Latest / State of the AI Union / The Data Advantage
Transcript
- 0:10Hello, everybody, and welcome to this week's episode of the State
- 0:13of the AI Union. I'm your host, Laura, and joined
- 0:17with me today is Thomas Hill. Thomas, would you like to
- 0:20introduce yourself? Sure.
- 0:21Yeah, happy to. Thanks for having me, Laura.
- 0:24I'm Thomas Hill and I'm a solutions architect for Dev Rev
- 0:27and I work on the pre sale side, meeting with customers.
- 0:31I'm setting up Pocs, getting to know their use cases, their pain
- 0:33points so that we can hopefully develop unique solutions to
- 0:37accommodate their challenges. So yeah, I'm happy to joined
- 0:40today. Thanks.
- 0:41Awesome. All right, Thomas, you know, one
- 0:43of the things you know, since you are in the field with
- 0:46customers, I'm sure one of the things that you've heard a lot
- 0:49is about is how they have built their own in house AI.
- 0:53And it's so easy right now to build, you know, AI that that
- 0:56does work actually and deploy it.
- 0:59I've heard the same thing from customers as well.
- 1:01Tell us more about what you've seen in terms of what people
- 1:05have built on their internal AI tools, What's been working and
- 1:08and why are they effective? Sure.
- 1:10Yeah. I mean, it's an interesting time
- 1:11for AI, right. And so it is possible.
- 1:13There's so many different use cases and that AI is creeping
- 1:17into, which is great. Like it let's face it, it's
- 1:20definitely helping improve our lives and gaining efficiency.
- 1:23And so I've seen first hand there's lots of examples, right
- 1:26there's copilot summarize calls for for coding.
- 1:29There is a Slack bot you know, back in the day there was move
- 1:32works. There's Jira bots, sales forces
- 1:35coming out with their own agentic work flows as well for
- 1:38various use cases, internal QA bots for, for HR or IT without
- 1:45having to pull a human in. So, so there's, there's really
- 1:48an unlimited amount of use cases where AI and agentic workflows
- 1:52are coming in. And so I've seen a lot of that
- 1:54with customers, how they build agents, what they want to see
- 1:57out of agents. And look, it's, it's not really
- 2:00that easy. Yes, you can do it and you can
- 2:04utilize existing tools. Of course, a lot of people are
- 2:07leveraging open AIAPI calls with their agents, but where you
- 2:11suffer is the context, right? You have limited context,
- 2:15especially from within your own company walls, right, Chad?
- 2:18GBT knows only what it can see on the web about your company,
- 2:22but it doesn't necessarily know all the internal workings, HR
- 2:24policy documents, support documents, cases, all these
- 2:28sorts of things. So companies are spending a lot
- 2:30of effort building their own home grown tools, which may or
- 2:34may not scale in the end. So yeah.
- 2:38Let's let's take an example of like the internal GPT, because
- 2:41I've heard that one, you know, across the board with a lot of
- 2:44customers that are doing that who typically like owns building
- 2:48something like that. Which department?
- 2:51Wow. It ranges a lot.
- 2:52So I met with a number of customers and it can be their
- 2:56development, it could be support, it could be a sales
- 3:00development team, right? That's writing tools for for
- 3:03salespeople and accounts people. But the interesting thing about
- 3:06open AI usage and not a lot of people or companies know about
- 3:10this is, you know, you're sending some of your critical
- 3:13data out of your walls, right to to these tools.
- 3:17And you don't know if, if you have a contract with open AI,
- 3:20where there are they keeping that data?
- 3:22You know, so there's, it ranges. There's a lot of there's a lot
- 3:25of teams within companies that are leveraging open AI.
- 3:27But again, I come back to that challenge of context, right?
- 3:30So you ask ChatGPT, let's say you start a new thread with
- 3:34ChatGPT. You start, you know, asking
- 3:36questions. It remembers within that
- 3:39context. But again, it won't remember
- 3:42from case to case, right? And it doesn't really know
- 3:45anything about any other data that you give it from within
- 3:49your walls. And companies have lots of data,
- 3:50right? There's Salesforce databases,
- 3:53there's Jira, there's Confluence, there's Slack, all
- 3:55these data sources. And how are you gonna really
- 3:58leverage that effectively when you use an open AI?
