Charlotte Ward: 0:13
Hello and welcome to episode 175 of the Customer Support Leaders Podcast. I'm Charlotte Ward. Today I'd like to welcome Taddus Labudis for a fireside chat. I'd like to welcome to the podcast today, Tadas Laboudis. Tadas, it's lovely to have you join me here today for a fireside conversation, which I'm very much looking forward to because it's speaking to a lot of my pain and challenges right now. But we'll get onto that in a minute. First of all, would you like to introduce yourself?
Tadas Labudis: 0:52
Hi, Charlotte. Thanks for having me on the show. I'm very happy to be here and talk about a subject that's quite close to me. My name is Tadas Labudis, and I'm the CEO and founder of Proudsite. And just a quick uh kind of background me, uh in my past life was a product manager, uh, and I came up with this idea to help teams understand their customer feedback, tickets and reviews, and so on. So I launched ProdSite about three years ago to help companies understand their customers. And you know, vSense developed the platform to automatically tag tickets, reviews, and other feedback, and developed some of our own uh AI algorithms to help do that automatically. So uh along this journey, I've tagged tickets for other companies uh for free, being paid for it, uh seeing how hundreds of different companies tagged their tickets and structured their taxonomies and actually you know participate in building some of the systems to automate this process. Um so um yeah, you know, would love to chat more about it.
Charlotte Ward: 1:56
That's awesome. Um, well, I'm I'm very excited to hear that uh as part of your journey you were out there tagging conversations for companies for free. Um I guess I guess that's every support leader's dream is like just to have someone kind of swoop in and sort this out for them for free. I can I can but live in hope that that will happen one day. But nonetheless, I gather along that journey you've you have uh gathered a huge number of learnings around tagging, and that's what we're here to talk about today, isn't it? It's about all those challenges around tagging and extracting, and they're thereby extracting like insights and knowledge from our customer conversations.
Tadas Labudis: 2:41
Absolutely. Um yeah.
Charlotte Ward: 2:44
So so I'm really looking forward to this. Um, I um I have a number of challenges of my own in this space, and I know everyone else will for sure. Um, I think that the place I'd like to start is so often the place I'd like to start, it's at the beginning. What as you uh if you're new to particularly to support leadership, I suppose, is likely to be the first time you've come across having to kind of set up and and and wrangle tagging in any kind of form for the first time. If you're new to this process, if you're new to understanding tagging as a concept, tagging as uh like how you operationalize on it and and everything else, where do you begin? Where do we start with tagging?
Tadas Labudis: 3:30
Yeah, it's a great question. And uh I think you know, with uh it might sound a little bit cliche, but with a lot of things where there's a lot of effort involved and it's a continuing process. You don't want to start with, okay, like what tool do we use for tagging, or like you know, what tags do we set up. You need to kind of start with why are you tagging tickets in the first place. So of course, um tickets are coming in, whether you like it or not, customers have problems, stable contact user support. Tagging is kind of an auxiliary task uh that's there for um that teams run for a variety of reasons. So it might be that you want to understand the contact reasons, why are people getting in touch? Um that might be helpful in helping you understand what drives the volumes, how can you manage it better, possibly inform your self-service strategy. You know, if you're building a knowledge base or you want to update it, um, that might give you content ideas. Uh, if you're working with a chatbot or some kind of automation, uh, you might want to use that to inform what you know what the first 10 responses are gonna be, or even the long tail. And uh it might also help you inform what kind of training you need to provide to your agents. So if you're seeing that a lot of your queries are about billing, maybe you want to train your agents more on those kinds of queries so they can handle them better and quicker. Um, and maybe uh another point related to the contact reasons could be forecasting. So, what agents do you want to put on uh in different ships? And you could use that data to align the balance of your team at any given time to make sure those questions are answered. Um, so this is kind of the main reason we see company stack tickets. Uh, but it's important to understand if that applies to you and whether you know that's something you wish to do and invest time in.
