AI Transformation: Democratizing Innovation at TELUS

Automatic transcript of the recording (English); speaker labels are automatic. Timestamps open YouTube at that moment.

0:00Sacha: Okay, welcome. I have an incredible guest joining me today, Alexandre Guilbault, the founder of the TELUS AI Accelerator. Alex leads a dynamic team of over 65 AI engineers driving innovation across all TELUS divisions, including mobility, home solutions, health, and agriculture. And there is leadership. The team has not only generated over 75 million annually, but also set new standards for growth and excellence by leveraging advanced AI technologies and machine learning and technology. And he ensures that each solution is tailored to meet specific business needs, delivering tangible outcomes and setting industry benchmarks. Alexandre, thank you so much for joining me today. Welcome.

0:45Guest: Thank you. It's a pleasure.

0:47Sacha: What were the key business problems you and your team aim to solve by leveraging AI at TELUS?

0:55Guest: Yeah, right now, I think with the new technologies coming in, there's an opportunity to democratize those technologies even further than just centralizing them into an AI engineering team. Now you can use really natural language to produce codes, text, to analyze text. So I think the best ideas will come from having more people thinking about this. So our current challenge is to make sure that we can put those technologies in a safe manner in the hands of as many employees as possible. We really think that having 100,000 people thinking about the best way to use those technologies is much better than having a small team of experts in the domain trying to figure out how those technologies can really transform business students. So through the time we figured out that AI can be applied really everywhere. Data is everywhere in every company. A little bit more restricted to structured data before, so data that would fit in an Excel sheet, for example. But right now with those technologies, it's much easier to tackle unstructured data.

2:01So So if you have a network diagram, if you have images, if you have images, if you have text, you can start to ingest that and produce value out of this and create insights and accelerate people's work. So really, our goal right now is mostly to democratize and create even more value with more people thinking about those technologies.

2:20Sacha: Beautiful. Alex, could you share an example of a successful AI-driven solution that you guys implemented?

2:27Guest: Yeah, for sure. There are so many, so it's difficult to choose one. So maybe let me just share a couple quickly. And I will split them into two different categories because really there's a huge difference between classic ML solution and JiraDB AI solutions. So starting with ML solution across the business, we've been implementing many models through the last 15 years. For example, in finance, we have a hardware theft fraud detection. being able to detect whenever somebody is trying to steal somebody's identity and get phones, for in the store. So we've implemented those models. They've been running and creating a lot of value, both for the customer and the company as well. Nobody likes to being sold and having its identity sold. We also did a lot of things for sure with marketing. Those are classic use cases, predicting a likelihood to churn, creating recommendation engine to better serve our customer and recommend the right products for the right people.

3:31And in engineering, for decades, we've been forecasting capacity demand in the network, trying to predict whenever it will be congested so we can prepare solutions before we reach really those bottlenecks. So those would be a couple of examples of classic machine learning models that we deployed. In terms of gerundive AI, over the last two years, almost three now, we've been developing various co-pilots across the organization to accelerate work. Some of the top-of-line use cases would be IT support. We have an IT team that we need to call whenever we have an issue, like a password issue, or we need a new software. And getting access to them takes a lot of time. There's long queues whenever you call. So we created a co-pilot that can answer 70% of the requests from the people. So that has been really helpful and created a lot of value. We created field co-pilots for technicians that are on the field whenever they have a question about a standard or they need to have more guidance on how to perform a work.

4:38They can chat directly with a co-pilot that will help them and provide access to the right documentation. And we created co-pilot, for example, for call centers. So our call center agents right now have access to, I think, over 100,000 pages of information about all our services, our products. If you think, for example, of all the different devices we have, and we need to support with all the different operating system and upgrades and all the various plans. Some are legacy, some are new. We need to keep information of all of that. And it's really tough for our call center agent to be aware about everything and to find this information across the 100,000 pages they have. So we created a co-pilot that understands all of those pages. So somebody can just chat directly with the co-pilot and it will gather the right information tailored to the specific needs of a call center agent. So

5:32Sacha: that's fantastic. What would you say are or were the crucial factors that led to its success?

