The agents can then act: sending an alarm, or, if a person gets too close to a dangerous machine, automatically shutting that machine down. Lyons also described a retrospective function. The agent can build reports on how many incidents are happening, so the company can better train its people.
Why "Teammates" Rather Than Surveillance
The host noted that a layperson hearing "security camera as AI agent" might react with alarm, and asked Lyons to explain the teammate framing. Lyons argued that the agents watch video around the clock on behalf of a human, a job no person would want to do. By providing that reporting, they make employees safer and more efficient. In that sense, he said, the agent acts as a second or third safety manager, or another operations manager, freeing people to focus on higher-level problems and strategy rather than mundane tasks.
Capturing Tribal Knowledge Before It Retires
Jake Hall raised the workforce problem: he cited 600,000 open manufacturing jobs and an average worker age of 55, with much of the workforce long-tenured and a large amount of tribal knowledge set to retire in the next four to five years. He asked whether a vision-based agent could watch operators solve problems and turn that unwritten knowledge into work instructions.
Lyons described today's AI teammates as junior employees: fresh and green, but trainable and coachable, able to perform tasks. In the very near future, he expects that with foundational models and more advanced reasoning, they will "graduate" to more senior roles. He said they can watch footage and understand what is happening and the context of the environment, follow a process a person is carrying out, and break it down step by step. Where he sees this heading is agents serving as training mechanisms for new human employees, so that tribal knowledge is not lost. This is framed as a direction he expects, not something he described as already deployed.
The host said the idea of an agent absorbing a retiring worker's hard-won skills and teaching them back to newer employees has become a recurring theme on the show over the past couple of months. Lyons added that current models are multimodal. Spot AI focuses on video today, but he said it will "very soon" be able to ingest other business data, such as ERP systems or standard operating procedure documents, giving agents context on how the business is supposed to operate, which could then feed into training people.
A General Platform Rather Than Industry-Specific Solutions
Asked which manufacturing sectors use the technology, Lyons said Spot AI works with many manufacturers, primarily on safety use cases, and that it spans the board: if you have forklifts or people walking around a warehouse, or an assembly line that might get backlogged, the company can help. He said the platform was deliberately built to be as general and flexible as possible, so an agent can take the context of a specific environment and apply it, rather than building narrow solutions for particular subsectors.
Safety Case: 600 Cameras and a 40% Reduction
The host asked about an example involving a roughly 40% reduction in incidents. Lyons described a large manufacturing customer where Spot AI deployed AI teammates to detect forklift near misses and missing PPE. Before that, the company had a single safety employee who spent eight hours a day combing through footage from 600 cameras, which Lyons put at something like 16,000 hours of footage, looking for safety incidents to understand their frequency and how to reduce them.
According to Lyons, the AI teammates were deployed within days and took over that review. She then only had to look at relevant clips. The agent surfaced, in his words, five to ten times more incidents than she had found manually. She spent her time improving training and working with people on the floor to make sure policies were followed. A month later, Lyons said, the company had a good understanding of its incidents, people had been retrained with real-world examples, and safety incidents had dropped by 40%. These figures are as reported by Lyons for this customer.
What You Need to Get Started
Asked what assets a company needs, Lyons called this "maybe the most incredible thing." Many vendors he and others had spoken with at Automate over the previous days were selling new infrastructure or machinery. In Spot AI's case, he said there are already more than 100 million IP or security cameras deployed across US businesses, most of them recording and doing little else. What Spot AI supplies is a small hardware box about the size of a PC. You plug it into the internet and power, and any cameras already on the local network are streamed in and become AI teammates.
Getting Workers Bought In
The host returned to the concern raised at the top of the episode: with agents looking for forklift and PPE violations, how do you keep it from feeling like Big Brother is watching? Lyons said much of it comes down to education. People have a natural twinge of fear about AI because of the sci-fi movies they've seen, but the data matters: a 40% reduction in safety incidents means people are safer, can do their jobs, and go home to their families without fearing a couple of weeks in the hospital.
