Discover how Deepomatic is transforming industries with cutting-edge visual automation! In this insightful interview, Thomas Thuillier, Managing Director for North America at Deepomatic, sits down with Alejandro Piñero to discuss how their AI-powered image recognition technology is revolutionizing quality control for ISPs, telecom, utilities, and more.
From automating millions of fieldwork photos to ensuring first-time-right project completion, Deepomatic helps companies scale efficiently while reducing errors.
Thomas also dives into why the time is NOW for ISPs to adopt this technology, driven by 5G and fiber rollouts, federal funding initiatives, and labor market pressures.
Whether you're looking to optimize large-scale deployments, improve decision-making, or simply understand the future of visual AI, this conversation is packed with valuable insights. Tune in and learn how Deepomatic is setting the gold standard for precision and efficiency.
Alejandro Pinero:
Welcome everyone here on Fierce Network, another digital interview, and I'm very excited today to be joined by Thomas Thuillier. He's the managing director for North America at Deepomatic. Thomas, welcome. Thanks so much for joining us today.
Thomas Thuillier:
It's great to be here. Thank you, Alejandro.
Alejandro Pinero:
Excellent, Thomas, let's start at the top here before we get into the detail. A lot of people will be familiar with Deepomatic. Some might not be as familiar as we want them to be here. Can you tell us a little bit in broad strokes about your solution and why you wanted to chat with us today?
Thomas Thuillier:
Yeah, of course, Deepomatic, for those who don't know, is a software provider that focuses on what we call visual automation. To put it very simply, it's what we call image recognition. It's a branch of artificial intelligence that deals with visual data. Basically, we can train agents to be able to automatically recognize visual patterns on photos or videos and we have been working on this technology for over 10 years now.
We are really the expert in that field, but we also have extended experience in the telecommunication industry and a little bit more broadly in everything that deals with distributed networks, utilities, such as gas, electricity, water. For the past six years, we've worked with a group of internet service providers, infrastructure operator as well as their vendor, the construction companies, operation companies across Europe, North America, US, Canada as well as a couple of companies in Latin America.
We originally are from France and we've worked with all the major providers there, such as Altice or Bouygues Telecom. Other famous names are going to be Vodafone, Telefonica, or Swisscom to name only a few. Why do we do this? I'd just like to give a quick scenario for you to understand exactly what that means when I say visual automation in the telecom industry.
Imagine you're an ISP and you have either an internal workforce or maybe a deep chain of subcontractors who perform work for you. They might be upgrading your network, building a new piece of your network, or even just connecting subscribers. You really want to make sure that first of all you can have an idea of if the job was done, if it was done properly, and also to be able to do that you want proper reporting.
You want to know that they installed the right equipment in the right way in the right place, et cetera, and this is very challenging when we're talking about those very large distributed assets that spans miles and miles and miles. It's almost impossible to imagine people just going out and check everything. What's been going on before our time is asking those resources to take photos of their work as much as possible, standardized photo for the same type of work so you can check the work.
Now if you're a Comcast or even in France, some of our customers take tens of millions of photos per year. You will not have the resources even if you hire 10, 20 people to manually look at them and that's really what Deepomatic is. We train those systems to instantly validate that the photo was taken properly, very important for your systems of record, and in the photo validate that the job was done properly. Really that's the gist of it.
Alejandro Pinero:
That makes a lot of sense. Thomas, can we tell me then why is now an important time for these ISPs and carriers to be thinking about scaling their quality control? Now you mentioned they're big companies, a lot of scale, but why should this be front and center now?
Thomas Thuillier:
That's a very good question. The key is that right now the telecommunication industry is really undergoing a transformation in terms of the technology that's being put out there. There is the 5G rollout. There is the fiber rollout, but even more importantly, especially as we're talking to a majority US audience, we have large federal funding that are going through, such as the BEAD for broadband access to underserved community. That really puts some sense of urgency in that deployment right now.
On top of that, there is also requirements for those works to be done on time and in budget, which is becoming more and more challenging as we have faced a period of inflation. We also have pressure on the labor market with fewer qualified technicians available. If you combine all of this, really you want to make sure you have the right tool to be efficient and quick.
One thing I like to tell my customer is whenever you have to send a crew to just even just document a job that was not documented properly just because you have no idea if really you put your asset in the right place, it's a crew that's not building something else. If they're going out to fix something, again it's a crew that's not building something else. You're looking at a lot of resources mobilized on repairing things, redoing things that should have been done right the first time.
I'll just put my little plug. Deepomatic really is about that first time right approach, first time right automation, but even beyond this, that's really the first layer of why this is important right now. The second layer is since we're deploying these new networks, it's really key to create a system of records that you can trust and we've seen that in the past. We've seen that in countries that have deployed their fibers already previously and that do not know for sure where the asset is, in what state it is.
