Turning Support Tickets into Product Insights with Tom Bachant (Unthread Founder) - Ep 21

Kyle James: Clearly, the reason you agentize something like this is because it's something you're going to run on some sort of cadence.

Tom Bachant: We can't have every engineer and every product person reading every single ticket that came in. So, let's use this to jumpstart that conversation.

Kyle James: How long do you think it would have taken you to compile all this together on a weekly basis if you didn't have it in a tool like this? And how good do you think it would be in comparison? Can we put an actual time/dollar amount on what this thing's giving you?

Tom Bachant: Yeah, I can say to compile one of these reports for us, which we do every two weeks, this would take me

Kyle James: How does your team review support tickets today? Do they ever make it past the, we'll look into it stage? Do you use the data to influence your product roadmaps, operational playbooks, or content strategy? In this episode, I sit down with Tom Bachant, founder and CEO of Unthread, an AI-powered support platform built directly inside of Slack. Tom's not only helping IT and HR teams automate internal ticketing, he's also the creator of a support ticket analyzer agent that's making waves on Agent AI. From building a ride-sharing startup in college to leading an acquisition by Cruise, earning Forbes 30 under 30 prestige, and now scaling an AI native company, Tom's journey is full of lessons on spotting real problems, dogfooding your own product, and knowing when to let AI take over the tedious work. And with that, let's get into this episode.

Kyle James: Tom, thank you so much for joining me on this conversation today. I can't wait to kind of share the audience like your background. You've got such an interesting story, founder, entrepreneur, um, you know, you founded ride dash, you know, transportation tech, uh sold it off to Cruise, um, and now you're building a company kind of in the uh support space. Um, would love to like walk us through that. Like, how do you go from one to another and and your origin story? Like how did you get to be, you know, um, this this this person who's out there just solving problems on a regular basis?

ed a ride-sharing app in like: 2011

Kyle James: And so it sounds like you just identify problems. Um, or or even the t-shirt. Like why the t-shirt? Like did you have was it was it witty comments that you kind of had on the front or or was it more design focused? Like I don't know, go go into that a little bit more. That's interesting.

Tom Bachant: Yeah, maybe I was a little bit of an edgy kid growing up and uh I grew up in Massachusetts. Everyone was wearing Hollister. I don't know if you remember when Hollister was really cool.

Kyle James: Yeah, the cooler version of Abercrombie. Yeah, yeah.

Tom Bachant: Yeah, yeah, exactly. You'd go in the store and it just smelled like cologne. Um, I was a bit of a skater punk kid and I thought Hollister was a little too mainstream. And so I wanted to make my own t-shirt brand and had a really clever name. I called it Not Hollister, and I registered the domain name https://www.google.com/search?q=nothollister.com and I made some shirts with like a little X through the bird logo. And uh found some other edgy kids who had the same thought as me and uh yeah, who really resonated with it.

Kyle James: That's awesome. That's awesome. All right, so that thread started ride sharing and and now like you've kind of built this company kind of on top of Slack, solving support tickets. Like tell us a little bit more about what Unthread does and and kind of how that makes sense as we get into like this agent you built to kind of support people with with with tickets this with trying to look and and analyze a bunch of tickets in their system.

Tom Bachant: Yeah, so I think as I got, you know, older and progressed in my career, I saw bigger and bigger problems. And for Unthread, what we were trying to solve is how do people get support within Slack? Um, so let's say you're trying to get uh a new laptop from your IT team and you, you know, send him a message, you're like, hey, you know, I need a new laptop. Generally these messages get lost in a sea of threads and you've got, you know, thousands of slack channels to check. Everything's bold on your left hand nav and you've got DMs everywhere and it was basically impossible to manage. So, how do we create an automated system for getting support and also providing support for other employees while staying natively within Slack? So that was the kind of first uh problem we were trying to solve. From there, obviously it grew a lot. Like when you saw what sort of problems are being solved by people who use Slack, like this laptop thing is just one example. Like people are doing mission critical operations by sending slack messages around and they're getting lost in channels. So we think about it as operational automation, all driven through chat. And AI is a huge part of that because we need to understand what are people trying to do? Uh what's the context of their request? Like why are they asking this in this particular place to this particular person? And then how do we actually solve this problem? Like in the case of a laptop, we know exactly where the user lives and what sort of laptop they need. So why can't we just buy one using an API to like purchase a computer and ship it over to them? So that's been really fun is just kind of like expanding how much we can solve this problem away from just, you know, tracking tickets all the way to actually like accomplishing tasks.

