#295 Fergal Reid: Why Your Bots Fail and How Agents Fix Your Customer Support

This episode is sponsored by AGNTCY. Unlock agents at scale with an open Internet of Agents. Visit https://agntcy.org/ and add your support. Why do so many chatbots fail in the real world, and how can AI agents actually fix customer support? In this episode of Eye on AI, host Craig Smith explores how teams move beyond scripted bots to production-grade AI agents that resolve real issues across chat, email, and voice. We look at what makes agents reliable at scale, how to configure them safely, and how to manage them like digital workers alongside your human team. Learn how leading companies approach agent onboarding and governance, which pitfalls to avoid, and which metrics matter most for success, including resolution rate, CSAT, and cost per resolution. You will also hear how to enable actions like refunds and returns through secure procedures, design human handoff that customers appreciate, and build an omnichannel rollout plan that scales responsibly. ​​Stay Updated: Craig Smith on X:https://x.com/craigss Eye on A. I. on X: https://x.com/EyeOn_AI

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  With a lot of AI, while we've built this very deep product targeted at customer service, we have made it customizable and then customers will start to stretch it because it can just do a lot, it can deliver a lot of value. But we don't really want our customers, you know, writing individual prompts and trying to get down into the machine learning. Because, you know, we're able to go and test and optimize Finn across thousands of businesses. And that just helps us get better quality than you get by targeting anyone.

• 0:29Craig S. Smith?

  Build the future of multi-agent software with Agency. That's A G N T C Y. Now an open source Linux Foundation project, Agency is building the Internet of Agents. A collaborative layer where AI agents can discover, connect and work across any framework. All the pieces engineers need to deploy multi-agent systems now belong to everyone who builds on Agency. Including robust identity and access management that ensures every agent is authenticated and trusted before interacting.

• 1:13Craig S. Smith?

  Agency also provides open standardized tools for agent discovery, seamless protocols for agent-to-agent communication and modular components for scalable workflows. Collaborate with developers from Cisco, Dell Technologies, Google Cloud, Oracle, Red Hat and more than seventy-five other supporting companies to build next-generation AI infrastructure together. Agency is dropping code, specs and services, no strings attached.

• 1:51Craig S. Smith?

  Visit agntcy.org to contribute. That's A G N T C Y dot O R G.

• 2:00Speaker 00

  So yeah, so my name is Fergal Reid. I'm Chief AI Officer here at Intercom. I've been in Intercom about eight years, have a technical background, a PhD in sort of machine learning applied to network analysis, and degrees in computer science and maths and been kind of working in. AI and machine learning for for a long time now. And I joined Intercom about eight years ago to kind of start building really the machine learning function in Intercom. So yeah, that's me.

• 2:27Craig S. Smith?

  Okay, and we're talking about agents today and specifically Intercom's agent. Uh, is it an agent platform or is it an agent?

• 2:40Speaker 00

  Yeah, there's sort of a blurry line between those two things. Uh, we would describe Finn as more of an agent. And however, it's a highly configurable agent. You can really adapt it to do a lot of things for your business. We're we're still a little skeptical about very pure agent platforms. And they tend to not deliver the quality that's needed today to accomplish any one task end to end. So we've kind of been a bit more opinionated than that. We've built an agent product, uh, which is Finn. But we've done a lot of work to make it very highly configurable and customizable. But we don't really want our customers, you know, writing individual prompts and trying to get down into the machine learning.

• 3:20Speaker 00

  Because, you know, we're able to go and test and optimize Finn across thousands of businesses. And that just helps us get better quality than you get by targeting anyone. So, yeah.

• 3:32Craig S. Smith?

  Yeah, and Intercom, what was Intercom doing before it it uh, launched Finn?

• 3:39Speaker 00

  Intercom is traditionally a SaaS company. It's been around, you know, twelve, thirteen years, something like that. And really launched the business messenger way back when. It's one of these messengers that appeared in the little corner of a website and you could talk to people. Uh, Intercom was one of the first, if not the first, uh, to build a product like that, which then kind of went everywhere. And Intercom kind of evolved from that into a customer support platform. So you know, people at the other end of that messenger, customer support agents, answering customer support questions at scale, giving really good customer service. Intercom's mission traditionally was, you know, make internet business personal, really, really kind of give great product experiences on the internet like you might get if you had.

