Resolution-First AI: Support as a System, Not a Headcount Line

Machine-generated transcript from the episode video. Speaker labels are not included and some names and product terms may be transcribed phonetically.

[0:00] I've learned that the best founders are always selling. Welcome to the GTM Vault podcast hosted by Rick Kleta. We uncover the strategies, the pivots, and the breakthroughs. Turning startups into giants. Let's crack open the vault and find out how. Today on GTM Vault, joined by Raphael Brochi, co-founder and CEO of Notch, a company that's reinventing customer service with resolution for AI agents. Notch isn't about chat bots or reply only AI. It acts like your best support rep, only faster, smarter, and always on. The results: millions of interactions resolved end to end. 40x faster response times, 80% lower cost to serve, and doubledigit lifts and conversion rates. Executives at companies like me, Mixtiles, and Guardio turned to notch to scale without headcount. Stay lean while improving seesat and free human capital to focus on innovation margins and growth. Before founding Notch, Raphael spent a decade in the Air Force led product strategy at

[1:12] AGT International and invested with foundation partners. Experiences that shaped how he builds AI that works in missionritical production environments. In this conversation, we cover why reply only AI is dead end and what resolution first unlocks. How autonomous agents are already closing millions of tickets for global brands. The GTM playbook for replacing bloated support orgs with lean AI first operations. What the future of customer support orgs looks like when AI agents go mainstream. All right, let's get right to it. You know what I see a lot? companies with solid strategy but terrible execution. They've got dashboards, reports, tools everywhere, but no one knows what to actually do next to scale their business. That's why I've been paying attention to what Zoom Info is doing. They're not just the contact data company anymore. They built a full system of execution. They're calling it GTM intelligence, and it actually works the list, writes the outreach, and triggers the play. No guesswork, no manual grind, just pipeline moving, predictable growth

[2:22] strategy that actually delivers. Check it out at zoominfo.com. You went from building a product in the insurance space to pivoting into CX. What painoint in CX convinced you this was the problem worth solving? Well, we just felt that support is just not going good enough. It wasn't about reducing headcount. It is a byproduct, but essentially I don't love waiting on a phone for some things that are pretty simple to handle. Not every kind of support issue requires empathy. A lot of those, I'd say 85 to 92% depending on the company, could be resolved simply by an AI that knows the company business. General FAQ bots are a dead end. Why?

[3:06] And what does resolution first actually mean in practice? So I'd say that the kind of bots that people usually see now are gen two what I like to call. So gen one was the were the fixed menu bots that we came to hate. Gen two are essentially GPT rappers is what they're called. So those are kind of tools that do an API call to an LLM whether it's by open anthropic or any of the other companies building foundational model. And what you do is you connect your knowledge base and you use some sort of semantic retrieval technique in order to get that knowledge. And what that does is it provides a very good foundation into answering very simple questions about the business. But if you look at the amount of time spent by agents in companies we work with, 90 to 95% of their time does not go towards answering FAQ. Those are very very simple stuff that they do very fast. So the kind of Gen 3 and what we've developed and are still developing are the kind of AI agents that are able to solve very complex processes and completely disintermediate humans from the work of tier one agents allowing them to focus on more high value work.

[4:19] Why do LLM cap out at 70 to 80% accuracy and what does it take to get to 95% plus in production? Yeah, so that's a good question. What we've seen is that usually when you talk to chat GPT or normal FAQ bots, as we just discussed, the accuracy is between 70 and 80%. But sometimes consumers don't really feel that because if you're just asking a general question, you might get information which isn't accurate and you're not going to know about it because you're asking about ingredients in the product or even if you get something wrong and you understand, you might ask again and get the correct question. And why does that happen? Is it because LLMs are not deterministic? Meaning that if you provide them a large amount of data or small amount of data and you ask the same question in the same way, it could be a difference of one or two words, you might get a different answer. Now, that answer does not mean it will be completely opposite of what you got before, but it could be a bit different. That's one. Second, it's very, very hard to kind of evaluate on a large corpus of data that a company has. So it could be from what we've seen between 200,000 and a million words is the average knowledge base of the company and on that amount of data you can ask a billion different questions.