Kyle James: What what I'm hearing you say is like all of this is probably more internal support, right? You're not necessarily helping businesses do support for their customers in like a support system ticket tracking. It's more of like, hey, we're the internal IT team or I guess it really could be any department, um, that that supports people and how do we internally because we have this much more richer amount of information about the user like you said, like where they live, um, to kind of make those streamlined conversations happen and and like the checklist or tasks that kind of need to happen, um, post conversation. Is that am I understanding you correctly there?

Tom Bachant: Yeah, the way we see it is Unthread is an AI automated help desk that's built into Slack. And we do have people who use us for external support because the same problem exists when you set up like slack channels with customers and definitely some people use Unthread for that use case. When I think about where does Unthread provide the most uh like automation for like getting your job done, uh that's really on the internal support side. That's where, you know, we can integrate with all your internal tools, we know your like SOPs and all your processes and that problem space is just really exciting for us. That's that's where we spend a lot of our time. But this idea of solving support over Slack can apply to both, for sure, both internal and external.

Kyle James: That's cool, but I want to ask like, how do we tie this into AI, right? Like I'm sure what y'all are doing, there's a ton of AI trying to put this stuff together. So like, tie that thread together for us all.

Tom Bachant: Threads. Is a pun. Uh, pun intended here.

Kyle James: I know, you know, sometimes I get a little punny here, but.

Tom Bachant: I'm always down down for some dad jokes. So, yeah, I think this idea of support is a great use case for AI. Everything is driven by language. So all these LLMs are optimized to solve this exact problem. Uh the first like most low-hanging fruit of how AI was used in this sort of use case was generating answers for people based on, you know, documentation. Like that's just easy, set up a little rag pipeline, get some documents into the AI, pass in a question, you're going to get some really great responses. And that was the first product we launched when chat GBT first launched their API, you know, two years ago. But because so much is happening through chat and so much is driven through text and like just like, you know, communicating with people, we've been able to take this to another level, which is not just generating answers for people, but, you know, understanding intent of what people are trying to do, understanding what solutions were to problems, and then creating this feedback loop of, hey, here's what solved that problem the first time around. Let's create some documentation around that so that then the next question comes in that's the same, we'll use that same SOP to solve this problem and then that'll be used the next time the question comes in. And so all of this is driven through communication and it's all just something that LLMs are extremely good at.

Kyle James: Well I think that this kind of piece of it really ties into this agent you've built, right? Like what are these common themes that you're seeing. So I would love to kind of, you want to show it? You want to you want to give a demo of kind of this this agent you've built to kind of like analyze support ticket systems?

Tom Bachant: So let's walk through uh the support ticket analyzer. So the reasoning behind building this was again, we have this huge set of data of what people are talking about, like what issues they're experiencing, what are the solutions to those problems, and then understanding a deeper level of like what are the categories of issue. Um is it mostly asking questions? Is it normally getting help from people? Um all that stuff is there and humans are generally pretty bad at looking through a huge swath of data, but thankfully, that's what AIs are extremely good at. So you can just feed it a huge suite of uh ticket data and um, you know, unstructured conversation history and let it do all the hard work of reading through it and parsing through it. So what we've done is factor in a prompt based on what we see people looking for out of their data through Unthread. Uh we designed our own custom prompt for this bot that really tries to pick out the most relevant information and present it in a way that generally people like to see it. And so, um that's what you're going to see here. Uh it's going to control basically like how it defines the ticket categories, how it displays um, you know, what the recommendations are to the end users and uh just like having some really clear takeaways because, yeah, having AI summarized data is a pretty easy problem, but having that summary be actionable and like useful information is really hard. So that's uh really what we've tried to optimize here.

Kyle James: And and we should probably clarify too that like anybody that can get an export out of any ticket tracking system, you don't have this, they don't have to be using Unthread. Um, it could be using Zendesk, Hubspot, whatever, um, that that if you get that, if you could pull a, you know, a CSV export of of however they want to structure their tickets, uh, this is something they could just run and test and and try to pull together threads that maybe weren't obvious to them, right?