Charlotte Ward: 5:20
That's that's that's a really interesting point, actually. Just do you even want to bother? Because it's a lot of work. And and if you're gonna put any effort in, you want to you you need the data that you get out of that, whether uh by data can seem a big word in those early days, I guess, if all you're trying to tag is three or four common themes through your tickets to begin to, as you said, maybe coach to them or to put process issues or whatever. But yeah, it's even worth it in the first place. That's a great place to start. Let's assume it is. Let's assume that you don't have the luxury of deep diving on every ticket and building this kind of super uh intuitive understanding of your customers' needs and challenges and your team's needs and challenges. Um, let's assume we need the system to do some of that heavy lift lifting. So you you touched on a few things there, and I think I think that that was interesting how you didn't it you you you really it was really about keep it simple, wasn't it? It it was just there's no point in overburdening this early on with trying to answer complex questions that you probably don't need if you're just getting going.
Tadas Labudis: 6:32
Yeah, absolutely. And you know, if you're starting with tagging, um, you know, but like with most things, you want to start small and then expand from there. Um and it might just be tempting to start tagging tickets right away. Um, but it's actually worth spending a little time up front thinking through your tagging strategy. So otherwise you you're gonna end up with uh with a mess that you know it doesn't give you what you want relating to the first point. Why are you tagging? And you might waste a bunch of time and actually associate uh tagging with bad emotions and kind of bad vibes and not do it in the future because it didn't work first time. So I think it's important to start with your goals. What types of insights are you looking to uncover? Uh what actions will you take if you uncover these insights? Uh so kind of just running through these scenarios in your head. Like, okay, if I know that 25% of my questions are about people updating payment details, like what action would you take? Will you write an article? Will you contact your product team and ask them to create a self-service solution for that? Uh will you uh provide automated responses to those questions? You know, running through those scenarios will help you justify each tag that you have in your taxonomy.
Charlotte Ward: 7:47
Yeah, yeah. Um, when I think about uh that sequence of like not only keeping it simple, but also understanding the kind of things that you might be trying to capture and therefore the kind of actions you might take. Um the the thing that I that always rings alarm bells for me on these kind of journeys is where when you're very early stages, how you scale something that you started simply. And I think you know, if I had a tag that was billing, for instance, billing problem is my tag. Um how do I as as we grow um get more insight from that? How how and when is the right time to actually layer on complexity and how much of that how much of that complexity should you try and build in early on while trying to maintain that balance of keeping it simple, which is like a mind, like it's it's really difficult to wrangle that concept in your head that you want it simple but also extendable and and future complex, right? Potentially at some point.
Tadas Labudis: 8:57
Yeah, so I think um when it comes to building the actual taxonomy, one one point that I should have mentioned is even before you start building that list of tags, it might be worth thinking about how you're gonna create the reporting, what an ideal report is gonna look like. Uh, you know, do you have the capabilities to create a report in your support tool? Uh like how does that work? Maybe running even a small scenario of trying to apply a few tags, see how they trickle down into reporting and whether that report is gonna be useful. And then also, like, when are you gonna do it? Are you gonna do it every week, every month, every quarter? Uh, how much time is it gonna take? Just kind of thinking through these practicalities before you tag because that could impact how you structure your tags, you know. For example, um there's two types of uh tagging taxonomies that we see. We have uh flat taxonomy, which is more like a simple tag list, and that's how you tag things in intercom, for example. There's it's a very flexible system, but at the same time, it doesn't have the hierarchical structure. Umorical structure allows you to create tiers of um tags that go kind of deeper and deeper uh and become more specific as you control down. So you start with something abstract like billing, and then maybe inside of it you have update payment details, churn, pricing queries, things like that. You could maybe have even a third tier or four tier if you want. But so obviously, right at the get-go, it seems like the more detailed, the more insightful that taxonomy is, the better, right? It feels like it's gonna be a better output for everyone in making decisions. But this this the other side of that coin is when it comes to tagging, I think consistency is what trumps everything. So I'd much rather have five broad tags that are applied consistently to every ticket that comes in, categorized into those five categories, than a hundred tags that have possibly time gaps when we forgot to use certain tags, or inconsistent use between agents, as they uh you know don't know what each tag is for, or maybe can't even keep a hundred tags in their in their memory as they're working through tickets. So the more tags you will have, the more difficult it's gonna be to maintain the consistency.
