Parahelp: The End-to-End AI Support Agent

1. 0:00 – 0:46

   Intro: What Parahelp Is Building

   1. SPSpeaker0:00

      [upbeat music] Really excited to be here today with the founders of Parahelp, Anker and Mads. Parahelp is one of the top agentic AI companies to come out of YC in recent years, and they've always been at the cutting edge of building really sophisticated agentic AI systems. So, uh, today, we're gonna talk about Parahelp. Uh, they just recently raised a, a great Series A. We're gonna talk about that. Um, and we're also gonna talk a lot about the technology behind the scenes and some of the cutting-edge stuff that they've just released in the last few months that I think is going to really change the game for how the top agentic AI systems actually work. We're also gonna talk a bit about the crazy

2. 0:46 – 3:12

   The First Version (and Why It Didn’t Work)

   1. SPSpeaker0:46

      origin story behind this company and a two-year pivot story that began in Denmark and ended in this, um, awesome Series A that they just raised. So thank you guys for joining us today. Thank you, Joe.

   2. ARAnker Ryhl0:58

      Thank you.

   3. SPSpeaker0:59

      Parahelp has an incredible origin story. Do you guys wanna take us back to Denmark and how this thing all happened?

   4. ARAnker Ryhl1:07

      So both Mads and I, uh, grew up in, in Denmark. Uh, and we often say we're from Copenhagen, uh, because that's what people know, [chuckles] but we're not really. Uh-

   5. SPSpeaker1:16

      [chuckles]

   6. ARAnker Ryhl1:16

      ... both close to Copenhagen. Everything is close to Copenhagen if you're from Denmark.

   7. SPSpeaker1:19

      [chuckles]

   8. ARAnker Ryhl1:20

      Uh, I grew up, uh, in a village of 200 people. Uh, and that also meant, uh, that I didn't really have anything else to do than be outside or code. Uh, so I began coding very early on. Uh, then as the classic consumer arc, I didn't really know what B2B SaaS was, [chuckles] uh-

   9. SPSpeaker1:37

      [chuckles]

   10. ARAnker Ryhl1:37

       ... and how to actually build anything of value there. Uh, so when I went, uh, to high school, I met Mads. Uh, we didn't go to the same high school. Uh, Mads started already in middle school, uh, by building his own fashion company, uh, and designing there, uh, and then doing marketing as well. Uh, so we met through a fellow friend, uh, co-founding, uh, an organization for young entrepreneurs in Denmark-

   11. SPSpeaker1:57

       Mm-hmm

   12. ARAnker Ryhl1:57

       ... uh, to get all the high school students that were interested in this together. Startups are not really a thing, uh, in Denmark.

   13. SPSpeaker2:02

       [chuckles]

   14. ARAnker Ryhl2:03

       I think, uh, one of the fun things we often tell people is that it's actually not even a word, uh, startup in Denmark.

   15. SPSpeaker2:08

       [chuckles]

   16. ARAnker Ryhl2:08

       Uh, so people just say, "How's-"

   17. SPSpeaker2:09

       In Danish, really.

   18. ARAnker Ryhl2:10

       In Danish.

   19. SPSpeaker2:10

       There's, there's, there's no Danish word for startup.

   20. ARAnker Ryhl2:12

       No.

   21. SPSpeaker2:12

       [chuckles]

   22. ARAnker Ryhl2:12

       Uh, people just say-

   23. SPSpeaker2:13

       Well, that's telling. [chuckles]

   24. ARAnker Ryhl2:14

       ... "How's, how, how's your project going?"

   25. SPSpeaker2:15

       [chuckles]

   26. ARAnker Ryhl2:16

       But both Mads and I, after we found, uh, startups, I found coding, um, and Mads designing and marketing, we had a hard time concentrating in school. Uh, so we co-founded this club, uh, then instantly started to work together. I realized, uh, that I was not very great at designing. Uh, Mads was, and we need someone there. Uh, so Mads became really great at Figma, uh, and product. Uh, and I then did the coding. Uh, and then we built a lot of consumer apps during high school, uh, where I think a lot of founders-

   27. SPSpeaker2:42

       Mm-hmm

   28. ARAnker Ryhl2:42

       ... uh, that's also where, uh, our love of building the best product, uh, was born. Uh-

   29. SPSpeaker2:46

       And these were projects. This was before you started the, like, startup that became Parahelp.

   30. ARAnker Ryhl2:50

       Exactly.
3. 3:12 – 6:05

   YC, Cold Emailing 700 Prospects, and Finding the Problem

   1. SPSpeaker3:12

      Yeah. [chuckles]

   2. ARAnker Ryhl3:12

      Uh, but that all of the-

   3. SPSpeaker3:13

      You got it out of your system early.

   4. ARAnker Ryhl3:14

      E- exactly. All, all of the nearby, uh, high school used. Uh, and then in, like, the last, uh, year, uh, of high school, I'd also started to code Solidity already in middle school, uh, so we're in the space. Uh, then NFT started to actually just explode in usage, and all of our friends were asking, "How do I buy this?"

   5. SPSpeaker3:29

      Mm-hmm.

   6. ARAnker Ryhl3:29

      "And this seems way too technical." Uh, and we didn't really trade a lot or were very into that, but we thought there was an interesting opportunity to make it easy for our friends who were not technical at all, uh, and a super interesting product challenge. So where we raised our first round and got into a real startup, uh, was right after I graduated, uh, and Mads had one year left of high school. Uh, we moved into a very small office above a McDonald's, and, uh, I actually slept there [chuckles] for a while, uh, in Copenhagen. Uh, then we built this investment app. Uh, we were the first ones to partner with Stripe in Europe after many attempts-

   7. SPSpeaker4:00

      Wow

   8. ARAnker Ryhl4:00

      ... uh, of convincing them how our system worked, uh, so you could actually, with Apple Pay, buy an NFT, and we then took care of all the technicalities. Uh, then we also learned the hard way when we finally launched and got through the regulation, uh, that the NFT market, uh, had gone from a few million users, uh, to a few thousand. The users that were left, we had not really talked with users, so as soon as we actually started talking with them, we realized, likewise it says, that these are not really users that care about anything else than the market going up or down, uh, and the product that we are very proud of. Uh, but that didn't really matter [chuckles] if no one-

   9. SPSpeaker4:34

      Yeah

   10. ARAnker Ryhl4:34

       ... wanted to use it-

   11. SPSpeaker4:35

       Yeah

   12. ARAnker Ryhl4:36

       ... uh, unless the market was up. We couldn't control that.