Tom Bachant: Yeah, exactly. So as much as we would love if people were using on thread, they definitely don't have to. And yeah, like you said, you can take any export, export a CSV from from Jira, from Zendesk, from whatever. Uh the way this prompt is designed and basically just how good LLMs are at parsing data, it really is agnostic to the schema and the structure of the data. And yeah, we're happy to say like, yeah, if you're not using Unthread, you still want to be able to generate some useful insights. That's great. I mean, hey, why not be able to do that. So, here's an example of some output uh from this uh bot. So what we want to say is, here's all this stuff that happened. Here's all the stuff people are talking about. Now, tell me exactly what I need to worry about and what I need to do about it. And that last part is important, like what should I actually do here? So for me, I'm a product person, you know, I'm always looking at like, what do we need to optimize? What features should we be building? What bug fixes do we need to prioritize and optimize? And so what the agent does is runs through your list of tickets, tries to understand all the different concepts. It itself builds a hierarchy of the different um, like categories based on frequency and based on perceived priority based on even like the tone and the sentiment of how people are talking about things. So if someone says like, hey, I this button is the wrong color, it's obviously going to be interpreted differently than like I'm locked out of my entire account. So I can see here the number one recommendation here is related to access governance and authentication. This is pretty common actually within like IT teams who would use Unthread. A lot of the tickets that they're doing are around giving access to software, revoking access to software, you know, setting up, you know, 2FA, resetting uh 2FA tokens, like all this sort of stuff is is actually pretty common. And so I can see what the most common types of issues are here. Uh it's a pretty, it's a high frequency and the AI has also said urgency medium. Like people are able to get get by and like do their jobs, but uh, you know, it definitely should be solved. And uh this will also link out to some given tickets uh in my backlog related to this.

Kyle James: Which is super useful like knowing some of the urgency priority, right? Like you get this wall of things you can do, but like even here it's kind of telling you, hey, we saw this a lot or this seems like it's a big problem. You might want to prioritize it more importantly than some of these other things. Uh people should never discredit that.

Tom Bachant: Yeah, yeah. And for a lot of our customers and uh, you know, other companies out there, they're doing thousands of tickets a week and like, I mean, that's crazy. I couldn't imagine trying to read through all that. That's like a full-time job in and of itself. And so again, like the LLMs are just so good at this, so let's let them do all that annoying work. So, uh, yeah, so this is the summary of the types of issues that have come up. And then again, the recommendation and the what to do about it is really important. So I can see here, uh really we're having trouble identifying the problem with authentication failures. So what we need to do is improve our error messaging and our logging around that because generally when people are reaching out with auth problems, they're getting some generic like auth failed error. So, okay, that's really helpful to know. That's pretty low-hanging fruit. If we had some better messaging, we can solve these problems a lot more quickly. We can understand exactly what the problems are that the users seeing and uh put that into a sprint for the engineering team to look at next time. Another one here, payroll, uh so a lot of like HR teams also use Unthread for doing employee support. So you might reach out and say like, hey, what are health plans this year? You know, how do I what's my deductible, whatever it is. And here I see we're getting a lot of questions about payroll and uh that's obviously pretty important. Uh that's a that's a high urgency one. People generally care a lot about getting paid for the work that they do. So let's make sure that that one makes it onto the uh the priority list for the next uh the next sprint.

Kyle James: Tom, I'm curious like what what inspired you? Like what made you decide that you wanted to create like a support analyzing agent here? Like clearly you had a business that kind of does this, but like what was the what was the reason? Were you just wanting to play? You just wanted to try some new stuff out with AI or or kind of like, did you have some sort of like marketing, um, brand awareness tour to like, hey, let's create something out here that can that could, you know, people could try to use and and discover us and then we can talk about how we can help you with the other stuff.

Tom Bachant: Well, I think I'm in a fortunate position where I am one of Unthread's ideal customers. So the problems that we're trying to solve are also the problems that I see in my day-to-day life. Like we're supporting our customers, we're supporting our employees. We are facing this exact problem and a lot of the Unthread roadmap, I can certainly say I unfairly influence what stuff we prioritize based on the problems that we see or that I see in my day-to-day. And this was a big one for me. And we definitely validated that it was a problem with other customers, but, you know, for me, I'm I'm trying to prioritize among the thousands of feature requests and bug reports that we get, hopefully not thousands of bug reports, and prioritize this across different customers and different problem areas. And it really is a a hard problem to solve. and as a founder, it's one of the most important things you can do is choose what stuff to build. If you build stuff that's low impact and potentially high effort from your engineering team, you're just not going to make it. So I looked at our own ticket history and I said, okay, this is something that AI is able to solve and this would provide me a huge amount of value. And so I've started taking this to our sprint planning meetings and to our general like grooming sessions of our backlog to influence uh how we think about building stuff. So that way we're always keeping our finger to the pulse on what customers are talking about. We're always building the right things that are the most uh urgent and highest impact. And we've seen a huge amount of success. So really what we did is, you know, we first, you know, dog food this product with ourselves. Uh we shipped it out to our customers, you know, in a beta sense and let them test it around. And then once we really honed in on exactly what people want to see and the value they get out of it, we basically open sourced the prompt in a way and put it into agent.ai so anyone could use this and get the same value that we get out of it.