• 4:23Speaker 00

  Previously, you know, you went into a coffee shop and you had that highly personal service. So that's kind of what we did traditionally. And we're sort of an early adopter of AI. We had a previous generation product before Finn. Resolution Bot, we put a lot of work into to answer customer questions. And we were really well positioned when the current wave of AI came to sort of take advantage of.

• 4:44Craig S. Smith?

  Yeah. And and you're squarely in the customer service space, is that right?

• 4:49Speaker 00

  That's where we are today. I I think, you know, our our future aspirations are broader than that. But today, definitely, the product we have been is targeted at doing customer service and does a really great job of it. Now, customers do stretch that. Finn can do a lot. It's a very powerful product, can do a lot of functionality. So we do see customers stretching it where they're sort of, you know, they buy Finn primarily for customer service. But then they start doing things like, you know, upselling or customer success. So, you know, it's definitely an area in which our product is is being stretched and being pulled broader by customers. We have some customers who use Finn today and they use it entirely in a pre-sales way.

• 5:31Speaker 00

  So, you know, I think with a lot of AI, while we've built this very deep product targeted at customer service, we have made it customizable and then customers will start to stretch it because it can just do a lot, it can deliver a lot of value for them. So yeah, that's roughly where.

• 5:45Craig S. Smith?

  And and the agentic capabilities, what what can Finn do? And is it is the interface through chat or or voice? I mean, it's text or voice or?

• 6:00Speaker 00

  Yes, it's it's it's a pretty full featured product at this point. We've got over fifty million dollars of annual recurring revenue and it's used by like thousands of businesses. I think we've got more customers than any of the other sort of sophisticated customer service agents. And so the product, we've really built out the product. We've been doing this for over two years at this point. And so, you know, we have a very full offering. We built just Finn for the messenger initially, then it would run across things like WhatsApp and all the different channels. But then we built Finn for email. We had to build like a different architecture there to really deliver very high quality performance.

• 6:40Speaker 00

  So we're kind of doing that eighteen months ago. Over the last year, we've built a cutting edge version of Finn that uses voice, uses like the latest sequence to sequence voice models. That's now running at scale in production. There are loads of customers now, loads of end users who are sort of like talking to Finn over a voice interface. And it's magical. It it feels amazing to, you know, be, hear a phone conversation where someone is is talking to Finn and it's just kind of answering their questions and and so yeah, the product is getting getting pretty broad and it's getting a lot of capabilities.

• 7:12Craig S. Smith?

  Yeah. And as I said, the the agent tech capabilities, I mean, Finn, can you give us a use case where where it's it's not simply a question and answer, it's it's active doing something.

• 7:28Speaker 00

  Yeah. I mean, i- i- i- it does a lot of things. And and you know, obviously, I would say like, I wouldn't say just question and answering. Answering questions is is a huge part of of customer service and, you know, we we've invested a lot in building the best kind of question answering piece. And there's a big difference between agent capabilities in that space. But yeah, Finn can do lots more than that. So we have customers who do things like, you know, if they have us issue a refund, they have us handle returns. And and, you know, we've built a very rich and deep product to kind of enable it to do that, to to kind of, you can hook it up to your external systems, you can hook it up to your APIs, you sort of give it the parameters around that.

• 8:14Speaker 00

  We've built a very rich, you know, uh, almost SaaS style, you know, product to enable you to to specify and to trend together the complexity of the procedures that your business needs in order to do something like actually issue a refund. Every time we talk to a customer, they're like, oh, I know my refund policy. It's really simple. And like two weeks later, they actually finally get them to write down the refund policy on paper. And it's this big sprawling thing with all these edge cases. We've had to build a functionality to do that and to enable people to use L L Ms in those like real production use cases. So yeah, so i i it's it's pretty.

• 8:55Craig S. Smith?

  And and on on refunds, because that's a pretty complicated use case. Uh, Finn actually then uh pushes the button to release money or wire money or uh.

• 9:15Speaker 00

  Yeah, i i it absolutely does. We absolutely have customers who have deployed this, you know, where i in like consequential areas, in areas where if it gets it wrong, their business will lose money, their end user will be extremely disappointed. We absolutely have customers who are using it in those ways. And, you know, it's a lot of work to do that. Like, it's a whole different ballgame when you start trying to do things like that. You have to convince the security team that, you know, what what if a customer tries to to jailbreak this? Or what if a customer tries refund or return a different customer's product?