   13. SPSpeaker4:38

       Yeah.

   14. ARAnker Ryhl4:38

       Uh, so we went into full pivot mode. Um, and then actually inspired, uh, by a- another YC company called Slope, uh, where we read, uh, how they had pivoted. We did this where we would sprint for three weeks, uh, on an idea, uh, to actually, like, be courageous enough in trying an idea that seems stupid perhaps-

   15. SPSpeaker4:55

       Mm-hmm

   16. ARAnker Ryhl4:55

       ... because it was only three weeks-

   17. SPSpeaker4:57

       Mm.

   18. ARAnker Ryhl4:57

       ... so the cost song fallacy wasn't that great-

   19. SPSpeaker4:59

       Mm

   20. ARAnker Ryhl4:59

       ... if it didn't work. And then we would write a postmortem, uh, if we didn't pursue it, uh, to force us to actually try and build and sell something. Uh, that's also where we traveled to San Francisco for the first time and visited some hacker houses and fell in love with it. Um, so in, uh, January of 2024, uh-

   21. SPSpeaker5:14

       And you guys were how old?

   22. ARAnker Ryhl5:15

       I was 21, uh, at the time, and you were 19.

   23. SPSpeaker5:20

       18, I think.

   24. ARAnker Ryhl5:21

       I think, yeah, 18.

   25. SPSpeaker5:22

       18.

   26. ARAnker Ryhl5:22

       20.

   27. SPSpeaker5:22

       Okay.

   28. ARAnker Ryhl5:22

       Yeah.

   29. SPSpeaker5:23

       Okay.

   30. ARAnker Ryhl5:23

       Yeah. So when we got into, uh, YC, I just turned 22. We traveled to San Francisco. We pivoted, uh, a lot, and in January of 2024, uh, so, um, almost two years ago, uh, we then moved into a hacker house in Berkeley, uh, where, uh, we also, uh, met the Same.dev guys that convinced us [chuckles] to apply for Y Combinator. And there we began exploring the B2B space, uh, and had this co-pilot, which we got into with originally. Uh, and then already at the retreat, uh, I remember we, we, [chuckles] uh, discussed with you, uh, and could feel that we had customers, and they were happy, but they were not really happy. And sales cycles were long, uh, because-
4. 6:05 – 8:40

   The Mid-Batch Pivot That Changed Everything

   1. ARAnker Ryhl6:05

      disparate pain. Uh, and then we had Sonnet 3.5 launching.

   2. SPSpeaker6:08

      Mm-hmm.

   3. ARAnker Ryhl6:08

      Uh, and suddenlyEnabling, and this was at least our thesis, [laughs] uh, semi-reliable agentic change if we sampled. Uh, and this is where we then did, uh, a few weeks later, I think, a classic mid-batch pivot of deciding to go all in on end-to-end resolution, uh, of support tickets, and then Mads sent, in just one week, 700, uh, personalized cold emails. [laughs]

   4. SPSpeaker6:29

      [laughs]

   5. ARAnker Ryhl6:30

      Uh, I think, and we basically kept that up. Uh, all of our, uh, initial customers, including Captions and Perplexity, were called, uh, since we didn't really have a lot of connections, uh, at all, if any, [laughs] in, uh, Silicon Valley. And we sent, therefore, emails to all of the companies we thought, uh, were cool, uh, and then pretty quickly realized, uh, that for especially software startups, which were all the companies, uh, we, uh, we admired us of, we could build a much better product, uh, if we focused on them, and therefore focused on building deep integrations to their tools, uh, and focused on building the best solution to the knowledge management problem they all had, uh, which really no other verticals, uh, in the support space had.

   6. SPSpeaker7:06

      And so the, the current idea for Parahelp, the thing you guys are working on now, that was a mid-batch pivot during YC after, like, a two-year journey that began with, like, Robinhood for NFTs in Denmark, and somehow-

   7. ARAnker Ryhl7:19

      That's right

   8. SPSpeaker7:19

      ... turned into, like, agentic AI for customer support teams.

   9. ARAnker Ryhl7:22

      So a, a lot of pivots.

   10. SPSpeaker7:24

       [laughs]

   11. ARAnker Ryhl7:25

       Uh, and, uh, I actually... Like, Mads and I have always, uh, and, and sometimes we thought too easily, uh, been great at pivoting, because as soon as we thought that this idea didn't really work and we had the gut feeling, we just couldn't work on something we didn't thought mattered a ton. That's also why we in some classes in high school had a hard time concentrating [laughs] if we knew the subject.

   12. SPSpeaker7:44

       Oh.

   13. ARAnker Ryhl7:44

       Uh, wanted to build on our startup, which we thought, uh, was way more, uh, important. So we pivoted a lot, uh, until then, uh, finding Parahelp, uh, and suddenly, uh, feeling, uh, the pull, [laughs] uh, way more than we felt, uh, ever before. Uh, and since then, uh, it's been practicing continuing obviously to stay alive, but as Dalton says, getting to the next level.

   14. SPSpeaker8:04

       Why don't you just tell us what is Parahelp?

   15. MLMads Liechti8:06

       So Parahelp is an AI support agent, uh, for customer service, uh, that resolves, uh, support tickets end to end. And that means using tools such as Stripe, uh, or similar, uh, to perform relevant actions and follow complex instructions.