Kyle James: Well and and as kind of you mentioned like the obvious user persona for this is like product managers, right? But I'm sitting here thinking like this is also useful for marketing people who are trying to write content. Um, it's also useful for sales people that are trying to understand like customer pain and like what are they running into to maybe in demoing things they should avoid or or things they should be prepared to address and answer. And clearly support people, this is useful because this is the things they're running into over and over and over again. But if they're able to like pass this on to product team like you, um, they don't have to waste their time talking about it because you could just synthesize and bring it all together. So there's probably a lot of different departments where you're breaking down some silos and kind of bringing some synergy together in ways that aren't always obvious and people don't really know how to do, but then they're like, yeah, these are important. And maybe they tweak the priority a little bit. Do you see some of that happening where it tells you this is more priority but maybe some conversation's like, uh, maybe this isn't as much or maybe it needs to be more.

Tom Bachant: And I think that's a good point about AI in general is that it's a jumping off point and then human judgment still comes into play. So we're not going to take this data and blindly follow it. We're going to use it as a good guiding point and we'll find that it's right almost all the time, but still we need to like factor in our own knowledge and you know, use our own judgment to a certain degree. So, yeah, that case you mentioned about, you know, maybe cross team you're seeing different things, you can use an AI to try to, you know, create one source of truth across all those teams. But I think what we've seen be more successful is each team finds out what's a priority for them and then you can have a conversation about it and really deeply understand each other's problems in a better way. Like if a sales team shows up to a meeting with this type of summary of like, here's everything I'm talking to my prospects about and here's the list of how important these things are to people, that's something that a product team can understand way more clearly than a series of Jira tickets or, you know, 500 gong recordings. Like you need to have this synthesized in order for people to make good judgment calls.

Kyle James: So, with that, I'm kind of curious, what was the first, or do you have one? like an aha moment where the first time you or the first couple times you ran this and certain things really stood out to you that just like completely surprised you like, oh, maybe we do need to do anything. and and like, have you seen some of your roadmap change from using this?

Tom Bachant: Yeah, I remember the when we first were validating this, again, we wanted to see like, is the AI close enough to human judgment that it it is this perfect jumping off point. Where it's summarizing things and discussing things in in the way that I would. And so we took a set of data from a week where we had a particular issue with one part of our system that we know was like pretty critical and we had maybe a bad user experience around it and we had a bunch of support tickets around it and like it was very obvious what the output should be. And I said, okay, well me as a human, I consider myself a pretty good product person. Like I think I know how to summarize this. I think I know how to judge the urgency and how to think about what the next step should be. So let me see how close the AI is. And honestly, it did a better job, I think of communicating it than I did. But the the general concept was exactly what I kind of saw as a person after reading through these tickets manually. And so that was kind of the point that was like, oh, I actually don't need to babysit this system. Like it it pretty much knows the same things that I know and it can communicate them sometimes even more clearly than I can. So it's kind of like running that control versus experiment sort of thing that you do in a science class. Like, I knew what the output should be and I'm going to try a bunch of prompts and see what comes out of it. And then when I get that outcome, I'm like, oh my God, this like this thing is like again, better than me in some ways.

Kyle James: And you kind of alluded to it, but I want to like circle back to it like it seems like, clearly the reason you agentize something like this is because it's something you're going to run on some sort of cadence, right? It's not a, I'm running this thing once a year. It it it sounds like you're saying like weekly, weekly is kind of the right cadence. Like, hey, let's see what happened in the last week because we've got some crazy number of tickets and we don't want it to get too out of control, but it's it's kind of like, I don't know, end of the day end of the day Friday or first thing on Monday. Let's come back and get a recap of what happened so that we can kind of make plans. Is is that kind of the the process that you kind of see working?

Tom Bachant: Yeah, I think it's great to bring up at a a retro of the week. Um, for us we do like a grooming session at the end of our sprint to try to make sure we're aligned on what's coming up in the following one and that's where we discuss it. So we can say, all right, we're about to plan a bunch of work for the team, but let's quick quickly do a catch up on what the heck was going on in the past two weeks and we can't have every engineer and every product person reading every single ticket that came in. So let's use this to jumpstart that conversation. Um, so yeah, I think definitely frequently doing this. Um, keep looking at your results and seeing, all right, have we been able to reduce the number of high urgency issues and are now are now we we're seeing more high impact feature requests, let's say. So that means maybe hey, we're building in better stability into the system. Um, so yeah, I think it's worth it to to frequently do this and then also compare the results uh like week by week of of how things are changing.