• 9:53Speaker 00

  You have a whole different level of complexity that you have to deal with there. You have to have like an API you can call. You have to like set up the API to work with Finn. And we've done a lot of work to try and make that easy. And we have like, I don't know, at least five years, maybe a lot longer of investment in this space. You know, we were building low code API integrations for customer service bots back before the current wave of L L Ms. And what we've done now is we have integrated the next generation AI L L M technology with all that sort of like battle tested, secure, and easy to set up interfaces of before. And of course, we support M C P. When we saw M C P coming, we were like, wow,

• 10:35Speaker 00

  this is amazing. This is a potential huge unlock. And so I think we're we're in the first wave of people to like really support M C P.

• 10:42Craig S. Smith?

  Yeah, yeah. And that's a model context, put a call that allows uh models to talk to each other or agents to talk to models, etc. The um. And in the customer, do you focus on a particular vertical? Or is this any customer service in any industry?

• 11:08Speaker 00

  No, we don't focus on a particular vertical. And, you know, we have lots of verticals that the product's very successful in. It's very successful in financial services. It's very successful in software as a service. And, you know, areas where people want to give really great customer service, where they're, you know, willing to pay for really great customer service, where it's worth it, where they want to use a very high quality product. And Finn does really well on. But we don't constrain ourselves to a particular vertical. And we kind of think it's probably a mistake to do that. And, you know, the reason why is that building a product that's horizontal like this, like there's so much commonality across the different use cases.

• 11:50Speaker 00

  The AI layer is, you know, you can invest deeply in the AI layer. We've done things, we've built custom models, we've done a great deal of testing and optimization. And so we really have this product that's like unified. And then we have thousands of customers running in production. And we're always improving it based on how well it does for those customers. And so like our product, our standardized product really has outperformed vertical specific competitors who have a much smaller volume. We really think that there are these really deep scaling and experimental laws in AI. You really want a standardized product envelope, a standardized product core

• 12:32Speaker 00

  that can be optimized and tested on mass across a very large number of customers, but then deeply customized for the individual business. And that's really what we've built. And so we're not in a world where we're telling customers to go and prompt the agent. And instead, we're doing, we've built a core that's standardized and that we have continually made better over time. But then with a customization there, it can adapt to a specific business.

• 12:58Craig S. Smith?

  Yeah, and most of the customer service solutions that I've talked to work in tandem with a call center or a BPO.

• 13:12Craig S. Smith?

  Does Finn do that as well?

• 13:16Speaker 00

  Actually, I mean, you can deploy it and we have customers that deploy it in every configuration. You know, it's a very mature product on a very mature platform, been in the space for a long time. Um, but I would say Finn is more likely to kind of compete with BPOs. Like, we typically see, you know, the typical pattern is a customer of ours will deploy it and they will often end a contract they had with a business process outsourcer that used to do maybe their sort of their tier one or their frontline support. You know, even a year ago, and the product was much less mature a year ago, even a year ago, we were

• 13:56Speaker 00

  talking to customers who were like, yes, I've deployed Finn and I have ended my relationship with my BPO. I still have my, my technical support team, I still have my own internal support team. But the BPO to which I was outsourcing, sort of the simple queries, um, is just gone and Finn entirely does that. And, you know, more reliably and more cost effectively. And that was even a year, year and a half ago, and Finn's really, really pushed up the quality level since then. And a lot of people's internal tier one support is what is now doing something else.

• 14:27Craig S. Smith?

  And, and so what happens when a customer hits a wall that Finn can't resolve? Is it routed to, to your internal customer support?

• 14:40Speaker 00

  Yeah, exactly. So, you know.

• 14:42Craig S. Smith?

  I, I should say to, to the customer's internal customer support, yeah.

• 14:47Speaker 00

  To the customer's internal customer support, absolutely, yeah. Again, we've been at this a long time and so probably like eight years ago, we were building Resolution Bot and previous generation tech. And while like the core engine's completely different and the modern AI is next generation and works much better. We were solving all those problems around escalation and how do you gracefully hand off to a human? And so we, we, we, we've, we've great solutions for that that are configurable and we have like almost a decade worth of technology investment in there. But again, we, we completely rebuilt, we built Finn completely from scratch, we were. At the door on G P T four launch day with our new G P T four powered product. And and so we, you know, it's it's it's an entirely new architecture.