When you're talking about just operation, regular, connecting your subscriber, but maintenance, if you have any events that requires you to go out, if your connection point is on the wrong side of the house where you put it on your system, or if your street cabinet is on the wrong side of the house, that can create a lot of trouble.
Deepomatic, with the fact that every single photo will be looked at by an AI that you can trust, and obviously we have ways for you to test that the output of the AI is trustworthy. That means you can have a tremendous level of confidence in the data that comes in from the field that you put in your system so that down the line you will really be able to make decisions, business decisions, operational decisions, that rely on something that is true and real.
Alejandro Pinero:
I mean it's pretty clear what the benefits are here. I understand the complexity and what the pressures are that drives for this to be important now for these carriers in particular. I guess could we talk a little bit about how the solution specifically works and what makes it different to perhaps other ones that are out in the market?
Thomas Thuillier:
Yeah, that's key for us right now, especially since we hear a lot about other systems like the GenAI, et cetera, that is able to handle a lot of different things. I just talked about trust and I think that's really the key when we're talking about those large systems. When your confidence level in the business output that the system gives you in terms of, yes, this job is done properly, a couple percentage points is going to be driving in terms of volume. It's going to be extremely large, statistically significant.
The approach of Deepomatic, and that's where our deep R&D background for 10 years really is key, is we have the technology platform to be able to train those very specific systems for the operation of our customer. Right now, we have more than 10,000 specific visual checks that we've specifically trained. Those things can be done on our platform because we can do it quickly and we can do it efficiently with not a super large data set of images. That's why it really works.
The first thing is we hear a lot about the GenAI. It does great things. It can help on the field if you have little questions. When we're talking about a systematic check of what is done where you want to be able to consistently trust the output you need something that we can say is a little bit more deterministic where really the AI is trying to give a yes or no answer on very precise points and not give it a general idea. That's going to be the first thing, but this is fairly new. It's on everybody's mind.
That's why I wanted to talk about it first, but if we look really at what people are doing right now, I would say in the US there is a need for transformation of the way fieldwork is handled. A lot of the time we see people still rely on a non-digitized tool and we have a lot of pen and paper. I'm not saying fundamentally it's bad. We have a lot of trust going in between people in that industry, which is great, but we can go one step beyond.
I think the obstacle for the digitization of this industry has been that people are aware that if I start asking my crews to document their job better, take more photos, like I said earlier, it's just going to create a glut of data that is really not very useful, especially since you contrast it. I was talking with a large East Coast provider recently who was telling me, "We were looking at our locate data in terms of making sure that people properly mark the ground and we have a backup that says, 'All right, we know where stuff is,' and we really marked it for future utilities that might do work around or in the area."
They had an issue. I think someone cut some cables and they're like, "Okay, let's look at what we have." They were telling me all their data set was bad to use not a bad word, but not usable, missing data, incorrect photos. Then, you're like, "What can I do? I have nothing to back my claim." Deepomatic is really the missing piece in the equation to say, "Now take the photo because we will look at every single one of them." The key difference I talked about the first time is you get that feedback instantly because you could also say, "I'm going to hire 50 people, 100 people and look at this photo," but that takes time.
Now the technician does their work. They take the set of photos that's required from them. If they're new, we can also tell them, "Look, you should take the photo this way," but they interact in real time with the system and they're human. They make sometimes a little mistake. We're not in the business of policing people, but they might just make slight incorrect labeling of the fiber or maybe they do something slightly off. We'll be able to make sure they get notified and they fix it before they leave the site. That's really the biggest difference that today no other system can provide.
Alejandro Pinero:
Sure, and I wanted to pick up on a couple of the points you mentioned there. First is you mentioned about training and identifying these images. Do you use this with your own algorithm? Is this something that you can use off the shelf? How does that work for Deepomatic and also how long does it take for that to go live? I think you mentioned you need a pretty small sample size, but can you maybe expand on that a little bit?
Thomas Thuillier:
Right, in terms of technology, we use our proprietary AI training systems, which I know for some of the biggest national ISPs it's going to be important to know that we're not just taking your data and putting it into somebody else's system. We have our own AI infrastructure. I'm not going to go into the details, but we really use what is at the top of the market in terms of neural network architecture. I will leave that technical speech to somebody more proficient than me.
When it comes to your question from an operational point of view regarding off the shelf models and training timeline, et cetera, like I said, we have an extensive knowledge of visual checks that are performed on this type of work. Over 10,000 very specific checks have been deployed in the past. We have a lot of pre-trained models today available for us to quickly put something in production in the hands of your crews in order to test the AI and accelerate the deployment.