5:40Guest: Yeah, I think the most important of all is the stakeholder alignment. I'm often saying the best model in the world is useless if nobody use it or nobody change their process out of that. So sometimes I feel with new technologies coming in, people are so impressed by the technology itself that they forget the principle that has always been there and is still there, that there's no meaningful change without technology, people and process. And you shouldn't forget about people and process. So I think a clear focus on the stakeholder alignment to drive the people education and to change the process is really, really important to drive a meaningful change. And the other thing we've figured out through the years is it's better to identify the right problem and not fall in love with the solution. A lot of times when we we discuss with other teams and they brainstorm around where they could apply AI, they think a bit like if they have a hammer and they were looking for nails to it.

6:45And sometimes it's a little bit too siloed or too narrowed when they think about what could be automated in their task. And automation is something AI can do for sure, but automation will just save you some time. It won't drive a big disruption. It So we're right now really working on creating brainstorming sessions where we double down on the right problem to solve before thinking about what's the solution. So we're not forcing AI necessarily, but AI will always come. Right now it can do so much, it will come in the solution at some point, but we don't want to force AI in the solution. We're thinking really taking a step back and defining the right bigger business problem before trying to solution it. And finally, one of the key thing as well is, ties to the others I've been mentioning, but it's to be outcome driven and not technology driven. One key thing we're trying to bring as well is not trying to do POCs. POCs are really popular in technology, but a proof of concept finishes with technology concepts.

7:52It's not necessarily creating value. So we prefer proof of value. And a proof of value, you might not even have the technology concept completely finalized. A proof of value, for example, is doing statistics and validating something in there and making sure that you have stakeholder alignment to do something about the statistics. And once you've proven the value, then you can start to build technology. But if you start with technology, sometimes they can take a lot of time before you complete the technology. And then you need to test if the market will react correctly to that. So really, we're changing this mindset from proof of concept to a proof of value.

8:27Sacha: To your point, there are so many use cases that AI can solve nowadays. How do you approach prioritizing AI projects and initiatives?

8:37Guest: Yeah, that's a really good question. I'll put maybe Generative AI aside for this question because it's really two different frameworks. And to explain that quickly, Generative AI nowadays is already a pre-trained model that you're calling through APIs. So it feels more like a software delivery solution. For software delivery, you know you'll be able to success, but you don't know exactly the path to get there. So that's why we developed the agile methodology. Like you break down the problem into smaller chunks and every week or every two weeks in a sprint, you develop a path forward to get to the final solution without knowing exactly all the processes that will get you there. It's not a waterfall already, but you can plan at least for a week or two. I think in AI, classic AI, when you train a machine learning model, it's more R&D. So you don't know the path to get where you want to get to, and you don't even know exactly where you want to get to. You know there's something in there. You know there's data. And while exploring data, you might get to some insights or some solution that might be useful for the business.

9:42So given this aspect of more R&D, more risk, we had to define another approach to our work that we call fail fast. We're trying to ensure that whatever we tackle, if we are to fail, we prefer to fail within a day or two rather than failing after six months. Because there's a huge opportunity cost. Like you can be successful if ever you work not just six months, but two years on a project. You'll be successful after two years, but you might have lost like two years of other opportunities you could have worked on. So it's really important for us to tackle a project and test the weakest link as fast as possible. This weakest link could be privacy regulation. So let's test it if it's the highest risk of this project. It could be stakeholder alignment. Do we have really everybody aligned, ready to apply those technologies? It could be the data cleanness or completeness. So there's multiple chunks in there and we figure out what's the highest risk and tackle this one first so that we can secure it and move on or already kill the project and move on to something else.

10:52So given this fail fast approach, you can see that like our priorities are really dynamic. Through time, the priority of today might not be the same tomorrow depending on what we learn today. We learn every day. We learn on every step of the project. So priorities are really dynamic. They change. But the way we measure them is based on something we call risk weighted EBITDA. So the risk factor is really included in our calculation. If we have an opportunity of 10 million, but our success rate is of 80%, then it's no longer 10 million. It's 8 million at this opportunity. So something else might become higher and that changes dynamically through time. So that's how we prioritize our project, risk in the calculation.

11:34Sacha: You also mentioned a consultant-based approach to tailoring AI solutions. Can you expand on that? Can you explain what it is and how it works?

11:43Guest: Yeah, for sure. One way we like to describe our team is not really AI experts, but a consultant expert that happen to be really good with AI. So whenever we come into solve our problem, we're not trying to force AI into the solution. We're trying to really define clearly the problem and define And as I said earlier, data is everywhere. It will for sure be useful in the solutioning and there will be an AI solution eventually. But let's not think about that first. Let's think about business problems first. So that's a little bit the way we shape our team and shape our approach. We're trying to really ask why multiple times before we get to the right problem. We try to take a step back and understand the business process and map it before just optimizing one aspect of the process we're trying to think globally about the overall process. And again, as I said earlier, we're not trying to force AI into solution.