He said the AI teammates are about supporting humans, making them safer and more efficient, and augmenting their abilities, and that there is no notion of replacement or monitoring in that sense. Spot AI partners with customers' safety and operations teams so they are well educated and can bring their people along. In his experience, even when there is minor resistance at first, people buy in fairly quickly once the teammates are in place.
Operations Case: Keeping a Distribution Line Out of "Red"
For an operational example, Lyons described a very large clothing manufacturer whose packaged goods travel down a distribution line. Volume rises at different points in the day, and the company classifies the line's state as green (everything fine), yellow (starting to clog, when intervention could prevent escalation), or red (full stop, meaning downtime and lost revenue).
Spot AI deployed an agent, using an IP camera already installed, that recognizes what a yellow scenario looks like, with slightly higher volume coming down the line. The agent immediately alerts an operations manager on the ground, who can tune the line to keep it from reaching red. Lyons said that over the course of a year this is going to save the company tens of millions of dollars in lost revenue from downtime, a forward-looking estimate rather than a reported result.
Where Agents Could Go: From Instructions to Objectives
Jake Hall asked for the 10,000-foot view of what AI agents might do five years from now. Lyons said "the sky is the limit." Pointing to foundational models such as Claude, ChatGPT, and Gemini that can take in many data sources, he predicted that AI teammates will soon ingest all relevant business and video data, and that users will give them a goal rather than a narrow instruction. Instead of "watch for forklift near misses and text me when you see one," the instruction would be "make this process 25% more efficient," and the agent would reason over its data sources to help plan and implement a strategy.
He said the digital world is already starting to get there at the enterprise level, citing OpenAI's ChatGPT deep research: you give a broad outcome, such as learning about a topic or solving a problem, and it spends 15, 20, or even 30 minutes reading and absorbing material from the internet before producing a full report. Lyons expects the physical world to reach a similar point very soon.
The Education Gap That Remains
The host said that whenever they bring up AI agents in a manufacturing context, few people feel they understand the concept. Most people's default picture of AI is help writing code or a copilot for generating ideas. They asked what it would take to make agents a widely understood concept.
Lyons called it a very good question and said Spot AI is grappling with it too. He credited ChatGPT with making agents more accessible to more people, and said that with prospective customers the company describes itself as "ChatGPT for video," which helps the idea click. But he acknowledged it is still something the company is figuring out, and said they will be working to educate people better over the next couple of years.
The episode closed with the three comparing their beers, a Mexican-style Vienna lager, a New Zealand-hopped pilsner, and a New England pale ale, before heading into the Automate afterparty.
How do you get buy-in from the team to make this not feel like a big brother is watching?
Big brother's watching over you the whole time.
A lot of it comes down to education, right? When a lot of people hear about AI, there's kind of a natural twinge of fear because of all the sci-fi movies that we've seen.
We've been talking a lot about AI agents on Manufacturing Happy Hour lately. And today we're going to be doing that with Spot AI, a company that does video AI for the physical world. We'll be looking specifically at the impact that AI agents are having on safety and operations. All this as we prepare for a massive Manufacturing Happy Hour party that we just threw in Detroit, Michigan, today on Manufacturing Happy Hour.
Cheers. Cheers. Absolutely. Yeah. Party time. Yeah. Thanks for having me on.
Well, Dunchadhn, this is going to be one of maybe the most active podcasts we've ever done because we literally have what might, well, what will be hundreds. I just don't know how many hundreds of people filing in for this event. So, should be a good time. But, Dunchadhn, you're with Spot AI.
This conversation is around AI agents. So, I think the best, most natural way to start would be how do you describe an AI agent as if you're having a beer with someone at Batch Brewing in Detroit. Let's start there.
Yeah, absolutely. So, an AI agent is artificial intelligence that can perform a task on your behalf, right? It's going to act autonomously. It's going to understand your environment and it's going to do something for you. So maybe in a brewery circumstance, you could have a video AI agent on a camera that's looking at your tap room. And if there's a line forming at the bar and there's no bar attendant pouring beers, it's going to play a message over the loudspeaker in the back to say, "Hey, we got a line. Somebody needs to come out and pour these beers." Right.