• 15:31Speaker 00

  I think we're on generation four internally of it now. But the but it it's connected to all those old battle tested systems that we've had for a very long time. And so escalation, yeah, we've we've had great escalation hand over to humans for a long time. Very recently, we shipped a custom A I model trained ourselves. That's great at detecting exactly when to escalate, exactly when to hard escalate, when not to escalate, and when to offer an escalation to the end user. So we've even built like custom A I technology in there to get particularly good at that. That that's a problem that we we carry deeply.

• 16:06Craig S. Smith?

  And I'm curious, uh, i- is is there some metric about uh what percentage of calls end up being escalated? Uh, and is it is it lower with Finn than with other solutions or?

• 16:25Speaker 00

  Yeah, yeah. So so we we have um our our core metric that we care the most about, we call resolution rate. And it's essentially the percentage of times when Finn is involved in a conversation that successfully resolves the conversation versus when, you know, you have to escalate to a human. And that for us is about sixty five, sixty six percent at the moment. And that's sort of our North Star metric. We care deeply about it. It was about thirty five percent when we launched Finn. And we have this like team, my team's about fifty people in it. We care deeply about constantly moving that metric up over time. We're constantly doing A B tests. We're constantly taking any new change to the A I of Finn, running it at scale in production,

• 17:09Speaker 00

  hundreds of thousands of users in each version, and then checking to see, does it increase the resolution rate? And over time, we've managed to, through that scientific process, get that resolution rate up and up and up and up over time. So it's now around sixty five percent, still working on getting it up more and more. And we're incredibly proud of that. We think that's best in class. And anytime a customer kind of runs a head to head between us and between one of our competitors, a customer always tells us a resolution makes higher, they choose us. And occasion, we even um test specific competitors and we've always sort of won those tests.

• 17:45Speaker 00

  And so, yeah, we're really proud of that. And unlike, everybody says this, and we just always tell our customers, please just do an A B test, do a trial and see how you get on. And I, I think, you know, I think we're gathering momentum at the moment in the market as, you know, legitimately being one with it with the highest.

• 18:03Craig S. Smith?

  Yeah. And when, when you say it's, it's highly configurable, what, what do you mean by that? And, and when you onboard a, a new customer in an industry that maybe you're not as prevalent in, what kind of data do they have to provide to optimize the solution?

• 18:27Speaker 00

  Yeah, so I, I guess, you know, like all these questions, the nuanced questions. So the first thing I will say is, you know, Intercom traditionally, we believe a lot in like self-serve, we really like to build products rather than services. You know, what a product, you can maintain it yourself, you can change it yourself as we make it better, it gets better for all our customers. A lot of other people are attacking this via a services approach, they're, you know, putting engineers in and custom engineering it, and then the product doesn't get better over time, which isn't great. So, you know, look, we, we really kind of built this in a productized way, and we believe in productizing it for everybody, and then increasing the quality over time for everybody.

• 19:10Speaker 00

  And so, you know, initially on day one of Finn, you could set it live, and we had customers that set it live with no intervention, no onboarding, no information from Intercom at all. They would just go, they would configure Finn in our easy to use product, and it would be live and it would do a certain resolution rate for them. And, you know, over time, we've had loads of customers that have done that. However, we have learned as we have gone upmarket and as the product has matured, that there's an element of business transformation to this that our customers need. So now we have sort of built the muscle really over the last sort of six months, the last nine months, we've really built the muscle that we have to have teams of people. So we have a R and D services org, we also built a professional services org.

• 19:55Speaker 00

  Whose job it is to help customers get successful. So if you're an enterprise customer, you're an upper mid-market customer. You need a lot of help. You maybe need to negotiate and talk to your security teams, do these deeper procedures. Maybe you have a voice deployment. You need to get that voice deployment live. It's on very big scale. You need that help. You need that services team. So we've kind of, we've been forced to build that muscle. Probably about six or nine months ago, we didn't have a great muscle. Nine, nine, twelve months ago, we didn't have a great muscle for that. Some of our competitors were starting to do it. Some of our competitors were starting to make headway at us because they were going and kind of by hand in an unsustainable way, probably.