   16. SPSpeaker8:21

       And who's using Parahelp now? Who are your customers?

   17. MLMads Liechti8:24

       Parahelp was built specifically for fast-moving software companies. Um, we quickly realized, uh, during our beginning of Parahelp, uh, that if we built a solution solely for them, uh, that would mean that we could build a 10X solution focusing on tho- uh, their problems.

5. 8:40 – 11:28

   Landing Perplexity as Customer #2

   1. MLMads Liechti8:40

      And, uh, today, Parahelp is used, uh, by companies like Perplexity, uh, Heygen, uh, Framer, ElevenLabs, uh, Replit, Vodoroom, many of the fastest-growing AI companies, and also top AI.

   2. SPSpeaker8:54

      So the thing about Parahelp is it doesn't just respond to tickets. It actually takes action to, like, solve the problem that the customer is reporting in the support ticket. Can you give an example of, like, an action that, like, you would take on behalf of a customer?

   3. MLMads Liechti9:07

      One example would be a refund request. Uh, so a customer comes in, and they request a refund, and, uh, the company then has a set of criteria that makes that customer either eligible or not eligible for the refund. Uh, and then what Parahelp does is that Parahelp, first of all, understands those criteria, uh, so we are able to follow complex instructions, uh, when, uh, handling such a ticket. And then Parahelp is also integrated into the billing system that they use. And so that could be Stripe, uh, for example-

   4. SPSpeaker9:41

      Okay

   5. MLMads Liechti9:41

      ... where if the customer is then eligible for the refund, Parahelp can go ahead and then process that action in Stripe.

   6. SPSpeaker9:48

      Parahelp is actually directly connected to the company's bank account, and the company has effectively authorized it to move money on the company's behalf without any human supervision. Is there, like, a human that has to authorize it, or, like, they trust Parahelp to just move money around by... because they know that it's gonna follow the, the policy?

   7. ARAnker Ryhl10:05

      It's actually, uh, both. [laughs] Uh, we've even been the ones often pushing, uh, for human authorization-

   8. SPSpeaker10:13

      Mm-hmm

   9. ARAnker Ryhl10:13

      ... for certain things, because we've had customers, uh, being very quickly to trust, uh, because we are, uh, extremely, uh, reliable. Uh, but we have a Slack approval flow for sensitive actions, uh, where the agent, the Parahelp agent, basically goes and says, "Hey, I've just requested a teammate to look at this. I'll get back to you." Uh-

   10. SPSpeaker10:30

       Mm-hmm

   11. ARAnker Ryhl10:31

       ... but then because we build the agent to function not like a workflow, but-

   12. SPSpeaker10:34

       Mm-hmm

   13. ARAnker Ryhl10:34

       ... an actual human agent-

   14. SPSpeaker10:35

       Mm-hmm

   15. ARAnker Ryhl10:35

       ... we just know that there's an ongoing, uh, request that's pending approval, so we actually often continue to chat with the customer-

   16. SPSpeaker10:42

       Ah

   17. ARAnker Ryhl10:42

       ... and answer all the questions. When, uh, one of the billing managers, uh, approves, uh, a Slack action, they just see a summary.

   18. SPSpeaker10:49

       Mm-hmm.

   19. ARAnker Ryhl10:49

       Then we go and execute, uh, the necessary actions and follow up on the customer. Uh, but there's also, uh, numerous billing actions for certain customers where we've had 99% success rate over a given time.

   20. SPSpeaker11:00

       [laughs]

   21. ARAnker Ryhl11:00

       When we first, uh, got into YC last summer, uh, we were building a co-pilot, uh, and then talked, uh, with you, uh, about how we're kind of trying to help every ticket for a customer, uh, which meant that we didn't really have a lot of success, uh, because then there were a ton of tickets that were too complex. Uh, and at the same time, we're very fortunate timing-wise, uh, because we saw Sonnet 3.5 come out, uh, and if we just sampled a lot of Sonnet 3.5 tool calls, uh, and basically simulated

6. 11:28 – 14:20

   What “End-to-End Resolution” Actually Means

   1. ARAnker Ryhl11:28

      multiple agentic actions at once every time we got a ticket, uh, and then used different heuristics or LLM as a judge to pick the best path, uh, then we could actually call tools reliably.

   2. SPSpeaker11:38

      Yeah.

   3. ARAnker Ryhl11:38

      Uh, and suddenly we got so down tau bench, and we could actually get to 50 or 60%. Uh, this was again last year. Now we are, uh, way higher for certain customers, uh, versus 20% [laughs] from using simple, uh, RAG. And there we built a, a tool framework internally we still use, uh, where every tool, uh, this also works with MCPs, uh, but have a validate function, uh, that can run deterministic code checks, uh, if needed.

   4. SPSpeaker12:03

      I want to go back to the list of customers that is, that are using Parahelp. Can you say again some of the really big customers that you guys have?

   5. MLMads Liechti12:11

      So a company like Perplexity, uh, happens to be our second customer, uh, that we onboarded during-

   6. SPSpeaker12:16

      Perplexity was your second customer ever?

   7. MLMads Liechti12:19

      Yes. Uh, during, uh, the Y Combinator batch.

   8. SPSpeaker12:21

      Okay.

   9. MLMads Liechti12:22

      Uh, and, uh, what we initially saw and, and what we initially just wanted to test out with Parahelp was that we could resolve these complex tickets. Uh, and that was the initial premise of Parahelp, uh, that we can do that, and that resonated a lot

   10. ARAnker Ryhl12:35

       With, uh, Perplexity. Uh, they were growing extremely fast, uh, and had a lot of tickets, uh, that, uh, they wanted to, to resolve, uh, to the highest quality possible. So that was what we, we found working with Perplexity, and then we just decided to double down. Uh, and since then, we have onboarded companies such as, uh, Replit, uh, also a YC company, uh, PhotoRoom, uh, Framer, Heygen, um, and ElevenLabs, many of the, the, the fastest-moving software companies.