[5:35] Okay. So it's very very hard to test out. That means that even if you run extensive tests and your evaluation mechanisms are pretty good. So there will be a lot of edge cases, a lot of questions that you have not answered that are not related to let's say a specific piece of information that you might have in your knowledge base, but kind of a connection of a few of those. more complicated questions and that means that you usually get the 70 to 80% uh uh accuracy rate and that's not good enough if you really want to handle the kind of complex processes for example following a company policy that at the end of it you have to call an API or do some sort of back office section to actually issue a refund in that case you have to be in the 97 98 99% now to get there that's that's the complicated part like everything in the parto principle right getting to 80% % is simple.

[6:26] Getting the additional 20% requires five times the work or 10x the work together. You must use different kind of systems and not just one LLM. You have to build some sort of orchestration platform meaning a combination of many small AI agents each doing a specific part of the task plus connecting them with some deterministic rules. Walk us through the agentic architecture workflows. deterministic rules, LLMs, how do they work together? So that's that's a good question. I'll say it really depends. Each kind of system or each kind of process or use case requires a different architecture. But I will say that the building blocks are exactly the kinds that you mentioned and for us there are two main building blocks. Either you use deterministic rules meaning rules uh that are either true or false that rely on data. So that could be data from the back office. For example, if the user that is talking to me is from the US, then I might do one thing and if they're from another part of the world, they might do another thing. I don't need an LLM to analyze that because that might not provide a deterministic response. So that's why we use deterministic rules in those cases.

[7:41] Uh and that's kind of like rulebased system. That's one. The other part is having the semantic understanding or what we call our textual conditions. For example, if a customer says, "Hey, I'm not satisfied with the item I received or I want to return the item I received." According to the company SOP, standard operating procedures, we might need to ask, okay, why do you want to return the item? Then the customer might say, "It's damaged. I don't like it. It arrived too late. I don't need it anymore." Now, those are the kind of conditions that are not deterministic, right? You cannot use some sort of data point to understand what the customer says. You have to analyze that. You have to analyze the textual condition here and according to that decide what are your next steps. So each one of those is is essentially a point in the process that might split your branch of the tree saying okay what do we need to do next?

[8:39] So when people think about or talk aboutic architecture which is now a huge buzzword essentially we're talking about workflows it's not that different than other rulebased system the main difference here is that how you go along within that branch or what are the things you have to actually define yourselves or things that could be generalized are very very different but at the end of the day if you want to solve a customer's issue you need to adhere to some sort of process process that is defined by the company. That's the SOP. Everything we do in life at the end of the day is a process. So just thinking that you can throw it all to an LLM and make it reach a very high percentage of accuracy is not the way to go from what we've seen in production environments where in customer support specifically, you must be able to adhere to the company policy almost at 100% accuracy rate. Otherwise, it's like zero. And that's that's the main difference there. So everything you talked about the agentic architecture, workflows, deterministic rules, LMS at the end of the day, the idea here is how do I make sure I can work through a specific customer issue or process and

[9:53] have as much flexibility as I can to not need to define every specific edge case in a very guided way. Otherwise, it will take us years to implement a company. How do you replicate performance, tone, and reliability of a company's best agents? Yeah. So, here I'll just tell people what I always tell our prospects. There is no silver bullet here. Some companies might say, "Hey, just provide us a list of your historical tickets. We'll get you up and running in a week." This never happens. At the end of the day, I would consider the implementation of an AI agent very very similar to how you would onboard a new employee. Meaning, it's not enough to tell the employee, hey, here's our ticketing system. Go ahead to that room, sit next to the ticketing system, go through a thousand tickets, come back to me, and we'll get you set up as an agent. That doesn't work, right? People need to understand the logic behind everything. So when talking about tone of voice or company policies, SOPs, at the end of the day, those are all guidelines. Even in tone of voice, you might have the general guidelines, right? Be empathic, be professional, use up to four emojis, write in short sentences, always

[11:04] apologize, never apologize, sometimes apologize. So the first thing we do is we really train ourselves, our team to be as good and as proficient as the company's agent as we go along. That's that's the first thing we need to do. Once we do that, we need to take the kind of data that we understood and transform it into the kind of language that our system uses inside the notch platform. Once we do that, from that point on, we start using a feedback loop with the company's agent to retrain our agents in an automatic in an autonomous fashion. Now after doing all of that then we can look at historical tickets and actually understand the logic behind what happened same as a human.