• 20:41Speaker 00

  They were getting their core teams to like build a product for each customer. We thought that was a bad idea, but we needed to compete with it. So we built this kind of services org. We've really upped their game there for those large enterprise customers. So. Yeah. So, you know, today we have customers that turn along. They get like fifty, sixty percent resolution rate out of the box. It's transformational for them. We also have a motion now for the big, really big enterprises where it's like, no, we can go deep. We've got a services team and do all that transformation with them, you know, that the business process change that is is hard and complicated and it needs to be gone through.

• 21:19Craig S. Smith?

  Yeah. And and when you go deep, as you say, are you, are you fine tuning the, the, the, the L L M behind the agent? Uh, or, or is it, uh, you know, putting rules in place uh, around the agent?

• 21:40Speaker 00

  So, so

• 21:42Speaker 00

  got it. So, so we, we, we do not fine tune the L L Ms on a per customer basis. And we're suspicious of people who say they do that because you need a very large volume of data before that makes sense. Um, maybe the biggest customers in the world, uh, that might make sense. But even then, we're much more in favor of fine tuning L L Ms that will work for customer service generally, because then you have this huge amount of data from many, many different customers. And in our experience, that that tends to outperform training for any one customer. So that's really the way we go. So when we talk about configuring deeply for an enterprise customer,

• 22:23Speaker 00

  we've built this product-based configuration layer on top of Finn, above the level of the L L Ms. So, you know, essentially it turns into prompts for the L L Ms, but managed prompts. You know, we don't believe that customers really benefit from having the ability to rawly prompt a large, complicated system like Finn, which is maybe ten or fifteen different prompts underneath the hood. Instead, we have learned the control that they need for their business. We've given them easy-to-use windows to get that control. So to change Finn's behavior, to change its tone of voice,

• 23:04Speaker 00

  make it compatible with their brand, but without them kind of messing around accidentally in the core L L Ms. So i- i- it's quite configurable, it's quite customizable, you can change, you order it, those things, you kind of can change the policy a lot, sort of the mechanism, you know, a lot of A I products, they accidentally mix up customer control over like the policy with customer control over the mechanism. And then you end up with all these failed deployments or mediocre performance. We don't do that. We kind of have a standardized mechanistic layer and then like

• 23:45Speaker 00

  configuration and policy layer that we have done the hard work of productizing that we expose. And that way, customers, they they get that improved performance of that kind of core engine getting better over time. They get that while maintaining the configuration layer they need. Sometimes there's a trade-off there, but you know, the hard work of building a product is to kind of give customers the configuration options that they need. And I think we've done that. We've been doing that for about a year and a half of Finn's lifecycle. I think we've, we've really done that. They can give a guidance without accidentally breaking.

• 24:19Craig S. Smith?

  Yeah, you know, this is a crowded space, as I'm sure you're.

• 24:25Craig S. Smith?

  You're aware more than me. Uh, how do you, is the market so large that it doesn't matter? Uh, there, there is, there's always another potential customer. Uh, if you, if you, you know, lose one, but, or is it, um, I mean, what strikes me about this space is, and I've talked to a lot of companies in the space is, uh, I, I still, uh, I don't, I don't think yet in my personal life, I've run into one of these, uh,

• 25:07Craig S. Smith?

  generative AI, a gentic customer service solutions when I'm dealing with companies. I'm still going through the, the phone menus. I'm still, you know, getting, uh, you know, the canned answers from chatbots. Uh, so I guess the question is, it appears to me that it's, it's, it's kind of an endless market at the moment. Uh, how do you deal with competition? And why is it taking such a long time to penetrate? Is it just companies of have bought a solution and they're going to amortize it out before they switch?

• 25:52Speaker 00

  Yeah, I, I, I think an interesting question with a couple of different pieces to it. Um, I, firstly, I think the market here is huge, like the, the total addressable market here is like, absolutely massive, right? So many people doing these kind of rote customer service tasks. It, it's going to go away and like all human customer service won't go away, but all the road stuff will. And, and even a lot of the stuff that's not rote, a lot of the stuff where it's like, asking a question that can be answered from any sort of knowledge or documented knowledge, we think is going to go away. And so we need a massive market here. In terms of competition, we legitimately believe that we have the best agent.

• 26:36Speaker 00

  And we have a lot of data to back that up. And, you know, we, I think we're starting to win in the market. I don't know anyone else that has. We, I, I need to pull the exact stats. We definitely have over five thousand customers, fifty million, over fifty million in ARR. We're growing at about like, you know, we're on trajectory for a hundred million in two quarters, growing about four X year on year. Like, so, you know, there's an adoption curve, but we're growing very, very well. And so, you know, feeling pretty good. There's always an adoption curve, right? If a business wants to buy something, it's not a consumer product. There's always an adoption curve. There's always, you know, stakeholder management, the existing solution, there's all these concerns, you know.