   11. SPSpeaker13:05

       And not just the fastest-moving software companies, but one of the things I think is interesting about that customer list is almost all of them are AI-native companies. Like, these are companies that were started recently that are themselves building agentic systems using LLMs. Why are they using Parahelp rather than building this themselves?

   12. ARAnker Ryhl13:20

       Most of our customers find either by building it or trying to enable the default AI they have in their ticket system.

   13. SPSpeaker13:26

       Mm-hmm.

   14. ARAnker Ryhl13:27

       Intercom, uh, has a default out-of-the-box AI.

   15. SPSpeaker13:29

       Yeah.

   16. ARAnker Ryhl13:29

       Zendesk, Front, uh-

   17. SPSpeaker13:30

       Yeah

   18. ARAnker Ryhl13:30

       ... et cetera. But what happens, they are just like a custom-built system, is that it works really well in the beginning, uh, for the subset of tickets where you don't need tools. And then what happens is that they're not built for knowledge management for fast-growing software companies- [laughs]

   19. SPSpeaker13:43

       Yeah

   20. ARAnker Ryhl13:43

       ... because they are not by default connected, and it takes a lot of work, uh, to all the context you have in Linear, in Notion, in Retool, and in Slack, uh, and that's changing every day. This means, uh, that it starts out really well, and then because nobody really maintains it-

   21. SPSpeaker13:57

       Mm

   22. ARAnker Ryhl13:57

       ... [laughs] nobody has built a great system around that-

   23. SPSpeaker13:59

       Mm

   24. ARAnker Ryhl13:59

       ... then it actually diminishing, uh, diminishes, uh, in return, a- and your resolution rate drops. And even worse, in some cases, customers actually get a, a really bad customer experience, uh, because it's out of date, uh-

   25. SPSpeaker14:10

       Mm

   26. ARAnker Ryhl14:10

       ... and there's not any built-in review systems.

   27. SPSpeaker14:12

       Mm.

   28. ARAnker Ryhl14:13

       Uh, which is what we've just done, uh, with our research agent that can do real-time QA, uh, on how our AI agent, Parahelp

7. 14:20 – 18:05

   How the Agent Takes Real Actions (Refunds, Fixes, Workflows)

   1. ARAnker Ryhl14:20

      agent, is performing. And then the other thing is that tools may sound, uh, pretty trivial to set up, and a lot of other AI, uh, support agent solutions also have at least, like, on paper, tool support. Uh, but then, uh, as you dive in, uh, you find out that tools are way more than just calling the API endpoint. It's having the proper guardrails, and it's also having the proper evaluations set up to make sure that when you have access, uh, to a Stripe tool, a- and therefore a, a really sensitive, uh, tool that's also connected to a customer's bank account-

   2. SPSpeaker14:51

      Mm-hmm

   3. ARAnker Ryhl14:52

      ... uh, then it should actually be called correctly, and it should always be called, uh, correctly. And having real-time monitor systems, uh, for that, uh, and an evaluation system, which is what we've spent probably engineering-wise the most time on building internally and just building more and more automated. Uh, but how do we run evaluations, uh, and how do we run them at all times to make sure we're not regressing? Uh, and that our agent, like today, we're handling thousands, uh, of tickets, uh, and queries daily. Uh-

   4. SPSpeaker15:20

      Per day?

   5. ARAnker Ryhl15:21

      Yeah.

   6. SPSpeaker15:22

      There's thousands of customer support tickets being closed every day by Parahelp's agents.

   7. ARAnker Ryhl15:25

      Exactly, and we're involved in, uh, et cetera, uh, which means that we can't monitor this by hand at all. [laughs]

   8. SPSpeaker15:32

      Yeah.

   9. ARAnker Ryhl15:32

      In the beginning, uh, we onboarded, uh, Captions AI as the first company, and then Perplexity, uh, AI a, a week later, uh, all of this within four weeks, uh, of pivoting. This also meant, uh, that we were unsure about a lot of things.

   10. SPSpeaker15:45

       [laughs]

   11. ARAnker Ryhl15:45

       Uh, I... And so we observed everything by hand, uh, and was, like, figuring out how do we run proper evaluations, uh, so we could actually sit and make sure the agent was steered in the correct way. Uh, and obviously we can't do that anymore. Uh, so having these automated systems built, uh, I would say is key. And what most customers that tries to build internally, even when they're great at AI-

   12. SPSpeaker16:06

       Yeah

   13. ARAnker Ryhl16:07

       ... uh, figure out that it's actually much more than just setting up a one-time, uh-

   14. SPSpeaker16:11

       Right, right, right

   15. ARAnker Ryhl16:11

       ... ChatGPT app-

   16. SPSpeaker16:12

       Yeah

   17. ARAnker Ryhl16:12

       ... or endpoint to our docs.

   18. SPSpeaker16:14

       Mm-hmm.

   19. ARAnker Ryhl16:14

       Uh, because making sure that evaluations continue to pass, we're not regressing, and staying in touch with both product and engineering and marketing and sales at all times.

   20. SPSpeaker16:25

       Yeah, an interesting thing about you guys is your customers are B2B companies. Their customers are... This is not like the customers of, like, DoorDash or something, or, or like Uber, where, like, the customers are just paying them, like, a few dollars per, per transaction. Like, their customers are very important, and their customer support tickets are very important, and so they really care about the quality of the resolutions to those tickets.

   21. ARAnker Ryhl16:47

       Exactly. Uh, and I think one of the milestones or metrics, uh, we're most proud of is that some of our companies have entrusted us with handling enterprise customers up to 1,000 seats. [laughs]

   22. SPSpeaker17:00

       Wow.