Kyle James: Oh, that's interesting. Yeah, you've got that Delta in there too that I didn't think about. But I think it also if if this is something that you're running frequently like it's totally thing to turn into an agent, right? Because it's it's got all it's it's a workflow of these things that need to happen. But I I'm I'm curious too, like how long do you think it would have taken you to compile all this together on a weekly basis if you didn't have it in a tool like this? And how good do you think it would be in comparison? You know, I'm just trying to like, can we put an actual time slash dollar amount on what this thing's giving you, you think?

Tom Bachant: Yeah, I can say to compile one of these reports for us, which we do every two weeks, this would take me hours, for sure. Definitely hours and instead it happens in, you know, 30 seconds. So, pretty big ROI there. The other benefit is that I know that I'm subject to my own bias. Like I said, I'll uh, I'll try to prioritize features that make my life easier as the founder of Unthread and how I use the application. So this also does a great job of being a an unbiased third party that looks at issues across the board and says, hey Tom, I don't care exactly what you do every day. I care about what's actually the most impactful thing that you all can work on.

Kyle James: Do you have any stories of of external people, you know, you mentioned kind of rolling this out to customers beta and then now it's kind of out there in the wild on on Agent AI, like, do you have any stories of anybody else that's come to you that said, hey, this is great, or hey, do you think you could tweak it to do this or or any of that feedback that's rolled in?

Tom Bachant: Yeah, I think the most exciting thing that we saw was someone requested that we add this functionality to our MCP, uh which we also just launched. Uh so people who are not even in the Unthread app, they're using Claude or they have some sort of enterprise tool that, you know, that they're using for their their internal LLM. And you know, we already had an MCP where people could query data and and sort of analyze it. The nice thing about what we built with this agent is that layer of QA and fine tuning what the results are where you don't have to create your own prompt and say like, hey, I'm going to provide you a series of tickets, please tell me the most high impact. Like you don't have to do that. We already have this pretty well fine tuned, although that's a term that I guess means multiple things. But the the prompt is pretty pretty well figured out. And so we exposed this over the MCP and that was really cool to see because you know, you can go to agent.ai and upload a CSV and you'll see these results. But now because the MCP already has access to your unthread ticket history, you can just say run the support ticket analyzer and we'll spit this out in a couple of seconds. So that's been kind of cool. It's kind of like the full circle of all these different tools you build and it goes to show how fast things are moving with AI. Like MCPs are only I don't know, less than a year old, I think. And uh three to six months. Yeah. Yeah. So we're already layering in other agents into the MCP and you know, all the buzz words are really kind of playing nicely with each other.

Kyle James: Well and and just so anybody out there is like, what is an MCP? Let me let me see if I can attempt and you can absolutely correct me. but it's kind of like this middleware instructions for AI about how to access APIs essentially. It's a middleware. It's like how do you connect two systems together? This is the system that kind of does that exposes the instructions, hey, if you want to access these things in the system, here's how to do it and here's the correct way to get it out.

Tom Bachant: Yeah, I think it's exactly right. It's like uh REST was this standard API protocol for engineers and for platform developers to say, here's how you communicate with my system, here's the data you're going to put in, here's the data you're going to get out. Uh and now instead of engineers doing this, you have AIs talking to each other and they seem to like this MCP protocol. It's basically the same concept as REST where you say, here's the data I want to get, here are the tools I want to use, here's the here are the endpoints that are actually going to be used under the hood to do this. And you don't really need to be so structured and how you call it, you'll let the AI do a lot of that work to figure it out.

Kyle James: It's AI documentation for APIs. That maybe that's an even clearer version for people.

Tom Bachant: That's a great way to put it. I actually haven't heard that exact wording, but that's I think that's accurate.

Kyle James: So what's next? What do you have plans to fully bring this into the product or do you have more enhancements that you want to add to it or or like I know there's there's something spinning on on like how do you take this from here, right?

Tom Bachant: Yes, there are always more problems to solve. So, we have rolled this out into our own product, uh, and we're seeing some really good data on that, how people are using it. Right now it's in a beta, so a next step is to roll this out broadly across all of our customers. But the next agent we want to build that we also will put up on Agent AI is around solving the problem of documentation. So prioritization of features and bug fixes is one thing, a huge never-ending problem of every product team and, you know, it's hard to really solve it perfectly, but um, it's just going to plague every software development team. So that was the problem we want to take on first. The next problem that everyone faces and that everyone dreads is writing and updating documentation.

Kyle James: Yes.