• 27:21Speaker 00

  Is it good enough? Can I adopt it? And we've been going through that, you know, for almost two years at this point. And it takes time, but we're seeing that very rapid growth and, you know, really the market is turning. And so we're feeling pretty good. And then in terms of like competition, yeah, we have competition that's growing really fast as well. And again, huge market really expanding. So good luck to them. And but we are setting out to win and then, and we think we've, we're really quite sure we've better technology, we've better products, we have a very big investment here. And so, um, yeah, that's, that's all compete and see who can give the best product and answer the most end user customer support questions.

• 28:03Speaker 00

  That's how we build. We build a dollar per resolution. We resolve the question, we get a dollar. If the end user is not happy and they talk to the human, we get no dollar. So that's sort of like putting our money where our mouth is in terms of our belief in, in having the best product and, and making it work. And it's a lot of vaporware out there. There's a lot of big claims out there. And our retention numbers are really good. And then let's see how the market plays out. But yeah, we want to build the best, uh, really good product for our end users and that we're going to make a good go of it.

• 28:36Craig S. Smith?

  Yeah. Uh, the, um, do you think, uh, on the adoption curve?

• 28:45Craig S. Smith?

  We're, we're still just ramping up. I mean, I know you're not in sales, but how long does it take a company to make a decision? Uh, to to to adopt uh, Fin for example.

• 29:00Speaker 00

  So I mean, thi- this market is very structured, right? You have the top tech companies. We have some great customers. Tropic is an amazing customer of ours and obviously really AI forward company. Other companies like AmpliJuts and Tisia, like really, really good and kind of tech companies. They make decisions quickly. They run. They're very sophisticated. They're very technical. They can run tests. They can really assess the quality of your product. On the other hand, we have customers who are, you know, more like a utility or a financial services company.

• 29:37Speaker 00

  And then they insurance uh, customers like, you know, th- th- th- they they will have more diligence to do. They care a lot about like. Is the regulator okay with this? Is this going to get me into trouble? Can you really be sure this won't hallucination production in a way that will cause me trouble? And so, you know, there's an adoption curve and different people in different industries are in different curves. So th- there's no one answer. I I can talk about the the growth of the product overall. And I think the growth of the space overall to, you know, the whole space is growing. And and so, uh, so yeah, I I think it's maturing all the time. It it is fast. It's a very fast growing space for B two B.

• 30:21Speaker 00

  But B two B takes time, you know, these are not consumer application.

• 30:25Craig S. Smith?

  And and do you think that adoption will accelerate? Uh, as as companies realize they're gonna get left behind if they don't upgrade their customer service. Uh, because there's competitors at at their level.

• 30:45Speaker 00

  I mean, again, the the the the most honest and clear and data-driven way for me to answer that is to say like, Fin has had. Double digit eh month on month growth for at least eighteen months, right? And you know, certainly, maybe, maybe, maybe one month, it's a little bit seasonality effects. But on, on average, trading three month, double digit month on month growth eh for at least eighteen months. And so like, that is an accelerating adoption curve. That is the kind of curve you see when a market is maturing and when the product has got like good product market fit. And so, um, yeah, I mean, I, I, I think, I think that's the best evidence I have of an accelerating market adoption.

• 31:30Speaker 00

  And you know, there's also the things, there's this viral dynamics, people, people do encounter it out there in the world. And then they look and think, wow, this is great. I want this for my business. And so, you know, I think we're, we're sort of seeing that. But we're still early here. This is a huge, huge, huge market and it takes time to penetrate through it. And, and you know, like voice, we have built, our, our voice product is a next generation voice product. It's built with sequence to sequence models. We have competitors, even very new competitors who kind of build the voice for like text to speech and speech to text. It's kind of janky. We've gone like for the newest models. It's really, really great. You know, we couldn't do that. We couldn't have built that product a year ago.

• 32:12Speaker 00

  So like voices is like the, the paint is still wet on voice, but it's really great and it's next generation. And so, um, yeah, I think you're going to see continuing adoption. Like AI for voice, there's no comparison with the best models now versus where they were like even eighteen months ago, you know? So, yeah, very exciting.