Charlotte Ward: 11:20
Yeah, yeah. I think I've always erred towards trying to do something on the hierarchical approach. Um and I think to me, it feels like that's the place where you get the answers to those really specific questions. The the um, for instance, the example you gave there, billing, this particular type of query, and then uh, you know, a risk or an action or something. Or um, and from my point of view, like thinking about that from a software point of view, I might choose something like um screen, like which which part of the product they were using, and then what they were trying to do with it, and you know, like what the challenge was. Was it an error, was it uh like an aesthetic problem, or was it uh like you know, like a function missing piece of functionality or something? So I've come across different hierarchies, and I think my natural tendency, because I I I like to kid myself that I think in a way that that is that ordered, but what human being does really, but and and that's I guess what you're speaking about, really the the fallibility of humans in such an ordered system is uh uh is is a risk factor, isn't it? So you can apply a lot of order as you were saying and get like really deep insights out of it, but that's useless if the humans fail.
Tadas Labudis: 12:44
Yes, and I guess one of the symptoms, if you're not sure if you're suffering from this problem, uh the first symptom is um you see a tag that's overused. So for example, you might have product issues, other, or like other query or unknown query. Um and if you see agents picking that, you know, 20% of the time, 30% of the time out of all your tickets, you know that that's kind of like a black hole where all the uh more difficult tickets are going into. So you're not getting any insight on that uh collection. Um so if you have that tag, it's it's a time to rethink like should we pair back and simplify our taxonomy? Um, or should we, you know, if you really want to have that specific detailed insight, should we invest more in training agents? Uh actually, you know, highlighting the value of tagging so agents understand why they're doing it and that why it's important. Uh I'm not sure if you've ever come across that, but that's what we see in a lot of companies.
Charlotte Ward: 13:48
Yeah, yeah. I mean, it's nice to think that um coaching agents to behave in manners that maximize the the like the the value you can the value you can extract from your tools is is the way to achieve all of this. But I think frankly, honestly, all of the training and all of the coaching in the world is not going to get over anything like that's really onerous and complex in a busy frontline role, is it? I I think you can you can be idealistic about that, but frankly, you have to have something that's usable on the front line.
Tadas Labudis: 14:27
Absolutely, you know, and if if you're tagging manually in a support tool, you're always relying on a human process. And you know, there's this kind of entropy effect where any human process will break down at some point, especially under pressure. So, you know, when the ticket backlog fills up, and that's what we've seen a lot with the rise of COVID, where maybe companies are pairing back on how many support agents they can hire in a given time or resource. Um, but the ticket volume is in some cases even going up because lines are having, especially the travel industry, cancellations, reschedulings, uh, just general stress of what's gonna happen with my trip. Um so you kind of get pushed from both sides, and tagging is often seen as this auxiliary task. Like it's not gonna help me today to tag a ticket because someone at the end of the month or the end of the quarter is gonna look at this data and make some decisions. But from an agent's perspective, that's not gonna help me get a promotion, it's not gonna help me today in working through this backlog. In the reverse, it's actually gonna slow me down, possibly. Um, so that's where uh tagging systems break down, then you have gaps in tagging or inconsistencies. Um, but obviously it's a it's a short-sighted uh approach because you know insights are actually essential for you to actually manage that support volume. So it's it's kind of the coming back to the first principles. If you know what problems you have, you can solve them. If you don't, then you're kind of you know flying blind. So that's why it's important for everyone involved in the tagging process, the manager who's implementing the system and responsible for the system, and the people participating in production of tags uh to kind of understand the purpose and importance, and of course, keeping in mind those ways to ensure it's a successful program. That that will also kind of help with that buy-in.
Charlotte Ward: 16:22
Yeah, yeah. And and I think that you know frontline agents can be major beneficiaries of that, um, if you get that right, if you're able to take the right actions from a tagging system that's working and that everyone has bought into. I think I think they can be major beneficiaries of it. Um, I think though that there are lots of other stakeholders, aren't there? Um, you know, particularly around product, let's say, or around billing, or any of like any of the you could almost capture anything in these customer conversations and take those insights elsewhere into the business. Um which I think begins to add more complexity, potentially, but certainly more responsibility to your frontline agents. And they and therefore, therefore they see even less of the benefits potentially in their day-to-day job. So the incentive is pretty low. Where do you where do you stand on um automation?