And so I've used AI agents in the past more for LinkedIn. So what I've done is I make a post every single day on LinkedIn and I've created an AI agent to say, "Hey, this is the general format. This is the general flow. This is the attitude. I like to have facts and bolds and all that stuff." So for me, I'm able to create an AI agent to make my life easier when I'm writing and creating content.
How can manufacturers begin to understand? AI is such a buzzword in the industry, but we're focusing on AI agents. How do AI agents really begin to make manufacturing lives easier?
Yeah, absolutely. And we focus kind of in two areas. One is safety and one is operational efficiency. So in the safety space, and we're primarily working with video inputs, right, from IP cameras, you can have an AI agent that is watching all of your camera film 24/7 in real time and it can understand things like are people wearing the personal protective equipment that they're supposed to be, right? Hard hats, safety vests, gloves. Are forklifts driving safely in your environment or are they driving too fast? Are they in restricted zones? Are there near misses with pedestrians, right? And the agents can then act on your behalf to, you know, send off an alarm. Maybe a person's getting too close to a dangerous machine. It can automatically shut down the machine. But then also, more retroactively, it can build a report for you so you really understand the number of incidents that are happening so you can better train your people and make them safer.
So, put this in the context of, as I've heard you describe security cameras as AI agents, as AI teammates, which I think we need to define, because if I'm a lay person and I hear security camera as an AI agent, my first impression is going to be I'm going to freak out. But flip the script on this. How are they AI teammates?
Yeah, absolutely. So, as I said, you know, these AI agents are watching video 24/7 on behalf of a human, right? And that's something that no human would really want to do, sit around and watch all that video. And they're then able to make the lives of the employees at the business better, right? They can make them safer. They can make them more efficient by providing, you know, this kind of reporting. And in that way, the agent really is kind of your teammate. They act as your second or third safety manager. They act as, you know, another operations manager to make life on the floor easier and allow your people to focus on higher-level problems and strategy rather than more of these mundane tasks.
When I think of problems in the industry, one is workforce, right? We have 600,000 open jobs in manufacturing. The mean average of people working in the industry is 55 years old, of which a majority of them have been in the industry for a long time, if they're around that long.
Honestly, I think we have enough people coming tonight to this party to fill the skills gap. If they could, 850 people.
Yeah, it's going to be a packed house tonight. But when we look at the tribal knowledge that currently exists on the manufacturing floor that will be retiring in the next four to five years, how can an AI agent with a vision system be able to look at operators, understand how they're going about fixing problems, addressing issues, being able to create work instructions that might not actually be on paper, but in the knowledge? How can an AI agent move that and really capture that data and information?
Yeah, and that's a great point. Where we're at today, we think of these AI teammates as junior employees, right? They're fresh, they're green, but you can train them, you can coach them, and they can perform tasks for you. But in the very near future, with the advent of foundational models and more advanced reasoning, these agents will be able to graduate to more senior employees. And what's so interesting is they're able to watch footage and truly understand what is happening, right? What the context of the environment is, follow a process that a person is following, break it down step by step. And where we see this going in the very near future is those agents can then serve as training mechanisms for other new human employees, right? So that tribal knowledge is not lost, as you said.
That's a key part I keep hearing about AI agents and hopefully something that manufacturers take to heart. A good agent can take all the best skills of the person that's about to retire, with all the rare talents that take years to acquire, and then in turn teach back to newer employees on the team. You know, that's something that comes up on Manufacturing Happy Hour quite a bit, or it has started to just within the past couple months. So I would expect that to continue to be a theme.
Absolutely. And these models right at this point are multimodal. They can take in lots of different forms of data inputs. So, you know, we primarily focus on video today, but very soon we'll be able to ingest other business data from, you know, maybe your ERP system or standard operating procedure documents, right? Other sources of data that give the agents context about your business, how things are supposed to operate, and then that can all be fed right back into people that you're training.
What industries do we really see this? I mean, manufacturing is broad, you know, and there's all different types, from medtech, automotive, packaging, pharma. Where do we see this technology kind of being deployed?