• 32:33Craig S. Smith?

  Uh, do you think that because the market, there are so many players in the market? Uh, I mean, obviously there'll be a period of consolidation at some point. How far do you think we are from that consolidation where uh the winners will emerge?

• 32:51Speaker 00

  Yeah.

• 32:52Craig S. Smith?

  They'll either buy up the, the smaller competitors or the smaller competitors will fold.

• 33:00Speaker 00

  It is so hard to know. I think we can only wildly speculate on that. I think huge markets with an absolute ton of value being delivered. And so like that would kind of tell you that like there'll be funding and opportunity to continue grow without a creative consolidation for a long time. On the other hand, um, you know, our product Finn returns, our product Finn returns a healthy margin to us. Uh, I believe there are lots of businesses in the wider AI space where their margins are more questionable. And so, you know, people are just optimizing for growth at all costs. And it is possible therefore that there will be a shakeout or a period of consolidation if something happens in the wider

• 33:44Speaker 00

  macro environment or in the environment of AI to kind of to drive a cooling. And overall, I would probably welcome a bit of a cooling. I think we've a great product. It's not just hype. It's got positive margin. So it's suited strategically if there was a little bit of a cooling. I don't know if we're going to get that. I do think that, you know, the deep seek moment was interesting. I think the Chinese AI models are very, very powerful and compelling. And so maybe that will lead to a cooling. Um, I just don't know. It's a big space and that's a real macro question.

• 34:20Craig S. Smith?

  Yeah. Well, that's interesting because what what kicked off this space really was uh, the uh, was generative AI and you know, ultimately the transformer algorithm. It, I mean, you're a technical guy. So looking, what what do you think has to happen?

• 34:48Craig S. Smith?

  For the market to go through another uh, surge, another, what breakthrough are you watching? What bottleneck are are you guys uh, working on?

• 35:02Speaker 00

  Yeah. I'd say I'd say a number of things there. Firstly, I think we're, I believe that we are in overhang at the moment here. Where like from an industry point of view, even if the AI models completely froze at their current perspective, at their current capability. We would have like at least a decade of huge growth industry to deploy all these things to many different valuable applications. So I really believe that. But I do think that, you know, the core AI capabilities continuing to improve, and they are continuing to improve.

• 35:38Speaker 00

  Is an accelerant to the whole space overall, top to bottom. And, you know, really interested in things like reliability improving. We're like constantly working in our procedures product, you know, the one that like takes the actions and calls APIs. We're constantly working to get the best reliability we can out of L L Ms. And, you know, we're working and we're investing in trying to improve that reliability at all these different layers. And also, we've really had some improvements recently in, you know, the sort of the the small models hypothesis. Which is this sort of idea that you can take smaller large language models.

• 36:18Speaker 00

  And you can train them to be specifically good at a given task you're interested in, really specialize them at that task. Kind of take a model that knows how to speak the language, knows how to retrieve a lot of information, maybe knows how to do some reasoning. And then you you fine tune it in a sophisticated way to be really great at a specific task. So we have a sub part of Bin where we we summarize the end user's question. It's really important to do that. Because sometimes, you know, before you go and go searching and you go trying to retrieve content, we really want to kind of canonicalize or summarize your question. We built a custom model that does that by kind of like fine tuning and one of the the kind of the Qwen models.

• 37:00Speaker 00

  And that's performance is really excellent. And that's replaced a call to a much bigger L L M with like better performance overall. So I think you're going to see more of that. I think that the whole tech stack, top to bottom, is like radically changing. There's innovation at all levels on it, whether it's things like replace an L L M with like a smaller custom model. Build a custom model from scratch. We've deployed a whole lot of custom models we built from scratch and Finn as well. And then also at the at the kind of the frontier L L M layer, they're getting like better and better all the time. So, uh, yeah, innovation are all.

• 37:35Craig S. Smith?

  Yeah, uh, on you, you mentioned insurance and banking. Uh, those are, you know, highly regulated industries where,

• 37:46Craig S. Smith?

  their data is valuable and proprietary.

• 37:52Craig S. Smith?

  You guys are,

• 37:55Craig S. Smith?

  excuse me, you guys are a SaaS product, uh, operating, operating from the cloud. Uh, what, what do you do when someone needs an on premise solution?