What I'll say is even between two different providers or construction companies that work in the similar region even we see a lot of differences visually. It's going to be slight differences in just the environment, the light, et cetera, but also the equipment, the processes. Our approach has always been that the best visual automation system is going to be the one that is trained on your data and that was always at the heart of building our platform.
To train or tweak the existing systems to work very, very high level on our customer's data, we need sometimes a couple hundred photos, a couple thousand photos. That's something that goes very quickly even if that data set doesn't exist. Most of the time it already exists. When we're talking about data acquisition we're not talking year long program of generating data. It's very quick and painless. When it comes to training the AI that's even faster.
Our infrastructure technologically speaking on the software allows us to train those systems very quick. The last part in terms of deployment is going to be integrating Deepomatic into your system. Deepomatic is going to have the biggest impact when it ties into your field management software. Already if you currently have online forums on the mobile device of your technicians, we don't change their practices. We try to make the change management as light as possible.
Now I'll just say, as a side note, Deepomatic can be used as a standalone. You do not have to have a field-ready mobility tool, but it's also important that we connect to the back office tools, to the GIS, the OSS, so that you can leverage in the most efficient way possible all the data that's being generated and also to tell us what we're supposed to be looking at. Again, Deepomatic has been built in a way where building these connectors through APIs is very easy and we can always just download simple CSV files with all the data that we can then plug into. Even that part can be quick.
To conclude, I'll say the long end of a deployment is going to be towards six months taking into account both having to do some data collection, having to do an integration both on the mobility side and the back office side, having to retrain a certain number of systems. To have a fully complete system, at the most it's going to take about six months, but even over this six months you will have first levels of APIs being deployed gradually.
You already generate some of the value. What I mean by that is the easiest step for us is to put in the field the systems that will validate that the photos are done properly. Just that already has a lot of value. Then, as we move forward, we can start training the AI on the specific visual aspects in the photo that says if the job was done properly or not. All in all, you get your value very quickly and the return on investment starts to be positive after just a couple of months.
Alejandro Pinero:
Thomas, let's close out here perhaps putting you a bit on the spot, but you mentioned those 10 years of experience for Deepomatic, an impressive list of clients in Europe and in the Americas. Could you maybe put everything that you've talked about here into context and maybe an experience that you can share of one of your customers with some tangible results?
Thomas Thuillier:
Yes, it's true that we haven't talked about the specific metrics yet. I'm glad you bring this up. I will try to follow a stricter thought here, not to lose everybody. I'll say the first metric that we see is if you're a customer, and we've had some of them, that currently require a photo being taken of the field to validate the job and document it, the first large impact of what Deepomatic will do is dramatically rise the compliancy of that reporting.
Today, on average our customers across 30 odd customers from again ISP, construction company side will have 97% of the data coming from the field correct, being properly recorded. The interesting part is that the 3% that are not correct, we know about them because the eyes looked at them and said, "This job's not done properly." We know. All in all, you have 100% of your job where you can make a correct size assessment, which means that even that 3% you can address. Reporting quality is really improving.
The second is we've talked about productivity efficiency. The biggest metric here is going to be the rate of truck rolls or revisit repairs, however you call them, to have to redo a job that wasn't done in the first place, a job that you first of all had to even be aware was not done in the first place just to remind that you might not even know. Here we've seen across our different customers, and we've done, it was barely over 12 months ago, an independent study with Forrester who's a leading market researcher.
They have what's called a total economic impact. They look at four or five of our customers and generate somewhat of an anonymized result just to not divulge potentially classified, sorry, proprietary information. The result is that on average across OSP build, network deployment, service activation we really work in all these areas. You'll have about 10%, 12%, sometimes 15% of the jobs that are not done right the first time where you need to send the truck rolls.
Deepomatic in the first three years eliminate almost 60% of those truck rolls. Considering that depending in your region you're talking about a truck roll at maybe $250 for something suburban, light touch repair, we're not talking digging back up something, to upwards of thousands of dollars if you're in more rural areas, you have to drive three hours out to do a job. Again, that's potentially millions in avoided costs annually. The interesting part is that the 60% reduction is what's attributed directly to the AI, meaning the AI caught a problem, got it fixed before the technician leaves.
What the report pointed out is that if you just compare pre-Deepomatic, post-Deepomatic, the reduction is even bigger, up to 80%, 85%. The reason why the reduction is even bigger is you change the behavior of the people on the field. Not to say that we're not doing properly, but now if someone before with the best intention in the world does a job, but makes a mistake that might be obvious, but doesn't realize it, and there's no proper documentation system, they might put it back in.