- 4:01Yeah. And, and I mean, I definitely
- 4:03see that there, there is a, there is a pressure from, from
- 4:06the top to, for the different departments to, to use AI to
- 4:10become more efficient. So I'm sure that, you know, an
- 4:12easy win would be deploying GBT or meeting summarizer.
- 4:16I mean, those are sort of easy things that we could build in
- 4:19house that are, you know, extremely useful.
- 4:22Totally. Where does it and you touched
- 4:25you touched upon use it where breaks was in the context and
- 4:30reasoning, right? Are there other factors that
- 4:33lead to a pivot point where a customer chooses to build
- 4:36something versus buy something? Yeah.
- 4:38I mean, I think there's on the surface it seems easy to just
- 4:42feed things to an LLM, but it starts to break down when you
- 4:46try to incorporate a bunch of different data sources, right?
- 4:49Let's take a classic example. I met with a customer yesterday
- 4:53and they wanted to have improvements around their build
- 4:55system. So when one of their development
- 4:58in the end development environment, so when they're
- 5:00build breaks, they have, you know, a Jira ticket that's open,
- 5:04they leverage Slack, they have a Jenkins build server.
- 5:08And so they're trying to piece together their own agents on top
- 5:14of open AI. But again, it's it's difficult
- 5:17to take a coherent snapshot of all these different tools data,
- 5:21send that to open AI, get a coherent response back that's
- 5:24meaningful and useful. In that example that you give,
- 5:28which I love, are they creating, are they using like the GR agent
- 5:33and then taking the other agent and having those agents talk to
- 5:36open AI, Is that how they are? Yeah, it's kind of a mishmash of
- 5:41sorts. Yeah.
- 5:43So maintaining it is. AI cut you off there or you were
- 5:46gonna say something? No, no, yeah, I was just gonna,
- 5:48Yeah, just answering that. Yeah, for sure.
- 5:50This is this. It's kind of a hybrid approach.
- 5:52And This is why they're coming to us because they, they've now
- 5:55realized just how difficult this is.
- 5:58It's not an easy thing to do this and to maintain.
- 6:01Beyond beyond, you know, beyond let's say like looking at Dev
- 6:08Rev. What are some of the other sort of key moments in a
- 6:13customer's life cycle, the way they might decide to buy AI
- 6:17versus build their own, like on this internal ChatGPT thing,
- 6:21Like when might they say like, OK, I should invest in another
- 6:25tool. Let's say it's a glean, right,
- 6:27versus build my own search. Like why can't I do do that?
- 6:30What what are sure? Does that come up?
- 6:32When do you think about those? Well.
- 6:34I think you that one keyword you just you just hit on the head is
- 6:37a search, right? Being able to understand across
- 6:41a number of different data sources what's going on real
- 6:44time. And not every tool will be able
- 6:46to give you that real time response, right or near real
- 6:49time response. So I think people start looking
- 6:54at products to buy because they realize that again, sitting all
- 6:58that kind of response data in your context and being able to
- 7:02summarize, especially look, I mean, companies have huge
- 7:05databases, right, Salesforce, I met with another customer that
- 7:09we have in the pipeline a month ago and their Salesforce
- 7:13database is over 30 years old. I.
- 7:17Mean they were, I think they were like the number customer.
- 7:21Yeah, exactly. And it's huge.
- 7:23It's a huge database. So building your own AI to be
- 7:26able to introspect that type of data and handle that type of
- 7:30data efficiently, it is extremely challenging, if not
- 7:33impossible. For anyone else, what did they
- 7:35do? Did they?
- 7:36Did they attempt to do something before before?
- 7:39Yeah, they have some of their top engineers actually working
- 7:42on AI. And this goes back to previous
- 7:43questions like, look, AI is the thing.
- 7:46And like a lot of CT OS VP's of engineering, they know how
- 7:50important it is to get on this train.
- 7:52So, yeah, this company, they, they put their best on it, give
- 7:55it, they gave it a good shot, you know, but then they came to
- 7:58us. What were they trying to achieve
- 8:01in terms of in terms of putting a on top of the sales force?
- 8:05Well, so what's interesting about big enterprise companies
- 8:08is the fact that they just get siloed over time.
- 8:11Not only their data sources, but people just don't know what
- 8:15other teams are working on. And I experienced this first
- 8:17hand at my prior company that I, that I used to work at, which
- 8:20used to be a startup and grew, which was great.