Yeah. So, right now we're working with lots of manufacturers, primarily on safety use cases. So, it really spans the board, right? If you've got a forklift, you know, if you've got people walking around your warehouse, you need to understand if they're doing things safely, we can help you. If you've got an assembly line, you need to understand if it's getting backlogged, no matter what it is that's going down, you know, we can help you. So we've tried to build this platform as generally as possible and as flexibly as possible so that the agent can go in, take the context from your specific environment and then apply it to solve problems, rather than building very specific solutions for specific subsectors.
I'd love to hear two examples, two stories from you, whatever way we describe it, because we've talked about the operational advantages that AI agents can bring as well as the safety advantages. And I know you have specific examples of both. Let's start on the safety side. Let's go back to that. I mean, I think I heard at least one of your examples, you're able to help reduce incidents by like 40% leveraging these, you know, AI teammates, these security cameras. Tell me more about this.
Yeah, absolutely. So, we brought in AI teammates to understand forklift near misses and missing personal protective equipment, primarily, for a big manufacturing customer. And previously they had a single safety employee who spent eight hours a day combing through footage from 600 different cameras. That's something like 16,000 hours of footage, trying to find examples of safety incidents, right? So they could get an understanding of how frequently they're happening and then figure out how to reduce them. And, you know, within days we were just able to deploy these AI teammates. It does all that work for you, right? She doesn't have to watch any more footage except for, you know, the clips that are actually relevant. And so in that way, she was then able to spend those eight hours, instead of focusing on manually combing through video, making the trainings better, right? Pulling in the, turns out, five to 10x more incidents that the agent could actually detect versus her, you know, manually combing through it. And then spend that time working with the people on the ground to make sure that they were actually following policies. And then, you know, a month later, not only do you have a very good understanding of the incidents, but your people have been retrained with real-world examples. 40% reduction in, you know, these kind of safety incidents, as you said. So that's a really cool example.
So where do companies start when they're looking at this? What existing assets do they need to have in place? What assets do they need to deploy if they want to begin testing out these AI agents on their floor? What do they need to have at a minimum, and then what do they need to add on to that?
Yeah, that's maybe the most incredible thing. We've talked to a lot of manufacturing folks at Automate, you know, the past few days, and a lot of the businesses that are selling at the trade show are talking about a lot of new infrastructure to be put in. Maybe it's new machinery or whatever it may be. In our case, there are already 100 million plus IP or security cameras deployed across US businesses, and most of them are recording and not doing much else. All you need from us is a small box of hardware the size of a PC. You plug it into internet, you plug it into power, and any cameras that are already connected to your local network get streamed in and bam, they're AI teammates. So, it's really that simple.
While we're talking about getting started, team buy-in is another thing. And I know we're calling the security cameras AI teammates, but what about the team on the floor as well? Particularly with the safety example, right, where it's looking for forklift issues, it's looking for PPE issues. That's the one that makes this question come to mind. How do you get buy-in from the team to make this not feel like a big brother is watching?
Big brother's watching over you the whole time to make sure you're following the rules.
Right. How do you make them realize, I shouldn't say make them, how do you get folks bought in that this is actually something that's helping them and everyone else?
Right. Yeah. And I think a lot of it comes down to education, right? When a lot of people hear about AI, there's kind of a natural twinge of fear because of all the sci-fi movies that we've seen. But when it comes down to it and you look at the data, a 40% reduction in safety incidents means that people are safer, right? And it means you can go to work, you can get your job done, and you can go to your family without, you know, having a fear of ending up in the hospital for a couple of weeks, right? And so I think that's the biggest piece. These AI teammates are really about supporting humans, making humans safer, making humans more efficient, and augmenting human abilities. There's definitely no notion of kind of replacement or monitoring in that way. And so I think that's really been the focus. And we work and partner with safety teams, with operation teams at these customers to make sure they're well educated, they can bring their team along. And even when, you know, there might be minor resistance at first, once those teammates are on the ground, everybody gets bought in pretty quickly.
What do you do? Let's go to the operational example then. Do you have a story that shows the operational improvements that come from having these type of AI agents?