   23. ARAnker Ryhl17:00

       Uh-

   24. SPSpeaker17:00

       Which is, like, prob- probably a six-figure customer for them.

   25. ARAnker Ryhl17:04

       Exactly.

   26. SPSpeaker17:04

       And they're entrusting your agent to talk to their customer.

   27. ARAnker Ryhl17:08

       And obviously we didn't start there. It wasn't an accident. That was, like, going up from 20 members-

   28. SPSpeaker17:12

       Mm, mm

   29. ARAnker Ryhl17:13

       ... and then 50, [laughs] and then 100, uh, and then seeing, uh, that we continued, uh, to perform.

   30. SPSpeaker17:17

       Just, just kind of following the progression of a human customer support agent, right? Which, like, probably starts with the small accounts and then graduates to, like-
8. 18:05 – 22:18

   Building the Agentic Architecture

   1. ARAnker Ryhl18:05

      the model in Frontier Labs, uh, were doubling down on memory as a concept.

   2. SPSpeaker18:09

      Mm-hmm.

   3. ARAnker Ryhl18:09

      Uh, so now an average, uh, Parahelp agent for a customer has 40 memory files, uh, I would say define-

   4. SPSpeaker18:16

      Mm

   5. ARAnker Ryhl18:16

      ... that Parahelp is-

   6. SPSpeaker18:17

      Mm

   7. ARAnker Ryhl18:17

      ... at all times figuring out how to improve proactively, uh, by running, uh, monitoring and evaluations, uh, as tickets come in. And every time the customers add a new feature or we see it automatically because it happens in Linear-Or in the help center, or someone texts us in Slack, then the Parahelp agent goes, just like a coding agent, uh, and thus file edits per memory file, uh, to make sure there's no duplicates. So the core agent is actually quite simple because it has all of these, uh, very general tools, uh, that can then, through complicated policies, uh, be used. Uh, so one agent can respond in fifteen seconds to a trivial question, uh, for a customer, and then we can actually have an agent working for three minutes, uh, because it has access to fifteen custom tools, and it can go and look up a request ID in Datadog, uh, to help debug with the enterprise customer. Uh, and it can go and look at SSO policies, uh, and at both internal and Stripe data and combine it, and then reason for a long time, analyzing screenshots, uh, or similar, to then get back to the customer, uh, with an answer that actually fits their way more complex question than just, "What's your pricing?"

   8. SPSpeaker19:25

      Recently, you guys released your next generation agentic architecture for Parahelp. Can you explain how it works?

   9. ARAnker Ryhl19:31

      Today, uh, we have two agents, uh, that we work on, uh, and build with our customers. Uh, we have the Parahelp agent, which works directly in the customer's ticket system. This enables customers to set up Parahelp extremely quickly, uh, because all you actually need to do is invite us as an agent to your ticket system. So you just send an email invite to the Parahelp agent, uh, and then we're live, uh, without engineers needing to install any SDKs, uh, or connect any channels. And then we have the Parahelp Assistant, uh, which is what we just launched a major upgrade, uh, to in terms of architecture, uh, which is what enables our customers, uh, to build and optimize policies, uh, set up tools, uh, and then test everything thoroughly, uh, before publishing to production. And the Parahelp Assistant is what has changed most, uh, the past six months. Uh, so in the beginning, and a lot of our customers, fast-growing startups, chose us and still chooses us because we can solve complex tickets. Uh, this meant that we were running, uh, a lot of complicated evals end-to-end, uh, when onboarding a customer. And because models were not intelligent enough, we had to do this manually in a lot of cases. We would run this, uh, via internal tools to actually make sure that the agent were optimized. Uh, and if anything needed to be fixed, we would often have initially, uh, to go and actually prompt optimize this forward-deployed engineering approach. This worked well, uh, but it also meant, uh, that we were finding ourselves writing a lot of custom prompt optimizations that couldn't be reused for multiple customers, and we're therefore taking a lot of time as we grew a lot and had a lot of demand, uh, to actually onboard companies. Uh, and suddenly we're falling behind in being the fastest solution, uh, for setup, uh, and a high resolution rate simply because we couldn't follow demand without hiring a lot of people, uh, that would do custom work per customer, which, uh, as someone that's, uh, very product-minded-

   10. SPSpeaker21:22

       Yeah

   11. ARAnker Ryhl21:22

       ... uh, we didn't really like.

   12. SPSpeaker21:24

       Yeah.

   13. ARAnker Ryhl21:24

       And then Opus four actually came out.

   14. SPSpeaker21:26

       I remember you guys came to me, and you were like, "Our biggest bottleneck is we can't hire enough engineers to be forward-deployed engineers to be able to onboard all these customers who want to use Parahelp."

   15. ARAnker Ryhl21:35

       Exactly. Uh, and then we tried, uh, to hire a lot. We're still obviously hiring, but we actually-- we also found, uh, that we were probably not the best founders at building a super execution and ops.

   16. SPSpeaker21:48

       Yeah.

   17. ARAnker Ryhl21:48

       We were the best founders at building a product first-

   18. SPSpeaker21:50

       Yeah

   19. ARAnker Ryhl21:50

       ... company.

   20. SPSpeaker21:50

       Yeah.

   21. ARAnker Ryhl21:50

       Uh, so the engineers we wanted, uh, to hire were also the best at building product-

   22. SPSpeaker21:54

       Mm

   23. ARAnker Ryhl21:54

       ... uh, not building a lot of manual work and being more consultant.

   24. SPSpeaker21:57

       Yeah.

   25. ARAnker Ryhl21:57

       Uh, so we wanted to keep a small team and still grow extremely fast. Uh, and therefore, we actually, over the summer, uh, had two months where we paused growth and rebuilt our agentic, uh, architecture, uh, because Opus four launched. Uh, and this model, for the first time, had enough intelligence on our internal benchmarks to actually be able to generate

9. 22:18 – 29:04

   Scaling to Thousands of Tickets Closed Every Day

   1. ARAnker Ryhl22:18

      eval sets on the fly.