- 8:23But over time, you know, company, the tools, the teams,
- 8:27they get siloed. So one of the challenges that
- 8:30they face is, you know, being able to leverage these data
- 8:34sources across both teams, I'm sorry, multiple teams and to be
- 8:39able to work together. And that's, that's one of the
- 8:41difficult things too is, is being able to coordinate
- 8:45effectively between these teams. And then you're asking to write
- 8:48software, agentic software that can, you know, work across all
- 8:52these different internal teams in house.
- 8:54So it's really trying to tackle that data sprawl and being able
- 8:59to tackle that problem that I just spoke about, about
- 9:03different teams being involved. Yes.
- 9:05So they were trying to basically consolidate all of their data in
- 9:11one system and get it ready for the LLM so that they could do
- 9:15search or something like that. That's right.
- 9:18Yeah, Exactly. Exactly.
- 9:20Awesome. Any other interesting use cases
- 9:22that you've seen for what companies are trying to do right
- 9:25now with AI and maybe some of the limitations they're
- 9:28encountering? Yeah.
- 9:29So I mean, there's so many different areas that I find.
- 9:34So it's really hard to kind of say one or the other, but
- 9:38definitely support a customer support is a big one, right?
- 9:41So having cuz every company at this point has a chat bot.
- 9:46So you go to a website, you try to purchase something, they're
- 9:49like, all right, let me click on this chat bot.
- 9:51And believe it or not, that's a big challenge.
- 9:54Like I see big companies, enterprise companies all the
- 9:57time and their chat agent is awful.
- 10:00And majority of the time it leads them to a customer, A
- 10:04customer support escalation has to be redirected to a human.
- 10:07So I'm seeing a lot of use cases in that area.
- 10:10And it's not even like their, their chat is such a basic chat.
- 10:14It doesn't even, it's not even AI driven.
- 10:15It doesn't leverage some of their knowledge base articles
- 10:19say. So this is 1 area that that is a
- 10:22huge advantage for us because we, we can tap into all your
- 10:26knowledge base articles, we can have access to all of your
- 10:29critical tools data, so we can provide effective answers from
- 10:33the agent experience. Yeah, and actually you said, you
- 10:36know, like the the chat bots are seemingly simple
- 10:40implementations, but often leave the customer even more
- 10:44frustrated than they were before talking with the chat bot,
- 10:49right, For sure. And you know, on a scale of like
- 10:53one being like, oh, this is easy to do.
- 10:55And 10 being like very difficult, What what is the
- 10:58typical like timeline for when a company starts?
- 11:03Oh, this is this is going to be easy.
- 11:04I'm going to start doing it. Then how do they get to like 10?
- 11:07We're like, OK, this is like, you know, maybe I should
- 11:10consider doing doing something different or calling in the
- 11:12experts for sure. Well, I think every, especially
- 11:16developers like I'm a, I'm a former developer myself and we
- 11:20always think that things are easy be on the surface like, OK,
- 11:23well, I think I can do this. So it starts out as like a
- 11:26three, right? And then you start working and
- 11:28then you start having to deal with some of some red tape or
- 11:31maybe data access issues because hey, I need to get some critical
- 11:36data from our Salesforce database crap.
- 11:39But they have critical financial information in there.
- 11:42So that door shut, OK. Or I don't have a license, I
- 11:46have to track down the API. Exactly.
- 11:48That's right. Yep, Yep.
- 11:50So that's always up. And then you run into
- 11:54implementation issues, maybe rate limits like how much data
- 11:58are you sending to the AI or have any questions throughout
- 12:01the company are being sent to your AI?
- 12:04Can you handle it? And so you finally get to the
- 12:08point where you're like, all right, you know, becomes three,
- 12:11becomes 5, becomes 8. And then, you know, it's, it's
- 12:15it's time to start looking for solutions out there.
- 12:17Yeah. Now let's go back to kind of
- 12:19what we talked about in terms in terms of how, you know, you can
- 12:23create those agents or you can introduce AI within the system.
- 12:27And that one right, that you had that's what they they, they,
- 12:31they did, you know, in the beginning, sure.
- 12:33When you do that and you create these like AI agents within the
- 12:38system and then you have those agents, let's say, let's say you
- 12:41could do it. Talk to talk to like a bigger
- 12:45LLM across of that. What are some like potential
- 12:49issues that might come in? Sure.