Tadas Labudis: 17:23
Yeah, you know, that that's um that's a good question because obviously that's unbiased. I'm working in the space, and obviously we have spotted the gap that we're trying to solve. But automation, a lot of the support tools now have rudimentary auto tagging features that allow you to tag incoming tickets. That's not something was available when we launched CrowdSite, but then kind of uh obviously these big players start realizing the benefits and started adding those. And I think that's that's great because I don't think there should be like one tool that everyone must use. You know, everyone contributing and solving these problems just makes everything better, makes it competitive. Um, so I would you know start leveraging those tools even without shelling out on anything more complex. You know, if they're categories you already know, uh, you know, you have buy-in for tracking them, there's stakeholders that are using those categories. Just set up simple you know, auto-tagging rules and let us do at least some of the work. So even if agents are falling down on tagging, then you'll have some insights trickling through through the automation. Then of course, um, you mentioned that you know your kind of natural inclination is to go for a hierarchical approach, get more detail or as much detail as possible. And that's I guess where third-party AI-powered systems can add that value. So, on one hand, you want you know that that detailed taxonomy, but you don't want to overload agents, want to make sure there's consistency. These AI systems essentially do that. So to allow you to build uh detailed taxonomy, uh, set up complex rules, uh, allow the machine learning classification to do its magic. Sometimes you might even have sentiment analysis or additional types of analysis, and then kind of let that run and take it off the hands of agents and put it uh, you know, add that system as an extension of your team that gives you the insights without the human uh pullbacks. So, you know, Pradsai is one of those systems, but by no means the only tool out there.
Charlotte Ward: 19:29
Yeah, yeah. Yeah, so I think I think this area around automation is one that I have always um found a little mysterious. You know, there's an element of mystique to it. And and I think with that mystique comes a kind of sense of somehow like trepidation or loss of control. You know, I have I I'm attracted to the ease of a system that I can just plug in and get insights from after some time, but. I think because of the lack of understanding of exactly how these insights are being pulled, I kind of lose the sense of connection with potentially what I might otherwise go and do as a human being, or at least this particular human being, which is to create that very hierarchical and structured approach to tagging. And I feel I think one thing that's held me back from jumping on the automation side of things is the lack of understanding I have about whether I'm going to be able to answer the questions that I have at this point.
Tadas Labudis: 20:36
Yeah, I think that that's a great question. And you know, when we are familiar with the manual tagging system, you have a lot of control over tagging. But then remember, if it's the agents that are doing the tagging, you have like one step there already. So, you know, you might have your own interpretation how it's done, but how the agents interpret and execute that, that's already um removed uh from you directly. So one important comparison between manual systems and automated or AI-enabled systems is instead of tagging each individual piece that's coming in, uh, like in a manual system, automated systems are based around rules. And sometimes they're disguised as uh you know kind of mysterious AI, machine learning, robots taking over the world. Um, but even even in those models that have been trained to classify things according uh to training data, you know, uh essentially behind the scenes that encodes some rules that perhaps not easily editable, uh, but those rules are still there. You know, something, uh a comment looks like it's about a certain issue that the model has seen before is gonna classify that by comparing it to what it understands about that kind of data. Uh, and of course, there's a lot of complexity and technology involved in research, but that's what it boils down to. So garbage in, garbage out, that kind of thing. So and also another thing is not all of these systems are are equal or use the same method. So there's a range of different um text analytics methods, the whole field of natural language processing um that you know that these companies are benefiting from. So, you know, on one hand, you have more heuristic-based approaches where um you can specify keywords that you want to track, and if those keywords are present in your text, that will be a match and it'll be included in your report. Um, the benefit of that is that you have very direct control. The drawback is you might not know what keywords to use, or you might use keywords that are overly generous and that might bring in false positives. Equally, you might miss some keywords and you know have uh have gaps in your knowledge, essentially not get the full recall on those messages. Um another approach that's quite common is um using machine learning models that classify data based on prior training. And there are companies out there that have gone into you know SaaS, retail, e-commerce, and built models specific for those, or sometimes general models that train on some customer data. And you know, the good thing about these is they can cover more nuance. Um, so for example, in Proudsite, we have uh developed a model that can detect uh product issue and distinguish it from a product complement or a service complement or someone talking about churn or billing issue. So what it's doing is essentially it's comparing your message to the model and what it's seen and giving a score and then saying, okay, it's I'm 90% confident it's a billing issue. I'm gonna apply the stack. So the benefit here over keyword approach is that you can go more nuanced. So there might not be exact keywords that you would define, but the whole combination of the sentence and how the words are aligned gives that model that insight. Uh, the limitation of this approach is of course that you know, if you're serving thousands of different companies that have their own different products, their own issues, you will not be able to forecast every possible problem in the world and have a model ready for that. So some mitigation methods are you know to update these models with some additional training data and get them more accurate, uh, and possibly even introduce custom topics that the user would train on their own workspace and kind of like run on that machine just for themselves. Um what we've decided to do in ProudSite is actually combine the heuristic approaches where we can discover popular keywords, allow you to control the keywords, different rules like inclusions, exclusions, uh, to get as much control over that, but also introduce machine learning models so we can say, okay, show me proc issues that are to do with login, sign up, and password. This is how you kind of find these clusters and maintain control. Uh, but by no means, you know, it's it's never gonna be perfect. The human system is not perfect, the automated system is not perfect. You're always balancing accuracy versus cost versus um the insightfulness of your data.