Tom Bachant: I'm sure you've experienced this yourself where you look at a document in a confluence page and you're like, this was updated five years ago, is any of this stuff even correct and should I even look at this? And then maybe someone asks a question, you look it up in, you know, your system and then provide an answer that's totally different than your documentation, but you forget to go update it. So the next iteration that we're going to build and we've already started to build this into the Unthread product, but we want to again open source this into Agent AI so everyone gets the same benefit that we've been building is this documentation generation agent. First of all, we analyze all the existing documentation that you have and AIs are able to really generate a good map of this in internally of like, okay, maybe you have these different folder structures, but an AI can also understand across two completely different folders, there's a common theme among these two documents that, you know, maybe humans haven't even indicated but they can really pick up on that. Once we understand that sort of taxonomy of how things are documented. Now when questions come in and questions are answered by people, we understand is this relevant to maybe a document that already exists, maybe five documents that already exist, or no document at all, to understand we should generate a new article for you and put it into your notion workspace or into Google Drive or, you know, on thread we manage our own documentation uh internally also, or the really cool thing is the updates. So look at a series of all of your documentation and again, maybe there's five different copies kind of saying similar things, we can know to generate a diff across all those five documents and say, here's what you really need to update. And the agent part is actually going in and again, doing that work. where it's one thing to say, hey, you should go up these docs, but like I'm not looking for a to-do list that an AI is going to generate for me. That's really annoying. So I wanted to just do the work for me.

Kyle James: Yeah, just show me a change log to approve, right? Like here's where I want to change and then like let me approve it before it goes live, but that would be huge. God, I I yes, I want that.

Tom Bachant: Yeah, yeah, me too. Again, totally biased. I build stuff that I really want to use. And yeah, it's again this case of the AI should do such a good job initially that you shouldn't just blindly take everything the AI does, but it should do such a good job that you look at it, you once over, use some human judgment and say, great, this is good and and ship it.

Kyle James: I I've seen people start to do this stuff. It makes a ton of sense whether it's, you know, they're in cloud code and like give me my we could see the the pull request, right? So we could see like what's the change log? What's the release notes need to be like? Okay, what needs to go into the the the report or the support document. Like it's so it's a bunch of different steps, but if you could even help with all of them, yeah.

Tom Bachant: And a lot of this is process driven. I think with code, it's yeah, there's a lot of really exciting problems or products that are coming out to solve this for code where you can say, okay, I can see that a button moved from like pull right to pull left. And so in my documentation where it says use the button on the top right, I'm going to change the docs to say on the top left. Uh there are definitely a lot of great like static code analyzing tools that can do this. But a problem that isn't really addressed now is around this process documentation. So it's like, hey, I just hired a new employee, how do I onboard them? What what steps do I need to follow? And someone's going to answer that question. They're going to answer probably the same question dozens or hundreds of times. And so is that captured and is there a feedback loop going of like, can we make sure that we've defined a clear SOP on how to do this without you having to like write the whole document yourself?

Kyle James: And it's going back to the original thread where you're building all this in Slack. It's it's I I I should have said this earlier, but it's interesting how Slack has become kind of this additional tool for internal communication, especially remote teams, right? where it's not all email, it's not all phone, it's it's all slack now. Um, and if you've got this this agent that's kind of listening into this and it's seeing these common things that are happening, um, yeah, help me put this documentation so I can give instructions to people and then ask questions to it. It makes a ton of sense. I could see that streamlining so many different things and saving so many people lots of time.

early, this was like in like: 2018

Kyle James: It it's such a good point and you're reminding me like I've I've taught whole entire teams that like, look the difference between Slack and email is if you need something and you don't need it now, email me and I'll and I'll and check your email, morning, lunch and dinner or or end of the day. But most people that's they they've been structured to do that and if you need to interrupt me with something that doesn't require work but just an answer, that's what Slack's for. And and like there's this whole like, you know, uh ethics or or or like the etiquette. Yeah, etiquette about how people do that communication now, um, because it's kind of become so invasive on on the way we do business.

Tom Bachant: Yeah. and I think slack etiquette is a really important thing and it's started to start to go out the window for some people. They're just like, now it's the wild West. It's like you need help, just message anyone on Slack, send a bunch of DMs, like people are just going crazy out there.

Kyle James: Totally. I I'm curious kind of, you built a couple of these agents now. Like what is what do you think building these agents has taught you how to how to how does that plow back into the company, right? Like as a founder and as a builder, um, how is that changing the way you're growing your company now? Um, do do do you think about like how can we agentize everything now or or is it just like we need the process, we need to get this done a couple reps and then we'll figure out how to do that. Like how what has surprised you there?