   2. SPSpeaker22:20

      Wow.

   3. ARAnker Ryhl22:20

      Uh-

   4. SPSpeaker22:21

      So it's effectively like an AI forward-deployed engineer.

   5. ARAnker Ryhl22:24

      Exactly.

   6. SPSpeaker22:25

      Yeah.

   7. ARAnker Ryhl22:25

      Which meant, uh, that suddenly we didn't have to do-

   8. SPSpeaker22:28

      Yeah

   9. ARAnker Ryhl22:28

      ... custom forward-deployed engineering.

   10. SPSpeaker22:30

       Forward-deployed engineering. Yeah.

   11. ARAnker Ryhl22:30

       Uh, so we're still big believers in forward-deployed engineering as let's figure out what works best for a customer and then take it back to the product-

   12. SPSpeaker22:38

       Yeah

   13. ARAnker Ryhl22:38

       ... uh, and build something generalizable. Uh, but since launching this, we haven't written a single custom evaluation set because our agent does it-

   14. SPSpeaker22:44

       Wow

   15. ARAnker Ryhl22:44

       ... for our customers.

   16. SPSpeaker22:45

       Does it for you. Wow. I wonder if this is the next generation of the forward-deployed engineering trend that has overtaken Silicon Valley in the last two years. The knock on forward-deployed engineering has always been that it doesn't scale, and, like, this is the answer. Like-

   17. ARAnker Ryhl23:00

       Yeah.

   18. SPSpeaker23:00

       It scales if they're all agents.

   19. ARAnker Ryhl23:02

       Exactly, and I, and I think, uh, it's, it's definitely what we believe as in-- as the next generation of vertical agents. Uh, we took a lot of what we saw and thought worked really well with code agents, uh, which are definitely implemented far more wide than other vertical agents, uh, which was, uh, in our, uh, eyes, being able to run deterministic code and then the agent to optimize, so it could actually work for much longer.

   20. SPSpeaker23:26

       Mm-hmm.

   21. ARAnker Ryhl23:26

       Our agent can easily work for thirty minutes-

   22. SPSpeaker23:29

       Mm

   23. ARAnker Ryhl23:29

       ... today.

   24. SPSpeaker23:30

       Mm-hmm.

   25. ARAnker Ryhl23:30

       Uh, or forty minutes optimizing its own policies and running tests, and then having a review system. So those two things, uh, we were extremely inspired by coding agents. Uh, and then we thought that worked well. And with Opus four, we could actually build something, uh, for something that was a bit more complex problem, uh, because our evaluation framework is quite complex in order to allow us to actually test complex, uh, scenarios. But Opus four could manage, so to say, all of that context and also reason about how to improve it. And then the other, uh, big change we launched was this research agent. Uh, so we can actually search with a grip-like tool, uh, that does agentic reg, uh, and also searches in semantics. Uh, we can search up to twenty thousand tickets at a time, and then we look and analyze closely with a lot of sub-agents, uh, to manage context, uh, up to five hundred, uh, tickets, uh, typically, uh, fifty to five hundred, where we actually open the full ticket, uh, obviously with all sensitive information, uh, removed. And then we look at exactly what was handled, when did the AI agent, uh, transfer, or how did the human handle it, depending on what we want to analyze. And then we provide a deep research-like report.Which sources, uh, and all the tickets, uh, we analyzed, and then we suggest improvements proactively, uh, that can be implemented straight away by the support manager by just prompting, uh, the agent in the same session. "Please, I've just added this tool. Uh, please have a policy that now calls this to solve this knowledge gap."

   26. JFJared Friedman24:58

       Mm-hmm.

   27. ARAnker Ryhl24:59

       Uh, or our, uh, research agent actually being able to see what human agents replied in groups and then, uh, automatically saying, "It just looks like we need to have this policy," which the human agent does, and then I actually think I can solve 5% more of these tickets.

   28. JFJared Friedman25:13

       An interesting kind of continuous loop that all this enables together, not only the, the research agent, but also the configure and testing agent, is that you can find these patterns in how humans resolve tickets, and then you can also configure them and test them and publish them and have this continuous loop that at any given time, uh, optimizes the amount of tickets that the agent is capable of handling. Uh, so it all kind of ties together, and that's the interesting fly-flywheel.

   29. ARAnker Ryhl25:43

       So one of the stats we've seen since launching this, uh, for Perplexity, uh, as an example, just two weeks ago, and having this flywheel automatically set up that will ping them in Slack and then prompt them to give some information, and then the agent testing it itself and optimizing, is that we've seen their end-to-end resolution rate, uh, go up about 12 to 15 percentage points, uh, depending on the day-

   30. JFJared Friedman26:07

       Wow
10. 29:04 – 34:40

    Onboarding Customers in Hours, Not Weeks

    1. ARAnker Ryhl29:04

       700 tickets at once and then to be able to handle. We have a lot of spikes [laughs] in our system.

    2. JFJared Friedman29:08

       Mm-hmm. Mm-hmm.

    3. ARAnker Ryhl29:09

       Um, but we use Hatchet, actually, uh, a Y Combinator-

    4. JFJared Friedman29:11

       Okay, cool

    5. ARAnker Ryhl29:12

       ... uh, company we just switched to, uh, which we're very happy, uh, with so far, uh, and, and great at handling this. Uh, but then we follow up on everything automatically. Uh, and what's quite cool, uh, and exciting about this follow-up, uh, feature combined with research, uh, and having context from all of these systems, uh, is that we can actually start helping our customers with also being a customer success manager, uh, and upselling, uh, customers or going back to churned customers and tell them about new features. Uh, because every time our customers launch a new feature or solve a bug, it can just say, or the research agent prompts them automatically, "Should we follow up on every customer, uh, or reach out to every customer," [laughs]

    6. JFJared Friedman29:50

       Mm.