Yeah, absolutely. So, we're working with a very large clothing manufacturer, and, you know, they have tons of packaged clothing that goes down a line, a distribution line, right? And at different points during the day, they'll have higher volume. And they basically specify it as either, you know, green, everything's good; yellow, things are starting to get a little bit clogged, right? Maybe we want some intervention to prevent a red, which is full stop, right? Downtime and lost revenue. And, you know, by deploying an AI agent that can understand what a yellow scenario looks like, slightly higher volume coming down, simply with an IP camera that was already installed, they're able to have an agent immediately alert an operations manager on the ground. They can come in, you know, tune the system, tune the assembly line to make sure they don't get to a red state, right? And over the course of a year, that's going to save them tens of millions of dollars in lost revenue from downtime.
So, when we look at the future of what AI agents could be, right, we're really addressing safety right now. We're making it easier for the human to work. Where can AI agents go in the future? Like, what is that 10,000-foot view that we would love to see AI agents do five years down the road from now?
Yeah, absolutely. And I think the sky is the limit. As we talked a little bit about earlier, a lot of the newest models, foundational models, we talk about Claude or ChatGPT or Gemini, they are able to take in lots of different sources of data. So very soon your AI teammates will ingest all of the relevant business and video data, and you can simply give them an objective, a goal to move towards, right? Rather than "I want you to watch for forklift near misses and when you see them send me a text message," it's "I want to make this process 25% more efficient," right? And then, using all of those different data sources, reasoning, understanding and analyzing, you know, your AI teammate will do that work for you, help you plan, help you implement a strategy to actually solve that business goal. So I think we'll be there in the very near future. And if you look in the digital world, we're already starting to get there at enterprise level, right? If you look at OpenAI's ChatGPT Deep Research, it's a very similar concept, right? You give a broad, generalized kind of outcome. I want to learn about this topic, right? Or I want to try and solve this problem. It'll go off. It'll spend 15, 20, maybe even 30 minutes scraping the internet, reading, learning, absorbing, and then it outputs an entire report to you, right? And that's where we're going to be for the physical world very soon.
Awesome.
In the current, in the present, what do we do to get more folks to understand AI agents? Because I'm surprised every time, like, it's not like people, people are getting to understand AI, but their
default, and my podcast listeners hear me talk about this frequently now. The default is, oh, AI can help you write code. Oh, AI copilot will help you create ideas, right? Those are, and it's good that folks, you know, have some basic understanding of that, but I'm surprised every time I bring up AI agents, particularly in a manufacturing context, there aren't as many hands going up, like feeling like they understand it. So, what does it take to just educate people in the presence where this becomes more of an understood concept?
Yeah, it's a very good question. It's something that we've been grappling with as well, right? You know, I think the advent of ChatGPT and getting that into more people's hands has made the world of agents kind of more accessible to more people. So when we're talking to potential customers, we like to use ChatGPT for video, right? That's what Spot AI does, is ChatGPT for video. And I think as you start to explain that concept, it starts to click more, but it's still something we're figuring out as well. And yeah, I think we're going to be working to better educate a lot over the next couple years.
Awesome.
I have one final question. What did everyone order? I mean, this podcast is about to end because we are about to jump into what will be a very big Automate afterparty here at Extra Innings, which, Jake, this is the second time we've collaborated on this. What did you get, out of curiosity, Dunchadhn?
Yeah, this is called the Chella Familiar. It's a Mexican-style Vienna lager.
There we go. That's a great style. That's one of my favorites. I love craft versions of that style. That's probably what I'm going to have next. I went with a New Zealand Pilsner. New Zealand-hopped Pilsner. So, a little hoppier. Jake, what about you?
I got an IPA. I can't remember the name, but it had some crazy add-ons to it, and it is delicious.
I'll do my best to remember to find the name of that for the outro.
Yeah, I'll try and do that. I see a menu actually right over here. Real quick, let me grab this. Here we go. Right here.
Awesome.
I got the Crushian Tides. Actually, it's a pale ale. So, it's a New England pale ale that counts with Citra anchovi hops aromas and flavors with fresh citrus, red candy, and watermelon.
I love the description.
That is complex. I saw that. That one was a 12 oz as well.
Yeah. A little bit more. So, it's good.
There we go. Well, cheers, gentlemen. We got a party to get to.
Absolutely. Thank you guys for having me on.
Of course.
Article published