Tom Bachant: I think it's helped to figure out what LLMs are really good at because we get the the benefit of seeing a lot of data flow through our system and through this agent and see how well is the AI performing at different tasks. And so we get to get a good sense of what the limitations are and where we can lean on AI a lot. And that influences our product too because if there are areas where LLMs are they struggle, like there are known areas, right? LLMs are not great at math, for example. So that's not a place we're going to rely on LLMs for and we can see that in how it tries to if we ask it to do certain things like, hey, run a, you know, quick regression on these numbers and, you know, response time metrics broken down by like it sometimes falls down, it's just not what it's optimized for. So we get to see what it's really good at. And again, that's where it's like synthesizing a ton of data and then we get the fortune of seeing like, well what are all the problems where this is, you know, something that humans are experiencing and um, apply that as like a prioritized use case so we can be sure that if we roll AI out, it's going to do something effectively, not just be kind of like a flashy add-on like we know it's actually going to be good at whatever task we give it.

Kyle James: Well and and to that point, I I'm curious, are there any other any tips? Um, you know, a lot of people that listen to this, maybe they built some ages, maybe they haven't and then they're like, I don't I don't know if I can do this or not. anything that you throw out there to like advise people that hey, you're going to do this the first time, um, to be aware of or things that you learn that like you wish somebody had told you sooner that you want to share?

Tom Bachant: A lot of it comes down to being a good communicator and you can lean on AIs to make you a better communicator to AIs. I think you're starting to see this now as a more common recommendation, but maybe a few months ago people weren't doing this as often. But we use AIs to refine our prompts that we feed into AIs pretty often because they're the best judge of what is an AI going to think about the prompt that I give it. So you want to start with a really great baseline, like super clear communication style, you want to be as as clear as you can about what's in your head and put it into words in a way that removes ambiguity and you know, you can be really verbose, that's okay. It's really about like making sure you're super explicit about what you need and then let the AI give you feedback and say, hey actually this could be better, this was a little bit unclear, specify this, put this into this order and that feedback loop feels kind of tautological or like redundant, but it helps massively. Uh they're just so good at helping you to understand how to talk to them.

Kyle James: It makes sense, right? Like it's like a translation layer like help me help me help you. But I also like from an efficiency standpoint, right? Like if you could take this thing that maybe I wrote extra verbose and like optimize it in a way so that it uses fewer tokens while still getting across the exact uh, you know, the the message and intent that I want, like I can see that totally being valuable.

Tom Bachant: Yeah, and AIs are so good at that, so we should, yeah, leverage them when we can.

Kyle James: Well I wouldn't be I be remiss if I didn't ask you about this because you did kind of tease that at the beginning. like you were recently featured on a tech roasted comedy show, right? So, so let's talk a little bit about that and some of the the the humor, right? Like especially when this is dropping, like Sora 2 just released last week and like Tik TOK is obsessed with everything and and you see so many crazy um, memes going out with with just humor and and and all sorts of stuff. How do you stay grounded and how do you keep your sense of humor with with kind of, you know, everything that's going on both in that show and just like in general in kind of the enterprise AI space?

Tom Bachant: Yeah, I think uh we can never take ourselves too seriously and there's a lot of really impactful stuff being built and a lot of big problems to be solved, but we're all just a bunch of squishy humans out here doing the best we can and we're all, you know, flawed in our own ways. So, yeah, the tech roast show is a fun uh fun experience where they they made fun of me for uh a previous company I started where someone left us a really bad review on Google Maps of all places. And then they they bought a uh a spite domain uh to, you know, kind of hate on us a little bit.

Kyle James: That's how you know you made it though, right? Like if someone actually went to that level, you know you made it.

Tom Bachant: I know, this and it's still up. They've been paying those, you know, $15 a year domain renewal fees for the past like 12 years. So that's how how much of an impact we must have had on them. But no, I mean humor is just important to me. I actually have done my own stand-up comedy uh, you know, from time to time and being on the roast show is just a a fun way to say like, yeah, I mean, come on, none of us can take us too seriously.

Kyle James: I completely agree. I mean, I love me a good bad dad joke, but I think it's also kind of like the ultimate like level up when you're when you're okay to make fun of yourself, right? Like you open it up like, hey, like anything goes like um, I I am comfortable in who I am in my own skin and and um, and we should all be that way, right? Like

Tom Bachant: Yeah, and I think if you're comfortable, it makes other people comfortable. I'm sure that's true on this show as you interview people, like that's probably why people feel comfortable. And for us, like, I don't know, if we're doing a demo or doing a sales call, you don't want someone who's just so full of themselves and serious. Like often times people are looking for an interaction that is pleasant in some way. You can still talk about things of substance without really, you know, kind of being too gravely serious about it.

Kyle James: Yeah, I I always joke with with Matthew our our show producer is like, yeah, it's my job to like get on here and ask the questions that everybody's afraid to ask but they want to know anyway. I don't care if that takes a hit to my ego like I'm still curious too, I want to know. Um and and the only way to know is to ask so you can't be afraid to ask those things.