- 12:51Well, even if you had the ability to create an agent and
- 12:55you're able to span a few data sources, which is good, it's a
- 12:59it's a good accomplishment. The next thing that you run into
- 13:02is how do you do this in a role based manner, like role based
- 13:05access, right? So if you're a support engineer
- 13:09and you're trying to ask the agent about financial
- 13:12information, do you have guard rails in place?
- 13:14Did you think about that? Like in the context of this
- 13:17user, what kind of like response are you going to give them that?
- 13:20That's a real problem and it's something that dev Rev that we
- 13:24put a lot of thought into. You know, I would say the first
- 13:28couple of years we knew this was going to be a really big
- 13:31challenge across multiple data sources.
- 13:34So role based access was was a really big thing.
- 13:37And then I and it's hard to get it right.
- 13:39Yeah. Are there like other?
- 13:41Are there other things around? For example, the the individual
- 13:45system, AL and M's becoming so finely tuned that they cannot
- 13:51correspond, you know, or or it's so contextualized to just that
- 13:55one thing. I don't know.
- 13:57Tell us more. Well, sure, sure.
- 13:58Yeah, yeah. I mean, you can build an agent
- 14:01to do 1 very specific thing, but what happens when you want to
- 14:04give it more skills? What happens when you want to
- 14:07expand the capability of that agent?
- 14:09Is it easy to do? How many man hours is it going
- 14:11to take? Are you having to rewrite, re
- 14:14architect your system just to accommodate these new use cases?
- 14:17And can you spin up another agent, you know, clone that
- 14:20agent, spin another one up, but maybe have it do a slightly
- 14:23different use case and attach it maybe to a different surface.
- 14:26You know, I, I spoke about the chat widget earlier.
- 14:30Maybe you have another surface that you want an agent to
- 14:33operate within in your company. And do you have the ability to
- 14:36do that? So these are some of the
- 14:38challenges that we thought of upfront and, and that are that
- 14:41are important. And I would say, you know, with
- 14:43Deb Revs, Workflow Builder and Agent Builder, you know, these
- 14:47are things that we put a lot of thought into on the design side
- 14:50before we executed. Yeah, and and you know, I, I
- 14:54think there there is a use case for internal AI like when it's
- 14:57very simple, like meeting suffer, that's very that's very
- 15:00easy and stuff like that. But it sounds like from you know
- 15:03what I'm hearing from you, the problem isn't so much like in
- 15:06the LLMS or in the AI models, like they sound very good.
- 15:10It's really more in the way that the data is accessed absolutely
- 15:16missioned and that if the data is in unified, it's very
- 15:20difficult to do meaningful AI. You know, I read a stat there
- 15:26where it says like 40% of AI project cost actually comes from
- 15:29preparing the data. Yep, that's a lot.
- 15:33And, and I'm sure, you know, there's additional like services
- 15:35cost on top of that and stuff like that.
- 15:37So most people I, I think where like ChatGPT became like very
- 15:42mainstream was not so much the models and models maybe existed,
- 15:45but it was in the fact that they had access to all of the data,
- 15:48right? So.
- 15:49That's. Right.
- 15:50Tell us more about what dev ref does differently.
- 15:53And you reference them early on in the conversation.
- 15:55But tell us more about like how does dev ref do it differently
- 15:58to solve for that data problem? Yeah.
- 16:01And and to your point, the model is really the problem, the
- 16:04blocker, the missing piece is connected trustworthy context.
- 16:07Deborah tackles that at architecture level, we maintain
- 16:10a unified knowledge graph that keeps the customer and product
- 16:13support and engineering data In Sync near real time.
- 16:16So like all these different teams can collaborate and we're
- 16:20able to you know, connect these the dots in real near real time.
- 16:25So resolve, we resolve identities.
- 16:27So you know, company XYZ is company XYZ everywhere and we
- 16:31apply permissions at retrieval. So each answer is both correct
- 16:35and safe. On top we expose semantic layer,
- 16:38so definitions like active users or churn risk are consistent
- 16:42across every agent and an action layer with approval so that AI
- 16:46can not only answer but also create tickets, post updates,
- 16:50draft outreach with Lineage. That foundation is what let's AI
- 16:54scale beyond a single Slack bot or Salesforce bot and you get
- 16:59faster, more accurate answers and you can move from insides to
- 17:03actions without any of the duct tape that you need to usually
- 17:07use. And also analytics.