[11:51] Not just positioned to replace bloated teams but your e-commerce case study showed it can actually grow revenue tripling VIP campaign response rates. How do you sell both cost and growth in the same story? Yeah. So, one of our clients who's a leading e-commerce brand, enterprise e-commerce brand, came to us with a problem after we've already implemented our AI agent for customer support and reach significant levels of automation. They said, "Look, uh, we have our VIP club and we want to do something with it." We started digging in. We asked is this VIP club? They said, "Look, every person that buys over a specific amount of money uh in a specific period, we add them to the VIP club and sometimes we send coupons or specific offers to the VIP club, but the response rates are normal, right? We don't see more upsell or more money from those kind of customers. Although we know that those kind of customers are customers that are supposed to spend more. So, what we said was, okay, let's change the narrative, do something a bit different. Let's not try and push. Let's do what people actually like when working with luxury brands. And what we found out is that customers really like the SA, the sales assistant function. And what a good

[13:05] sales assistant does, like any salesman, is they build rapport. They build relationship before they try to sell. So we said, let's do the same thing. Now, instead of sending your outgoing SMS campaign and just saying, hey, welcome to the VIP club. Take 20% off. What we said is let's build an AI agent called Ella. She'll be the personal essay. Now, the first time a person joins the VIP club, instead of just pushing an offer, just tell them, "Hey, this is Ella. Welcome to the VIP club. I'll be a personal essay. If you need anything, let me know." And people started talking to Ella, asking questions, asking about offers. And what we've seen is that after three or four interactions, once we pushed those outbound campaigns with the kind of coupon and the kind of offer that the company wanted, the response rates and conversion rates were 3x the amount. So in terms of how do we sell CX or growth, I'd say both areas are valid, but in 90% of our cases, a CX is the wedge. So that's usually the first place where enterprise companies look at for cost savings. Uh and that's where we come in. There's a lot of work there. Uh

[14:14] but once we come in, we grow, we expand, and we usually move on to two, three, four different use cases. What metrics resonate most with execs when deciding to adopt seesat cost to serve conversion lifts? Yeah. So, usually the execs that are I'd say more within the weeds. So, the VP customer support, that's definitely seesat and cost to serve amount of agents. when you go higher up to the COO that's a bit more broad right so it could be general satisfaction and it could be the fact that they want to see less outliers what are outliers so what sometimes really bothers management is the kind of things that happen when customer support goes south very bad reviews reports to the BBB or other kind of functions or even DMs or emails to the CEO of the company that in addition to the kind of goals they have usually from the board of the company they need to cut costs so I'd say it's a combination between ces set and cost to serve you cannot talk about cost to serve if you cannot first make your customer make our customer the company feel convenient with the quality so what we say is we're going to keep your seat

[15:30] at the same level or higher average is 10% higher and then we're going to be able to help you create more efficiency by up to 50% of your customer support within x time depending on the complexity. Which industries are adopting fastest and what's your early ICP wedge? So I think that was really surprising when we started when GPT 3.5 and then four came out. E-commerce was definitely the fastest. Usually those are industries and and companies with low AOV and high uh transaction volume which fits perfectly to use AI tools. But in terms of enterprise, what we've seen that surprised me, finance works are adopting pretty fast. And I'd say that the more commoditized an industry is, the faster they're going to adopt. So I think what we'll see is that telecom finance are adopting the fastest even health even though regulations are very tough in those industries specifically finance and health but still that's one of the only ways uh to really differentiate yourself in a commoditized environment is by offering better service. So the kind of companies that will be able to adopt fast and live and