    7. ARAnker Ryhl29:51

       "Uh, that ever requested this feature?"

    8. JFJared Friedman29:53

       Mm.

    9. ARAnker Ryhl29:53

       Uh, and then we can launch this agent and say, "Hey, this is now available, uh, on our Pro Plus plan. Uh, so if you upgrade, it does it automatically."

    10. JFJared Friedman29:58

        That, that's, that's a super intelligence version, right? Like, being able to go back and respond to everyone who ever asked about a feature over the last two years now that you've finally shipped it.

    11. ARAnker Ryhl30:06

        Exactly. Uh, and I think we're going to see more and more of this once support automation is actually at these, like, 85 to 95%. [laughs]

    12. JFJared Friedman30:14

        Mm-hmm.

    13. ARAnker Ryhl30:15

        Uh, and, like, the rest of those tickets you want, uh, a human to solve.

    14. JFJared Friedman30:19

        Uh-huh.

    15. ARAnker Ryhl30:19

        Then the next, uh, frontier, uh, to actually have the best system is how do we make support, make it as revenue-enabled, uh, as possible, uh, and how do we actually prevent churn? Uh, so we're also launching, uh, for our Intercom or Zendesk front, uh, agent a tool that can give a customer a discount. Uh, so instead of churning, uh, you can get a discount.

    16. JFJared Friedman30:39

        Mm-hmm.

    17. ARAnker Ryhl30:39

        Uh, or send the customer an upsell link.

    18. JFJared Friedman30:41

        Mm-hmm.

    19. ARAnker Ryhl30:41

        Uh, and then you can easily define the policies you want around this.

    20. JFJared Friedman30:44

        Mm-hmm.

    21. ARAnker Ryhl30:44

        And suddenly for certain customers we haveHundreds of thousands of tickets resolved already, uh, which means that the Parahelp agent has hundreds of thousands of emails of potential customers that could be upsold to business plans-

    22. SPSpeaker30:58

        Yeah

    23. ARAnker Ryhl30:59

        ... or expanded significantly more. Uh, but the average, uh, team doesn't have the resources to do this of the overview, uh, but suddenly, uh, you actually do because you have Parahelp.

    24. SPSpeaker31:09

        To your customers, how much of the value prop is, like, reducing costs spent on, like, hiring human customer support agents slash the fact that they just, like, can't hire enough good human customer support agents 'cause, like, they're hard to find, versus this stuff? Essentially, being able to do stuff with Parahelp that, like, would be impossible with humans, actually being able to increase revenue by being able to upsell customers in, like, a previously unscalable way.

    25. MLMads Liechti31:32

        It's much more the latter. Uh, and that also ties into us only focusing on fast-moving software companies.

    26. SPSpeaker31:39

        Okay.

    27. MLMads Liechti31:39

        Uh, so in many cases, they don't have a big team today. Uh, the, the cost of support is not necessarily super big, uh, but it has to be big if they have to keep up because they have to hire a lot of people. So it's more so implementing Parahelp, and then Parahelp being that hundred agents that they would otherwise, uh, hire.

    28. SPSpeaker32:01

        Mm-hmm.

    29. MLMads Liechti32:01

        Uh, and then as we also just talked about, having this united brain of all those agents-

    30. SPSpeaker32:07

        Yeah
11. 34:40 – 40:52

    Raising the $18M Series A

    1. MLMads Liechti34:40

       So it, it becomes this much more implemented tool throughout the organization because in many ways, uh, that's software company specific, but in many ways for software companies, support is the epicenter of customer knowledge, um, and product knowledge, uh, which is the starting point for a lot of work, uh, for the rest of the organization.

    2. SPSpeaker35:03

       Can we see a demo?

    3. ARAnker Ryhl35:04

       Definitely. Uh, and I actually have it pulled up right here. Uh, so what we're seeing, uh, is the Parahelp Assistant, which is what we just launched, uh, recently, uh, to be also in an app and not only in Slack. This is not what's responding to customers directly. This is the Parahelp Agent, uh, we call it, which lives in a customer's ticket system, and this is instead the Parahelp Assistant, which our customer, in terms of their support managers, use, uh, to actually create new policies, create new tools, uh, test, and also where deep research lives in terms of seeing the results, uh, of the Parahelp releases you're making.

    4. SPSpeaker35:39

       So the assistant is like the AI for a deployed engineer. It's replacing what a previously a Parahelp em- the job that previously a Parahelp employee would have had to do, talking to the customer and figuring how to update the Parahelp system.

    5. ARAnker Ryhl35:52

       Yeah.

    6. SPSpeaker35:52

       You've built a AI assistant that can just do that.

    7. MLMads Liechti35:54

       Mm-hmm.

    8. ARAnker Ryhl35:54

       And, and being what we believe is the future of how to manage an AI employee. Uh, instead of having workflows, uh, you have-

    9. SPSpeaker36:01

       Yeah

    10. ARAnker Ryhl36:01

        ... this much more cloud code-like experience.

    11. SPSpeaker36:03

        Mm-hmm.

    12. ARAnker Ryhl36:04

        Where, as you can see here, so this is a, a demo environment, uh, for Replit, uh, as an example. Uh, then we have an example, uh, of a... And we see this pretty common, uh, a support manager saying, "What happens today if a customer inquires about an enterprise plan, uh, on Replit?" And then, uh, our, and the Parahelp assistant actually launches, uh, and very much like a coding agent. It searches through everything it already knows, and then it goes, "There was actually, uh, no knowledge found." Uh, and it basically just tells, "Hey, this is, uh, what we do today. We don't have this custom, uh, policy, uh, set up yet." And this is, uh, an example, uh, of how a support manager, uh, would then-... direct, uh, the Parahelp agent to act differently. They give us a policy, uh, to follow, so if a customer is inquiring about enterprise pricing or features, uh, and wants to know more, uh, as an example, get a pricing quote, et cetera, uh, please direct them here. Uh, and then follow this policy, uh, in terms of what to ask for. And then we go, uh, and just like a code gen agent, we actually update... Uh, so this is all happening, uh, right now, it's released, uh, but pre-release, it's all happening in a sandbox environment.