• 38:08Speaker 00

  We, we don't currently provide an on premise solution. And, and, you know, that, that's a long trade off in SaaS. Obviously, you know, we describe ourselves as an, as an A I company now. And I think it's, it's going to be a long trade off in A I as well. And, you know, and it's just, it is a velocity trade off. If you go on premise, it's very difficult to keep your product development, life cycle iterating fast enough. And there's just, there's so much value to be unlocked in the world for customers that are willing to trust the cloud. That, that's where we're focusing for the foreseeable future. And, you know, the customers we have in these regulated industries, they, they, they've crossed that bridge. They've crossed the bridge to the cloud a long time ago.

• 38:51Speaker 00

  And not, not every bank has, but many have. And so we think there's, there's a ton of market there. And we're not going to kind of slow down and take the velocity hit of, of going on.

• 39:01Craig S. Smith?

  Yeah, that's interesting. Oh, yeah. I just asked because you talked about smaller models. Uh, whether you were looking at the, uh,

• 39:11Speaker 00

  We, we, we, we could, we, we, we, we could deploy. We, we have enough of a stack now. Dash will run entirely on models that we could deploy, that we could deploy to somebody's, cloud if we wanted to, or on premise if we want to, but it's just strategically not the direction.

• 39:26Craig S. Smith?

  And, uh, on the, on the A agent capabilities, uh, and you alluded to this at the beginning about how customers are stretching, uh, how they're using the product. Are you looking beyond customer service? If you have this this powerful solution that presumably could run other business processes?

• 39:54Speaker 00

  I think that is a very interesting area that I think we're not quite ready to talk about just yet. But it is absolutely an area of interest to us. How could it not be? We have this product, it's excellent to customer service. And then we have customer stretching it, using it for pre-sales use cases, getting a ton of value out of it, using it for success use cases. And so it's a very interesting and insightful question strategically. But I think we're not ready to talk.

• 40:22Craig S. Smith?

  Okay, well, this is fascinating. And if people want to give Finn a try, it's fin.ai or what is the?

• 40:34Speaker 00

  Yeah, absolutely. If you go to fin.ai and again, it's a very self-serveable product. We have put a lot of energy into that. And if it's an enterprise customer, we definitely say get in touch, reach out to sales. If it's like an SMB, we'd say, give it a try yourself. You can probably unlock a ton of value. We do, we do the work of making sure it's self-serveable and productized.

• 40:56Craig S. Smith?

  Yeah, and actually I do have another question. Is this targeted for and priced for, you know, Fortune? One thousand or something, or is, does it, is it affordable for a small and medium sized company?

• 41:18Speaker 00

  It is absolutely affordable for a small and medium sized company. We really pioneered outcome based pricing here where we sort of, we charge a dollar per resolution. And we're quite proud of that. And that means that if you have thousands of port queries a month, you know, and it's taking you, you know, several humans' worth of time to do them. And you can afford to pay that one dollar. Also, if you have a huge support organization and you're outsourcing to a BPO, it should also work for you. So we're pretty happy that that price really scales with the value that it's going to deliver for the customer. And, you know, we have history around that. Intercom kind of got its pricing wrong in the past in a previous generation. We've really learned from that.

• 42:00Speaker 00

  We really believe very strongly in fair, transparent, outcome-based pricing and seems to be working pretty well for us. Are we like? Squeezing all the value out that we could, possibly not, but that's not even the right way to think about this. This is a growing market and growth is all we care.

• 42:17Craig S. Smith?

  Build the future of multi-agent software with AGENCY. That's A-G-N-T-C-Y. Now an open source Linux Foundation project, AGENCY is building the Internet of Agents. A collaborative layer where AI agents can discover, connect, and work across any framework. All the pieces engineers need to deploy multi-agent systems now belong to everyone who builds on AGENCY, including robust identity and access management that ensures every agent is authenticated and trusted before interacting.

• 43:01Craig S. Smith?

  AGENCY also provides open standardized tools for agent discovery, seamless protocols for agent-to-agent communication, and modular components for scalable workflows. Collaborate with developers from Cisco, Dell Technologies, Google Cloud, Oracle, Red Hat, and more than seventy-five other supporting companies to build next-generation AI infrastructure together. AGENCY is dropping code, specs and services, no strings attached. Visit agntcy.org to contribute. That's A-G-N-T-C-Y dot O R G.