- 17:09Analytics is also a huge factor there.
- 17:11So it's not just about agents and workflows.
- 17:14It's about how can we look at this data overtime or trends
- 17:16overtime. That's also a big win for Debra.
- 17:20How? How does you talked about the
- 17:22knowledge graph and you talked about like that being able to
- 17:26provide people with the, the organizations with the context
- 17:30that the LNM needs, right? Can you tell us more about how
- 17:33the knowledge graph works? Like, how is it that we're able
- 17:35to do that? And why?
- 17:37Why is it, you know, difficult to replicate, let's say?
- 17:40For sure. Yeah.
- 17:41I mean, I think when you hear about knowledge graphs, people
- 17:44automatically assume like a graph database, which couldn't
- 17:47be further from the truth. We don't use a graph database.
- 17:49It's something that we put a lot of thought into.
- 17:51So we have a specific ontology and schema in the knowledge
- 17:54graph with vector database on top.
- 17:57So when we ingest this data, it's weighted and we once you
- 18:01lay AI on top of that, that's when you really get the power of
- 18:05search and being able to traverse the knowledge graph in
- 18:08a fast and effective manner. So the secrets in the sauce, But
- 18:12yeah, this is the knowledge graph is by far one of our most
- 18:15powerful pieces and that allows a gentic experience to be
- 18:19unparalleled to other experiences.
- 18:22And and so you, you mentioned that, you know, we put it into a
- 18:25vector database, which is really what the graph is.
- 18:27It's not a graph. It's, it's a, it's a vector
- 18:29database. And am I understanding it
- 18:31correctly that you know, when most other kinds of data
- 18:36ingestion are sort of more flat files, They're, they're not
- 18:40vector based, so it's difficult to make those connections?
- 18:43Yes, absolutely. And some don't even bother to
- 18:45make a knowledge graph. They do it.
- 18:47They do API calls, so they're not even ingesting the data.
- 18:51And then you hit rate limits, right?
- 18:53What happens when you hit your Salesforce database so many
- 18:57times or some other database and you're taking API tokens at the
- 19:01same time, right? So the benefit of having the
- 19:03data itself and organizing in a way that you can traverse it
- 19:06easily is extremely important, especially when you talk about
- 19:12the agentic experience and making that smooth.
- 19:14Yeah, and maybe for some of our less technical guests on the
- 19:16show, tell us what in a vector database is.
- 19:19In short, it's a bait way to weight answers, right?
- 19:24So and it's a way to correlate data.
- 19:26So being able to match certain questions to their answers more
- 19:31appropriately. And that's kind of the dumbed
- 19:33down version, but that's essentially being able to
- 19:36correlate data in a weighted. Manner.
- 19:38So if I understand correctly, what we're doing is dev ref is
- 19:40taking the data, ingesting the data, they're putting it in this
- 19:44like holistic vector database known as the knowledge graph
- 19:47that allows the LLM to access all of these nodes in a way that
- 19:53can be contextual. Correct.
- 19:55Yep, yeah. OK, got it, got it.
- 19:58And are there any kind of what, what are there any kind of
- 20:02limitations right now or challenges when it comes to
- 20:06prepping this data either with dev Rev or or with a a
- 20:10different, you know? As far as data prep, I mean, one
- 20:14of the things, one of the other tools that we've put a lot of
- 20:17effort to into is our data ingestion tool, which we're
- 20:21calling AirDrop. And it helps customers when the
- 20:26time comes to bring dev Rev into their environment to connect to
- 20:30their data sources and to do that import, right.
- 20:32So we take care of rate limiting, rate throttling, data
- 20:35ingestion. How do you map certain objects
- 20:37into a dev Rev object, right? So if we have customers mapping
- 20:41to customers, if we have tickets mapping, mapping the tickets or
- 20:44bugs mapping the tickets in the GR case.
- 20:46So we have a very beautiful UI driven workflow to help
- 20:51customers do that mapping, right?
- 20:53Because that that's, that's what matters in the and what data
- 20:56should you import? Maybe there's a bunch of tables
- 20:59in your sales force database that are completely empty or you
- 21:02know, have very low usage. Do we want to import those?
- 21:05Probably not. So that's, you know, the data
- 21:08ingestion and having a nice workflow to do that and tool to
- 21:12do that is also very, very important.
- 21:15And is there a use case? I know that right now.