12:43We're trying to drive business value, not concepts. We're not successful just because we deploy AI. We're successful if we create value at the end for our customers, if I can say. And one other key aspect is really working hand in hand with the business. We're not working as a silo group that just deliver technology. We're creating squads, including them so that they can shape all bit and test the solutions we're developing as we're developing it. It's really interactive. We get something out and we test it right away to get some learnings out of it.

13:18Sacha: You know, new technology and AI is not accepted at the same speed by everyone. What have been some of the biggest challenges that you faced in driving AI adoption?

13:30Guest: Yeah, this is a really good one. I think resistance to change is one of the major blocks to really create meaningful value with AI. We tried multiple things through the years and we did some what worked for us. It might not be the best solution, but what finally worked is... Well, let me explain first what failed. I think it's important. Given we had this consultant-based approach, we were coming to the business with the AI solution, but also the people and process change that needs to happen. We were doing the change management and the AI. We were trying really to bring everything and drive value for all of our clients. And we figured out that there's more resistance to change whenever we come with everything. People like to think of us as an IT support team, a bunch of geeks that are really good with technology, but don't understand the business process. Like we're not... And indeed, we're not. We're not marketing experts. We're not finance experts. We're not RF engineering experts.

14:31But we have a team that is eager to learn and will ask the right question to get to the real problem at the end. And I think this approach of trying to solve it all for them created some kind of resistance. They were like, okay, yeah, you build good AI, but let me deal with it from now on. And that didn't really work because it wasn't clear who was owning the change management, who was the prime responsible of this process change and training of people. So what finally organically grew through the years with various business units and now we're adopting that as a framework for all of our engagement is to identify people that are being change champions in each group. So if we go to a specific business unit, let's say the marketing team, there's multiple VPs in there, but we'll find one VP that will be our change champion. So the VP with his team will be responsible of making sure that we really create value all of the solution we drive. They will be responsible of challenging the standard school. They will be responsible of integrating the solution in the process of people.

15:35And they will be responsible of helping educate the people about this new technology and what it's bringing. They are not officially called change champions. I think it's so different from one business unit to the other who will be responsible of that. Sometimes it's a project management team. Sometimes it's a strategy team. So it's always different. We call them change champions for us because this is what we need from that. This is what we are, the portion of the work that we want them to all to facilitate this transformation and minimize the resistance. It's been now two years that we've been adopting this and that has been really successful for us. So we really have a partner that inside of the business will help to drive this change.

16:19Sacha: I'd like to double click on your team. How do you structure your AI engineering team to foster like cross-functional collaboration? Like what specific roles or skill set do you think are critical?

16:32Guest: Yeah. One thing that has been key for us is to bring talent together and not isolate them in different silos. So I think traditionally in software deliveries and I have to say traditionally because I think we're even evolving from there. We used to like to structure teams as like you are the team responsible of maintenance and you're the team responsible of development, but that didn't work. So we moved to DevOps where the people developing are also maintaining, makes more sense. And we had the same kind of approach. We didn't want to split data engineers in one team, ML engineers in another one, and software engineers in another one that like you need to bring the product from one to the other, expecting some delays, having intakes, that blocks, not being able to be as agile as we would like. So every squad that we create have all the right talent and have the same kind of talent in each squad. So currently we have two squads responsible of the customer solution, two squads responsible of B2B solutions, including health and agriculture.

17:34We have one dedicated for our network engineering team, one created for our finance team. And each of those squads are only assigned to a group because we want to facilitate communication. We didn't want somebody from finance to come and give, and depending on the workload, like refer them to a different manager or a different team, or else they will get confused at some point. It will become really complex to understand like who's working on what at what point. So it's the same talent in each team, but we just facilitated communication by adding squads dedicated to a specific business unit. So people in finance will get to know exactly who's the manager, who's the team, and they know who to reach out to. But each team has the same talent, and we have mostly people we call AI engineers. We've changed a little bit through time. We call them data sciences, machine learning engineers. Right now we think AI engineers is the right way to describe them, given that engineers try to take something that is more raw and create value for business. So that's the role of an engineer.