Tom Bachant: Exactly. Got to be open to asking questions that some people think are dumb. Like yeah, who cares? Let's go ahead and ask it.

Kyle James: So Tom, as we kind of wrap up, like how can the community support you? Obviously, you you've put out an awesome tool here that people can kind of play with. are are you looking for more collaborators? Do you want, you know, you want you want feedback? Obviously customers, right? Like people get in here and play with this thing and if they want to come on and and use the whole product, I I'm sure you'd always be welcome to that. Um, but but what else? How how could people support you? What's the best way for people to connect with you?

Tom Bachant: Yeah, I think like any product, uh feedback is always the best. Uh I really thrive on that. Some people have reached out with some really helpful feedback of like here's some things I'd want to see and that's what's also provided some feedback on what we want to build in the next agent, like what are the other problems you have? Like where does this fall short and um, yeah, what other ways can we provide value to you? So if you're using the tool and there's some crazy idea you have, like I definitely want to hear about it, the crazier the better. And yeah, if you find it useful and you have a bunch of other support uh that you need to manage and you're looking for a a help desk solution, then definitely recommend checking out Unthread. But we're happy just to have people use the tool and again provide some feedback on how we can make it better.

Kyle James: Yeah, unthread.io. I'll make sure it's in the show notes for everybody.

Tom Bachant: That would be great.

Kyle James: I always like to leave it at this. Tom, what did I not ask? Anything, any kind of final parting wisdom or or something that you're like, you thought we should talk about, we didn't today.

Tom Bachant: I often get asked like what feedback I give to like founders just starting out because I guess I've been a founder now for like 12 years or something, 13 years. People often ask like what what what advice would you give your younger self and what what has actually been helpful? And I always like to tell people that as cliche of an expression as it is, the idea of having a growth mindset which kind of relates to this point of not taking yourself too seriously to the point where you always think you're right about everything, like you should be open to being wrong about a lot of stuff because you will be and always trying to learn and grow uh without getting stuck in a rut because yeah, where Unthread is today is so different from where we were, you know, three years ago and if we hadn't constantly had this feedback loop um where we're trying to learn and grow and figure out where we're wrong, what assumptions do we have that were wrong, like we would have just been stuck in an old failing business by now. So, um, I always say that's super important and the the phrase growth mindset has been co-opted and some people think it's really corny, but I think it's super important.

Kyle James: Hey, if you're not growing, you're dying, right? I mean.

Tom Bachant: Exactly.

Kyle James: So well now I'm curious like how much have you seen a pivot? Like it sounds like AI has pretty significantly pivoted the business then.

Tom Bachant: Yeah, uh definitely. We started solving this problem of just how do we organize uh like slack threads that are coming through. Like how do we organize a new issues being risen in a DM or in a in a slack channel? Um, and that didn't really require AI was a lot of like workflow management, like how do we there's some natural language processing that's involved there, but uh we didn't have access to the LLMs we have access to today. So once we did have that access and then the idea of agents coming around, which is this idea of autonomous LLMs that can actually check their own state and, you know, check their own work and how far along they are in completing a task, that's when we started to move into this whole like operational automation phase. But the initial version was just, you know, just ticketing, that's really how we got started and then, yeah, AI just opened up a bunch of doors for us.

Kyle James: And like we've talked about this whole episode, like all the stuff you could do summarizing and and and bringing things together and then finding needles in haystack that you couldn't see in this mountain of data like, boom, like it sounds like you found a whole, you know, a whole pivot that that's got a long way to go still, which is awesome to hear.

Tom Bachant: Yeah, and I'll say we got lucky that we picked a space where again, LLMs are really good at this task. If we were doing something like, yeah, again related to math or something, maybe we couldn't leverage AI as much as we do now, but we are fortunate that the the problem set we chose to solve just happens to be something that LLMs are really good at.

Kyle James: Well, Tom, I really appreciate you taking the time and hearing your story and kind of your your journey through this and thank you for sharing such a wonderful cool tool for everybody to be out there and play with and I I can't wait to hear how some of the people that watch this episode come and like dump their their support ticket database through and see how they're able to like prioritize and and set their roadmap up to for the future. So thank you once again.

Tom Bachant: Uh, thanks for having me. and yeah, like I said, I would love to hear any feedback from people. Uh, I thrive on that sort of stuff. So hope hopefully I do hear back back from some people on uh what we can do better.

Kyle James: Well, everybody out there, thank you for joining us for another episode of Prompted Builder Stories. Note that you can listen to this on audio. We do have it out on all of your favorite audio platforms as well as YouTube. So check us out there. Give us five star reviews, subscribe, thumbs up, all the good things. and until next time, keep growing and we'll talk to you then.