- 21:18I know that right now our customers, you know, they, we,
- 21:23they use, they use their verve for, for the agents on top of
- 21:30it, right. And that's really like the
- 21:32finalized version of why they come to us anyway to solve the
- 21:36business problem. Do you think there ever be a use
- 21:40case where customers are are coming to us and saying, hey, we
- 21:44want to do something else internally, right, Like we have,
- 21:47we have a great team of AI engineers, but can you like
- 21:51prepare the data and put it in your knowledge graph and then
- 21:55let us deploy our own terms or whatever.
- 21:58Do you think that's that's a foreseeable future?
- 22:00I mean if you're talking about deploying agents within dev Rev,
- 22:04is that your question? No, like not deploying agents
- 22:08within dev Rev, but deploying external agents on the dev Rev
- 22:11data but just using the vector database as the like.
- 22:16That's the use case. I I don't see it immediately.
- 22:21I mean, I will say though that we allow you to use your own
- 22:24models. So if you've trained your own
- 22:27LLM, we can actually leverage your, your model.
- 22:30I mean, we, we leverage a different, different models in
- 22:33house. But if you, you want to bring
- 22:36that to the table, then by all means.
- 22:38But yeah, I don't there are have been some in use cases where
- 22:43let's say the customer is using the sales force agent, support
- 22:47agent and it's really highly integrated into their
- 22:50organization. And we went the extra mile and
- 22:54we were able to actually integrate our agent, but be tied
- 22:58to their their sales force in the workflow, which was pretty
- 23:01unique. So in some cases, you know,
- 23:04we've seen that they've just been so bound to that, that
- 23:07specific workflow. So it's hard for them to
- 23:09transition to yet using switching their entire support
- 23:12team to using ours. So we did kind of a hybrid
- 23:15approach. But yeah, I don't see a, a lot
- 23:17for now of customers asking to write their own external agents
- 23:22to, to use our knowledge graph. But who knows?
- 23:24Yeah, maybe in the future that could become an interesting use
- 23:26case. Right.
- 23:27And and if like, you know, if if I was a buyer and I was thinking
- 23:31of building, you know, building versus buying an agent or an AI
- 23:38platform like Devre, what are some questions I should be
- 23:42asking myself or my team? Wow.
- 23:45I would say, how much time are you spending waiting through
- 23:49different data sources to accomplish what you need?
- 23:51How much manual effort or how many what?
- 23:54What processes are broken and why are they broken?
- 23:57Right. It could be, hey, I have like
- 24:00every single support case that's opened on my chat interface.
- 24:06Let's say 70% of those get redirected to a human.
- 24:09So where are you losing your time at?
- 24:11Where are you losing your man hours?
- 24:13Where can AI help with with the knowledge that you already have
- 24:16within your company, right within your knowledge base,
- 24:18because the power of LLMS are incredible.
- 24:21So I would say, you know, where can you get your efficiency in
- 24:25your in your man hours back? That's that's probably the
- 24:28biggest question. Mm hmm.
- 24:30Because AI is just going to continue to get better and
- 24:32better at these things, and it's important to find those gaps
- 24:35within your company. Awesome.
- 24:37Thomas, thanks so much for talking us through that.
- 24:39Was there anything last words from you that you want to share
- 24:41with the audience? Yeah, no, I'm just excited.
- 24:44I'm really excited about this phase of our company and AI as a
- 24:48whole. So looking forward to many more
- 24:50meetings and to learn about where companies are using AI and
- 24:54where they want to go to the next level.
- 24:56And hopefully, hopefully we can be a part of that journey.
- 24:59Awesome, thank you. So I think in summary, I think
- 25:01what we talked about was really Federated AI is going to cause
- 25:05AI chaos the same way we got to with SAS.
- 25:10We're serious about scaling AI and consolidating that.
- 25:13We need a solution to help us get the data ready and
- 25:18consolidate that in a way that is, you know, AI ready.
- 25:22And it just so happens that Defra does it, you know, what do
- 25:24you know? Yay, exactly.
- 25:27Cool, Thomas, thanks so much for joining us and helping us unpack
- 25:31that problem. And to all our listeners, if you
- 25:33enjoyed this content, please leave us a comment like our
- 25:38session and let us know I will see you next time.
- 25:40Thank you. Thank.
- 25:41You, Laura, appreciate having me.