[16:44] learn to live with the growing pains of adopting a new technology I think will gain a very significant edge on other companies in the field. What's been your most effective GTM motion and any channels or plays that didn't work? Yeah. So I think one of our main GTM channel that we use, I don't know if people will call it a GTM channel, but that what works best for us is the fact that we're managed service solution. Meaning when we uh talk to a company, we don't just sell a platform, we sell the solution. We're in charge of the delivery. And when you provide excellent service and when you reach the 70% automation, usually we get three to five very very warm leads from that company. So for us, I'm looking at the managed service aspect. How many people I'm going to put to work on a specific company also as a as a as a marketing expense, right? So if I have the ability to say, let's put two people on this project, I much rather put a third person and say, okay, that's an additional $70,000 on the cost of the project, but this might get us three more clients worth $2 million, for example. So, I rather put this kind of

[17:56] effort than spend an additional $70,000 on LinkedIn or Google or anything else. And are there any channels that you tried that didn't work? Well, we keep it pretty focused. We're not a huge team. Uh, so I'd say LinkedIn, Google, what I mentioned about our unique strategy. Content definitely works. So, those are the kind of emotions that work best. But again, we appeal we appeal to mid-market and enterprise. So, it's very very referral heavy. And how do you position that duality of selling both a product and a managed service? Well, I think that the kind of um way to sell an AI tool right now for enterprises is that at the end of the day, this is a workflow product, right?

[18:44] So a company comes to you and say, "Hey, we have this specific process that would love you to automate. Essentially, that's what RPA did before. Now RPA is incredibly not robust and not flexible. Takes a long time to build and cracks easily. But essentially, if you break down CX or any other department, it's all processes." So a company now with the technology that exists in AI might give you a broader range of processes to automate. So the entire CX department or the the entire tier one. Now to do that even if we worked with 10 different companies in that field 10 different finance company the 11th company will still be different meaning you have to really and clearly define and sometimes customuilt things for that company. So I'd say those two things building the product or a platform which is essentially a back office platform to help our managed service team implement faster those things go uh hand hand in how quickly do you see autonomous AI going mainstream 2 years 5 years 10 two to five is my guess I don't know exactly when and why I do think that even now there isn't one company in the

[19:58] world especially in B to C that is not talking about AI specifically AI in customer support. Now again talking and doing are different. We see that PC's usually last 3 to 6 months. Then from post PC to production it might take a year. A lot of projects fail for a lot of different reasons and good reasons sometimes. So I'd say that right now we're in the first inning. companies are either starting PC's or have just finished PC's are in the first phases of going live and within three to four years I think they're not there isn't going to be any major companies that have not launched an AI autonomous agent at one of their department what's overrated right now in AI for CX and what's underrated yeah so I'll say AI SDRs okay it's not exactly CX but what I see CX is not customer service It's customer experience, customer interactions. I think the reason I see AI SDRs is overrated just because being an SDR is probably one of the toughest jobs in the world. Not a lot of people can do cold calls successfully or send great cold emails. And I think that at

[21:10] this point you want uh to work on jobs that a lot of people can do, right? So even though the promise is huge, we have to think, okay, what happens when AI SDRs are great and people start getting a thousand phone calls a day, no one can answer the phone or you might be able to talk with your prospects AI agent and that's going to be a barrier. What's underrated in AI? Nothing is really underrated right now. I do think that some things more in the deep tech space which I'm not going to touch in terms of technology are underrated and LLMs kind of took all the glory although it's it's at the end of the day a small part of the entire AI space if resolution first AI becomes the norm what does the average CXO or look like in 2027 I'd say that looking at a company at one of our company that has reached maturity in terms of of our product adoption and let's assume that's the way that uh the the industry will behave. So what happens in this company is that tier 1 agents shrink by 85 to 95%.

[22:22] Shrinking the entire CX department by 70%. out of the people that are no longer working as tier one in CX, I'd say that about a third of them become or have different positions related to AI or a third to a half of those are usually pivoted to other areas in the business. Saving costs on CXs is great, but taking quality personnel that already know your business and pivoting them more towards revenue generating operations is usually better. So what I see is that you have a small number of people that have a command and control platform like Notch. They're operating I'm sorry an army of agents. Now they don't really operate them. Those agents operate autonomously. That's going to be like looking like a developer looks at logs, API calls, everything just works.