Charlotte Ward: 25:29
And and as you just said, it takes some maintenance, doesn't it? Because I I was I was going to I was going to ask you about how you future proof this, like really early on, whether you're going the manual route or the automated route. Um But I think you just touched on it there, actually. And that is that in in both cases, it needs maintenance. You might have to come and refine rules or add a layer, or or as you said earlier, peel back a layer if you if you're just finding strange anomalies with things falling into buckets where they just don't belong. Um and and maybe that's the future proofing. Maybe the future proofing is that you can't do it all up front and that you actually have to maintain it.
Tadas Labudis: 26:10
Yes, I think you know it's um neither process is completely perfect, but it depends on what you want. You know, if you want a detailed taxonomy that's done for you in a reliable way that you can control at the rule level rather than you know telling people what to do, um, then an automating system will allow you to reap those benefits for a cost. Um, but then I think sometimes what we of course deal with um is um kind of over-glorifying of the manual system because people are familiar with, they think it's it's kind of the bait the benchmark. But um, what was really interesting in the early days of prod size when I would do free tagging for companies as a way of providing the service, testing whether that's actually needed, um, we would get about 20% coverage. So a company says we are doing manual tagging in our system, we're running through our algorithm, and we actually see that only 20% of tickets are tagged. Uh so you know you might have consistency within that 20%, but you're not getting the coverage. 80% of the tickets are the black hole.
Charlotte Ward: 27:13
Hmm, scary black hole. Um, yeah, I mean, if you I I guess you know more data, even if it has imperfections, is always better than no data, right? I think I think that, as you said, the coverage is probably key coverage and keeping it as simple as you need it to be in many respects, I think perhaps does does kind of give you some elements of future proofing as well. I've got one final question for you, which is not so much about future proofing, but it's about um dealing with the past. Um, this is a nice it's it's nice to think about these things like very theoretically and very operationally if you're doing them from the for the first time, but I'm sure I'm not the only support leader out there who has ever inherited a tagging model uh or a mess of tags that has no coherency. Um and what do you do with that if it's not actually providing you with what you need?
Tadas Labudis: 28:17
Yeah, that's a tough one. And you know, as humans, we always think that whatever we touch is is perfect and what we inherit is is inherently worse. And in some cases, it's absolutely right, you know, that might be the case. And I think in those cases, the instinctive reaction would probably be just stop tagging altogether. It's broken, I don't understand it. Like let's let's not do it until we figure out the new system. But you know, at that point, what might happen is there might be people in the company that are relying on these tags as imperfect as they might be. Uh, you know, maybe there's a product manager there who's looking at that for product issues, or uh marketing managers looking in the pricing comments or whatever. So you might be depriving them of that without knowing. Um, and another thing, you know, even if it's imperfect, it might still be adding value, might still give you these like broad contact reasons or whatever the tags might be. Um, so I think what I would do is I would continue running the process and try and understand as much as possible about what value people are getting, what's missing from the system, coming back to the goals or why we're tagging, uh, and then kind of start designing in parallel a new system where you have all those fundamentals for place, you know whether it's gonna be a flat taxonomy, how many tags you're gonna have, what those tags are gonna be at the starting point, and then uh just create a transition plan. It's like, okay, guys, kick off meeting, you're gonna have these new tags. Uh, this is what we're gonna do, and this is why we're gonna do it. So that kind of would reignite without losing that insight in the meantime whilst you're figuring things out.