    13. SPSpeaker37:15

        Mm.

    14. ARAnker Ryhl37:16

        And we go and update files.

    15. SPSpeaker37:18

        Mm.

    16. ARAnker Ryhl37:18

        Um, so this is, uh, a classic diff-

    17. SPSpeaker37:21

        Mm

    18. ARAnker Ryhl37:21

        ... that a, a customer, uh, would see. Uh, this is just to see that the agent is working. Uh, typically, they just review the entire output, uh, and actually let the agent, the assistant, uh, agent review, uh, and test. Then it goes, and as you can see, uh, this is a pretty great example because we actually update in many places, uh, where there was conflicting information-

    19. SPSpeaker37:41

        Right

    20. ARAnker Ryhl37:41

        ... uh, which is what we found other solutions can easily miss. Uh, or customers, uh, our customers can easily miss because they have to find it manually. Uh, so we go and update both subscription and billing, uh, FAQs and the policies we have there. We update the discounts, uh, policies because there was something conflicting, and we update product, technical, uh, and feature questions policy. Our Y assistant agent actually implements, uh, this in a draft mode and says-

    21. SPSpeaker38:07

        Mm

    22. ARAnker Ryhl38:07

        ... "I suggest that we test it." Uh, so what you see here is that we have this multi-agent system, uh, to manage context where we have four modes. Uh, and we started in ask mode, which is what do we know today? Then the customer said, "I want to do this," and it automatically switched to configure mode, which is-

    23. SPSpeaker38:23

        Oh

    24. ARAnker Ryhl38:23

        ... the edit mode.

    25. SPSpeaker38:24

        Mm.

    26. ARAnker Ryhl38:24

        And now it says, "I think we should test."

    27. SPSpeaker38:26

        Mm.

    28. ARAnker Ryhl38:26

        "Let's go to testing mode."

    29. SPSpeaker38:27

        Okay.

    30. ARAnker Ryhl38:27

        Uh, and the last mode we have is this deep research mode. Uh, so this is actually what enables, uh, customers with these, like, four sub-agents to switch between modes in one session.
12. 40:52 – 47:40

    Advice for Founders

    1. ARAnker Ryhl40:52

       get a full refund. Uh, or within 14 days you can get a prorated, uh, refund," then the agent obviously have to make sure, uh, that it can actually create context necessary to test that scenario. Another thing we've seen is that for all of our customers, uh, we know when there's an outage, uh, so we are connected to their status page. [laughs]

    2. SPSpeaker41:09

       Oh.

    3. ARAnker Ryhl41:09

       Uh, so if you want to test-

    4. SPSpeaker41:10

       Mm

    5. ARAnker Ryhl41:10

       ... outage handling-

    6. SPSpeaker41:11

       Mm

    7. ARAnker Ryhl41:12

       ... uh, you obviously have to be able to-

    8. SPSpeaker41:14

       To simulate an outage

    9. ARAnker Ryhl41:14

       ... and the agent to actually simulate that there is an active outage-

    10. SPSpeaker41:17

        Yeah

    11. ARAnker Ryhl41:17

        ... on the outage page.

    12. SPSpeaker41:18

        Yeah.

    13. ARAnker Ryhl41:18

        But there's also how do we respond when there's not an active outage, because then we obviously want to troubleshoot further.

    14. SPSpeaker41:23

        This is super cool, guys. I feel like I'm looking at the future of how all agentic AI companies will do customer onboarding. It's all gonna be through a system like this. And it's not just onboarding. It's the continual maintenance. To your point earlier, just deploying a system like this is just the beginning, then you have to keep it updated forever. If humans have to do that, it's just not ever gonna work. It requires too much human effort, and, like, the answer is to automate this as well with something just like the Parahelp assistant that you guys have built.

    15. ARAnker Ryhl41:50

        That's the bit, uh, [laughs] and what's, we've been extremely excited to see the response, uh, of so far.

    16. SPSpeaker41:56

        Yeah. You guys are really at the cutting edge here. Like, there's a small number of companies that are playing around with this idea, but it's definitely not a mainstream thing yet. You're one of the first companies, I think, to get something like this really working at scale in the real world.

    17. ARAnker Ryhl42:09

        We were pressured by the urgency [laughs] we saw-

    18. SPSpeaker42:12

        Yeah. [laughs]

    19. ARAnker Ryhl42:13

        ... in terms of growth is everything. Uh, and either we were doubling down on trying, uh, to hire a lot of forward-deployed engineers-

    20. SPSpeaker42:20

        Mm-hmm

    21. ARAnker Ryhl42:21

        ... uh, like many others. But our, I think, gut feeling was that-

    22. SPSpeaker42:24

        Mm-hmm

    23. ARAnker Ryhl42:24

        ... uh, we should build product that scales [laughs] -

    24. SPSpeaker42:26

        Yeah

    25. ARAnker Ryhl42:26

        ... and where we're not writing custom code.

    26. SPSpeaker42:27

        Yeah.

    27. ARAnker Ryhl42:28

        But we also want to grow, uh, so we need this to work today.

    28. SPSpeaker42:31

        Yeah.

    29. ARAnker Ryhl42:31

        Uh, this is what caused the Atrillion Sense Summer Sprint, uh, I would say, uh, of getting this out.

    30. SPSpeaker42:36

        Okay. Thanks so much for showing us the demo. I feel like I got a peek into the future there. Let's talk a little bit about the Series A. Um, who led the round and, uh, tell us a bit about it.

Episode duration: 47:40