[23:16] Suddenly something happens. You get an escalation. you're able to work through those kind of escalations and push changes incredibly quickly. So what happens now? I see CX as just an arm, right, of a company. That's the part that actually that's where the rubber meets the road. That's where your product and service meet your uh end consumers. And the problem right now is that kind of feedback loop, the kind of insights that get to CX sometimes get lost. And then if something happens that you need to move fast to to counteract an incident or a bug or something like that, it takes an incredibly long time to digest that and afterwards um decide on what needs to be done. Once you have a command and control system that is able to immediately digest different insights from what happens in the field and suggest what you should do in order to counteract that effect, I think we're going to see a multiplier in the amount of feedback that companies get that would actually help them build better products much much faster.

[24:25] For founders and exacts listening, when should they start planning or design around this shift? So I'd say once you reach your product market fit, doesn't matter in which industry and once you have an operation that is starting to scale, you need to really think about how you build your processes in a way that is very organized. So I don't think that means you have to really understand how LLMs work and think about how to arrange data that will be good for an LLM. As a rule of thumb, if the data is well organized for a human being, it will be very good for an LLM. Okay, so that's what you need to think about. That's one. Second, try and decrease the amount of areas or the amount of gray areas in your processes. Okay, so every time you say it depends, this is not going to be good once you scale anyway, right? Once you reach a 100 people within your support org, the it depends part needs to be as small as possible. So start handling those things early.

[25:26] Make decisions. They can always be changed, but going through all of those gray areas once you have 300 people in your support team might take months. So, I'd say starting to scale, think about how you build your processes, your kind of business documentation, what we like to to call the business operating system as clearly as possible. It'll help you grow as well, but it'll also help you implement tools faster. How did your Air Force and investment background shape the way you lead a deep tech GTM focused startup? Um I think in general having different background than what you're doing now is very helpful. Uh you have different point of views. Obviously the air force uh gave uh me a lot of discipline and the way to think about complex processes that have a deadline. And in terms of investment it's also great it's always great to understand how the person in front of you thinks. So we are a VC backed business. to how VCs think. But in general, I'll say that starting a new business, a new startup is different. I

[26:33] think that starting each startup is incredibly different. And advices and those things, I don't see them as advice. It's just personal experience. What's the biggest GTM lesson you've learned about selling AI in production environments versus demo environments? I think in the Gen AI space, it's incredibly simple to build a demo that will look very, very, very, very good, very fast. And we see it with tools that are I'm not talking about bots. I'm talking about every kind of tool out there, right? Building an app now with Lovable or Vzero is incredibly fast. But obviously, once you go into to to the the small bits and bites, you see that it's very very different from production environments. And in customer support, it's no different. So, we might talk to an exec or uh we might say, "Hey, yeah, you know, why do I need to pay you guys?

[27:24] We built this tool internally in a week and we've already reached 12% automation." What they don't understand is that they're going to hit the wall very, very fast. And once you see a small amount of success and you hit that wall, that's incredibly dangerous because it'll take you a very long time to understand and change your mind that this doesn't work. So, I'd say not overpromising, knowing what you're getting into and making sure you understand and your prospect understand what's the difference between the demo that you're showing or what they're doing internally to how things actually look like in production environments. Want to move on to the rapid fire section now. First business you ever started? Notch. One GTM channel you double down on right now with 10K an additional implementation manager for a specific company. Biggest AI tool you personally can't live without besides Notch. Ah, that's simple. Chad GPT.

[28:19] One overhyped AI trend founders should ignore. Aist. One book, podcast, or resource that's influenced your GTM thinking? Crossing the chasm. If Notch had a mascot, what should it be? That would be my co-founder's dog, Shams. In one sentence, the future of customer experience is hybrid. Raphael, this has been a fascinating look at how resolution first AI is changing the game, moving past reply only bots, resolving millions of tickets end to end and transforming CX into a lean growth driving function. You can learn more about Notch and book a demo at Notch. If you enjoyed this episode, subscribe to GTM Vault on Spotify, Apple Podcast, or YouTube, and share it with someone in your network who's rethinking the future of customer support. Until next time, I'm Rick Ketta. Thanks for listening to GTM Ba.

[29:12] Tech founders and VCs, careers, lessons. GTM