Charlotte Ward: 29:55
That's definitely my temptation is just to like wipe the slate clean. Just as you say, put a halt there. Let's just forget it for a couple of weeks while we figure out what we actually need and and and start again. Um, but there are just so many dependencies, and and that just speaks to what we were talking about earlier. There are so many stakeholders outside outside your own team, right? And I think that gosh, as painful as it might be, we have to uh we have to help those guys as well, right? Absolutely. Yeah, thank you so much, Tadas. I think this has been uh a really interesting journey through where to begin to to hopefully where to end up and and uh and all of the the potential stopping points on the way and and uh challenges and and ways to get around them. Um if you have one final piece of advice for our audience before we say goodbye, um what would it be?
Tadas Labudis: 30:57
Yeah, so I think uh one thing that we see a lot is companies understanding the value of data that they're gathering, and some of that data might be passively kind of streaming in for support or other channels. I think you know the future is where companies kind of develop an edge around having superior insights. So if you understand your customer segment, your customer problems and can solve those more effectively than your competitors, you will have that advantage. And you know, with so many products popping out from everywhere, things are becoming commoditized. It's kind of harder to maintain that technological edge in some cases. Uh, and I think you know, customer insight-led companies, data-driven companies, but in a real sense, will actually be the ones that thrive. And I'm just gonna encourage it, you know, whether you're gonna do manual tagging, you know, kind of spot checking analysis or invest in an automated system, you know, just start thinking about uh that data and what it can do for you.
Charlotte Ward: 31:58
I couldn't agree more. Thank you so much, Tadas. Uh it was a pleasure to meet you. Thanks for joining me today.
Tadas Labudis: 32:03
Thanks for having me, Charlotte. It was a pleasure. And I hope it's going to be useful to your audience as well.
Charlotte Ward: 32:11
Oh, it's it's useful to me, if no one else, which is why I couldn't have been more pleased to have you on to talk about it, because um I have one of my friends, one of my friends, colleagues, guests, um, Simone Secchi, who's the head of support at Doodle, um, has a manual tagging taxonomy and he has a very rigid approach. He's like, it's three three-level hierarchy. It's this, it's this, it's this. And thus far, he's been kind of my model for maybe that's what I should be aiming for. But I've inherited such a massive mess of tags. Um, and uh, I don't I've I've been struggling like in my own head, how do I get from there to there? And yeah, which is kind of yeah, the the final that was why I wanted to ask you that final question is like just what do you do? Because I'm just every day I look at our like the little box on our tickets, I think. I just empty them all.
Tadas Labudis: 33:07
How many tickets do you roughly get on your system at the moment?
Charlotte Ward: 33:11
Um 200 a month, 220 a month, maybe to on a busy month, 250.
Tadas Labudis: 33:17
Yeah, so one thing uh I I did include in my article, but I didn't mention here is how volume impacts how detailed your taxonomy should be. So imagine with Doodle millions of you know free users or like kind of prosumer type of users. Uh I imagine they're getting tons of volume. Uh in a kind of B2B context, and you know, we are here ourselves, uh, we don't have a ton of volume. It's actually it does it's not worth having a very detailed taxonomy because first, if you wanted to retrospectively tag stuff, it's actually easier to do than doing 10,000 tickets. So if you have a raw taxonomy that's you know classifying things into 10 categories, you can then take that category and analyze it further when necessary without doing preemptive work. And also another point if you have 200 tickets and a very detailed taxonomy that let's say has 100 tags, the distribution is gonna be so that you know this month you have two mentions of that, next month you have three. Those are not statistically significant insights anyway. So uh I wouldn't beat yourself up about you know having a very detailed or very comprehensive taxonomy, just start small and then grow as as you need new tags.
Charlotte Ward: 34:34
That's actually really interesting. So make it rather than trying to impose structure on it, actually just think about reducing it because we've got this kind of monster of a it's not a system, it's just people throwing in tags when they feel like when they feel like there's something new to say, which means I do get these single data points every month. Um yeah, maybe just regrouping again is the way to go, like really paring it down. Yeah. Yeah. Yeah. Oh bear that in mind. Thank you. We should have put that on the podcast, but I I might that's it for today. Go to customersupportleaders.com forward slash one seven five for the show notes, and I'll see you next time.