18:35If you think of an electrical engineer, for example, they take electricity and create value with electricity for business. We think AI engineers is the right way to call them because they take AI and bring value to business with it. So we would have per squads around six to eight AI engineers, always in the IPM or product owner that will own the solution and on the communication with the business units. One data engineers, the one that engineers per team just to facilitate creation of pipelines. They are not necessarily enterprise grade at this point. Like it won't be, we do have another team that is like doing enterprise grade data engineering, migration to the cloud. But inside of our squads, our goal is not to make something enterprise grade, it's just to get to solutions fast. Like sometimes hacking our way through the solution just to prove there's something. And if it's successful, then we will move to enterprise grade. So we really wanted to have data engineers inside of the squads so that they can help to streamline and and deploy something fast to test the market. And we have not always, but we're trying to have more and more software engineers.

19:39So pure software engineers that happen to be good with AI or be interested in AI. But that will complement a little bit the integration of our solution so that we can have a visual interface or integrate that into another visual interface or creating APIs. So software engineers are also really helpful in those quotes. Maybe just to complete, you also asked like what kind of skill set we're looking for. I would say an interesting thing is a lot of people are thinking AI is really special, a specialized skill. Like you need to have PhDs, people that have studied and work really for a long time in that domain. We took a different approach. And that's partly because we're not doing pure research. We're doing great things to develop and create value. So our approach has been mostly centered around soft skills and not even technical skills. For us, this domain is moving so fast. If you don't learn every single day in six months, you're a dinosaur.

20:42So even if you hire the best person in the world, but he doesn't have the curiosity, resourcefulness, humility to continue to learn and question others around and read, they will be out of the game within six months and a year. So for us, it's much more important to have those soft skills, communicating well with their peers, sharing, being curious about new stuff, being resourceful. I think those soft skills have really helped to center what kind of talent we want in the team and people that are really good on those skills have been extremely successful in creating value and keeping themselves up to date.

21:24Sacha: Honestly, I couldn't agree more. I always think and I always say to people that hard skills, you can always learn. Soft skills, a little bit harder. So I think I really agree. Earlier, you mentioned that you try to always ask why and focus on business outcomes. Can you elaborate on how you instilled that mindset within your team?

21:48Guest: Yeah. One way to instill this mindset has been to not own this as a leader in the team, but empower the team to do it themselves. We have been doing over 150 projects per year. And if everything needed to go through me and I was the one like taking decision there, I would be a huge bottleneck. I don't think we would be able to deploy 20 or 30 projects a year. So I'm really trying to manage this team as a bunch of intrapreneurs. Everybody in the team is an intrapreneur and they are building this team. We're building this team together. We're really not hierarchical. The success is everyone's success. We are all successful if our team is successful. I think it helps now that AX Raider is getting known more publicly and we share about it. I think it helps everyone in their own character and they are proud of what they're doing. So really cultivating this intrapreneurship inside of the company, giving them enough room to try stuff, to fail at stuff without like...

22:52Actually, we're encouraging them to fail. It's good if they fail. They shouldn't be too hard-headed and try to always be successful or else they're not innovating. So really this culture of being intrapreneur, being ready to challenge status quo, to innovate by themselves has really helped us. And I think that that's one of the key points that made us successful.

23:16Sacha: What advice would you give to other companies that are struggling or looking to embark on their own large-scale AI transformation?

23:25Guest: Yeah, if they're starting, I think it changes through time depending on your maturity. So this is not something that fits every company. But if you're starting from scratch, I think you need centralization of AI. Sometimes companies are trying to hire data scientists here, data scientists in marketing, and data scientists in finance. And they are all split inside of the company. And in this field, no talent alone is able to stay up to date with everything. Personally, I'm also trying to stay up to date. I'm reading every day, 30 minutes to an hour, and I'm not catching everything. Like I would need to read three hours per day to be updated on everything happening. So I think it's really important for AI engineers and data scientists to be surrounded by other talents if they want to grow and to enjoy their work. Because they won't like it if they are isolated in a marketing team, for example. So I think really AI talent attracts talent. So centralizing will help any business to build this team.