Now if they did a job and make a mistake that seems obvious, they will see it before they take the photo if it's very obvious, right? Nobody in their right mind will take a photo that gives you real-time feedback that says, "Hey, you forgot to plug the cable," if I see I didn't plug the cable, right? They'll fix it even before they take the photo, which means that that doesn't actually count as AI catching it, but you really instill that different mindset in terms of how to do the job. The impact is extremely big here.
The last one, we look at everything, a higher rate of compliance for reporting. You have higher first time right rates, a lot less resources spent on repair and maintenance. The third one is today you do rely on a lot of your internal resources for manual quality control in some form because everybody has to do some form of quality control. That might be having a dedicated team. That might be having some of your most proficient resources, managers, spending significant amount of time, several hours per day, looking at what their crew is doing just in their truck, sometimes on their computer just going through documents.
That's something that you can just eliminate in large part since the AI will do it automatically. Now it's not to say you don't want any human in the loop, but now we were talking about that 97%, 3%. Now it's the AI saying, "Oh, can you please look at these 3% because I think something's wrong?" That doesn't require much resources. Even for our biggest ISPs, we're talking sometimes one FTE. I'm saying most of the time it's less than one FTE spending time interacting with the feedback from the AI to make sure that everything was done properly.
Really to conclude, when you deploy Deepomatic what happens is you really see a whole range of benefits, but the first one is just going to be the money you spend on quality control gets slashed. Right there, that already pays for the cost of Deepomatic. On top of this, you spend less money on repair and control and then you see an acceleration of your work. You have gains in productivity, reduced time to revenue, and an edge on your competition. All of these means that the impact is extremely important.
Alejandro Pinero:
Brilliant, it's hard to argue with data and statistic points like those ones. Thomas, thanks so much for sharing. These are topics that we've covered at length here on Fierce Network and through Broadband Nation, our sister program, as well around the complexity and the cost associated with some of these deployments and sending crews back out and the cost that involves. It's great to see the Deepomatic AI solution really making a difference there. I really do appreciate your time joining us here to explain and put that into perspective.
Thomas Thuillier:
I do have, if we have one more minute, I realize you were asking about maybe a specific story and I have one to share. I can't necessarily name names, but I'd like to give-
Alejandro Pinero:
Sure, we'd love to hear it.
Thomas Thuillier:
It's an anecdote, but always puts you in the mindset of the impact that you can see. One of our longest customers in France had been trying to put a systematic quality control system in place that was very thorough, but obviously without the automated system. We're talking here on the service activation side. In France, it's a little specific. I won't get into the detail, but those ISPs do a lot of them.
There's a lot of change of service. Several million service activations per year requiring up to 10 photos per service activation, that's tens of millions of photos being generated and they work with five different vendors. Now you have a crew of about 30 people who spend almost all their time trying to look at about 5% to 8% of the sample of photos to try and determine is each of my vendors meeting the requirements of the SLAs?
They come to us and they tell us, "We have one of those five vendors doing a terrible job. We're in a fight with them. We want them to pay penalties, but it's very difficult." We deployed the Deepomatic system across 100% of the jobs, not 5% or 8% sample. After running for a couple months, the conclusion is that the one vendor they had been looking at was actually doing a perfectly fine job compared to the others. It's just that at some point if through your sampling you think someone is doing something bad, you put more effort looking at them.
You find more things that are wrong. There was one of the other four that was actually doing significantly worse than everybody else and they had no idea. I think what I'm trying to say here is that tool benefits everybody in the ecosystem if you're obviously working in good faith because we have some customers now who have much better relationship as the construction company with their ISP and vice versa. Some of our most advanced companies as customers have integrated Deepomatic in their billing system.
What that means is, as a vendor, we all know there is a lot of pressure on your working capital. It's difficult to manage ISP. Sometimes that takes six months to pay an invoice, but at some point if the ISP can have a system that tells them, as a trusted third party like Deepomatic is, this job was done properly, instantly some of our most advanced customer then automatically pay the invoice for that vendor. That means the vendor has more working capital, can hire more people, can take on more jobs, build faster. It really is good for the entire ecosystem. I wanted to conclude on that.
Alejandro Pinero:
No, that's brilliant. It's great to put it into perspective and I think that'll really put it into a specific use case for everything we've heard here. Thomas, thank you so much again for joining us. I certainly have learned a lot and have a lot of ideas now about our reporting and where we need to go a bit deeper on. I really appreciate it and I'm sure our audience also found this enlightening. Thank you so much for your time.
Thomas Thuillier:
Always a pleasure, thank you for having me.
Alejandro Pinero:
To you, our Fierce Network viewers, thanks so much for joining us. We'll be back soon with another interview coming your way. Until then, take care, bye-bye.