24:26Or else if they hire just one or two, they will probably lose those people after six months, a year, and they will start from scratch. So it's really important to assemble all of them together. And also, I think, to put more emphasis, as I said earlier, on soft skills. They need to find curious and go-getter people over experts know-it-all that will tell people, like, they got it all figured out. Finally, the other thing that is really important to get started is to encourage failure as long as this is fast. Sometimes I see this bias in companies to finance a project, but then expect a return on this project and look at it that way, like one project at a time. Whenever you do that, for example, you put one million up for creating an AI solution in a company, and you're looking for the return on this one million. So if you're looking for a project, you're looking for a project to not see return. There's more chance that a project will fail in data science, that it will be successful. Failure is due to various things, as we discussed before.

25:28It's not just the technology. It can be the data. It could be privacy, regulation. It could be security. There's multiple... Or change management is the biggest factor. But there's multiple reasons why a project might fail. But some studies have proven that 80% of the data science project in companies fails to deliver value. So that's four projects out of five that will fail. Finance thinks about funding one project and expects a return on one project. They are more likely not to see the return, and they will stop funding AI. So having a mindset of failing, and it's okay to fail 80%, even 99% of your project, as long as you fail them fast and don't have a huge cost of opportunity. If 1% of your project is successful and drives millions, then you're successful as a team. So it's important not to look at project, but as like a timeframe and value invested and testing out multiple things. It's more R&D than development work.

26:23Sacha: I'd like us to look into the future a little bit. What emerging AI trends do you see or are you more excited about?

26:32Guest: Yeah, it's always tough to predict the future in that field. Like we're always surprised by something that comes out of the blue. But there's something clearly happening right now. So is the right time to ask that? We've seen two years ago a big new wave of AI. So before that, it was mostly surrounded around predictive AI. So building machine learning models that predicts an outcome or forecast or identify categories. But it's mostly prediction. Two years ago, we started to enter a generative AI world. So generative AI is helping us to use natural language to ask requests from an AI. It helps as a human-machine interface. It tackles structured data. It can do multiple new tasks that we couldn't do before. And we're still just at the beginning of this. I think there's way more value we can create on generative AI. But even then, I think we're already entering a new wave. Like it's accelerating, it seems.

27:33So we've been just two or three years in that generative AI era. And we're already starting to move into this new era. And I want, for my team at least, to have generative AI as a mature technology that everybody can use. And start to think ahead with this new wave that is coming in. That many are calling agentic AI. So what agentic AI is, is generative AI that won't only act as a counselor. It will just answer a question whenever you ask it. But it will actually perform action for you. So today, you can use a chatbot. It will give you a lot of information. But that's all it can do. Agentic AI will be asking a bot to do work. So AI will move from being a counselor to a peer. That will be like everybody will add their peer as an AI, per se. Maybe multiple peers. And we'll be able to say like, I need this access and this code there. Just go to it and come back whenever you have done it. And the AI will actually be able to write some code, get access, do some tasks, test and file some time until it finds a solution.

28:39And then come back to you with a solution. So this will be really a big revolution in everybody's work. Like it will change really the way we interact with the machine. We'll transform all the work. But it's still really complex. The framework to get there is really complex. How do you ensure governance over those models? How do you monitor access of those models? How do you ensure that there's no failure along the full process that you won't be controlling anymore? Software, you really control it with APIs and gateways and all that. But an AI that will sound like that can make mistakes. It's still a predictive model that runs behind. So there's still a long way to go in that genetic AI. But already today, you can start to create, to give access to some tooling for the AI and you can start to perform some small action. So I would say that's what I'm the most excited about and that's what will drive the biggest next wave of AI technologies.

29:35Sacha: That is really exciting. Alex, I'd like to give you the floor to share any last word with the audience.

29:42Guest: Sure. I think right now we're poised to see big, unprecedented change in our lives and in our work lives as well. And we can't really predict how it will go. So I think the best way to prepare is to find the right talents and foster the right culture. Given you don't know exactly what you'll need, soft skills that are more applicable to various different fields are much more important. As I mentioned before, I personally think curiosity, flexibility, resourcefulness are some of the most important qualities your employees can have today to be prepared for whatever change that will come. Another thing that I've been voicing a lot with my management team is we're not just innovating in terms of technology, but we need also to evolve in our management practices. We won't be able to innovate with teams and create new stuff that we've never seen before if we're not ready to do new stuff that we never did before in the management.

30:49If we structure too much a team, we're killing innovation. So we need to be flexible. We need to reinvent ourselves in management in order to reinvest ourselves in technology.

31:01Sacha: Great. Alex, thank you so much for joining me today. It was a real pleasure.

31:06Guest: Likewise. Have a good day.