Your AI Strategy Is a Data Model Problem

0:00 So Jake and I were at the UpData Partners CEO Summit the other week and the presentation really,

0:08 really landed. It was a room full of CEOs. We presented on the idea that your AI strategy

0:15 is really a data model problem. And we felt like it resonated so well that we wanted to record the

0:21 presentation, go through it, and then just share it with the audience and all of you.

0:26 So relevant in the market right now where everyone's just throwing out Claude licenses,

0:32 trying to increase their usage where they can and try to maximize the value that they get in return

0:37 from that token spend. And I think it's transitioning now from how much can we do

0:44 with these solutions like chat GPT and Claude to how optimal can we use them going forward.

0:50 And summarize everything, semantic layer, write down your definitions, find a good place to store

0:56 it that you can put your AI to, context graph, connect the dots of your data between systems,

1:02 even within systems. Man, once mythos comes out, once chat GPT launches their new model,

1:08 this is going to change everything. More horsepower isn't going to build the map,

1:12 more intelligence isn't going to give you the context. And then making Usain Bolt run faster

1:17 isn't going to solve the toddler leg. If you're in a $20 million plan with a 20% conversion rate,

1:23 increasing the conversion rate by one point will give you another million dollars in the plan.

1:30 Do not let AI author your business logic. It is not there yet.

1:46 So Jake, thanks for being my wingman out in DC and then my wingman for this as well.

1:53 And I'm really excited to get into it. Yeah, same here, man. Like you said,

1:58 this is just so relevant in the market right now where everyone's just throwing out Claude licenses,

2:04 trying to increase their usage where they can and try to maximize the value that they get in return

2:10 from that token spend. And I think it's transitioning now from how much can we do with these

2:17 solutions like chat GPT and Claude to how optimal can we use them going forward. So this is kind of

2:25 the first step in that. And I'm really excited to take this out to the broader market with you.

2:30 Awesome. Awesome. Well, let's dive into it. We're going to go through the same presentation we went

2:34 through in the crowd. So we'll just go through the slides. We also have a few giveaways at the end.

2:38 So if you hang on for the end of the podcast, you'll have a QR code where you can go download

2:45 a lot of the materials that we have access to here. And then just you can get everything going

2:52 in your company. So first couple things we're going to go through. We're going to really anchor

2:56 this into why this is important. So this again, in the context of revenue operations, marketing

3:03 operations, go to market for B2B SAS and AI companies, we all inherently have the FOMO of

3:11 like, hey, we need to be implementing AI. We know it's important, but let's remind ourselves of the

3:18 why. Then we'll go into what it actually is and what we're talking about, break it down in a really,

3:25 really detailed manner, show you the art of the possible. So once you have a really strong data

3:29 model in place, what are the type of AI agents and workflows you can build and work on? We'll give

3:34 you a full roadmap on how to deploy it, how to set it up in your company. We'll give a quick reminder

3:39 of why it's important, and then we'll have the giveaways at the end as well. So before we fully

3:45 get started, I think this was one of the inspirations. It was asking a super on paper, simple

3:52 question, which of my customers are likely to expand this quarter? Which seems like a really

3:59 natural thing. You would just type in a cloud and be like, hey, I'm connected to Salesforce,

4:03 I'm connected to HubSpot. It even has access to, you know, Gong for my transcripts. I should be

4:09 able to ask this and just hit go. And then I think what happens that I think we don't appreciate

4:16 is there are a lot of layers to this question and a lot of things that the AI is going to need to

4:22 know like, what is a customer? It seems like a really simple question, but where is my source

4:29 of truth for what a customer is and what it isn't? What does expansion mean? Are we talking about

4:37 committed contracts? Are we talking about more usage? What do we mean by expand? And then what

4:42 drives likelihood? What in your world means a company is likely to expand? So all of those things

4:51 are struggles an AI model will have if you're trying to ask this question. So Jake, on the

5:01 presentation, this was I think my favorite part, you going through the metaphors, because I think

5:05 it helps anchor like how wait, shouldn't the AI just be able to figure this out? But anyway,

5:10 I'll kick it to you to go through some of these. Yeah, absolutely. And a lot of this, you know,

5:15 is lessons learned the hard way, as Anthony was alluding to asking a simple question, the AI sees

5:21 all of this data and all this information. But again, it doesn't really have the understanding.

5:26 And we're going to talk a lot about that. And a lot of the times, what we thought,

5:31 even six months ago, was man, once mythos comes out, once chat GPT launches their new model,

5:38 this is going to change everything. It's just going to work. And what we found is that's just

5:43 not the case. The model is already trained on all the publicly available information that's ever

5:49 been created in history. It's a mind boggling amount of information. But what it doesn't have

5:55 is the context and insights that make your company unique, how you run your business, how you operate,

6:01 everything you need to guide it to that next stage. And we really think of this like an F1 car,

6:07 it is a highly tuned optimized performance machine. That's what these providers are bringing to these

6:15 LLM models. They're best in class, best in breed, and they're making incremental gains on things

6:21 like coding in different areas. But an F1 car, if you don't give the driver a map, he's going to have

6:28 no idea where to go, right? We could be driving in the complete wrong direction for miles before

6:35 we realize, hey, we need to course correct here. So a lot of what we talk about today is going to

6:39 be how you build that map in a very, very simple way. Another one that I really love, I really love

6:47 this one, is you could hire the smartest person that you've ever met. And let's say you haven't

6:52 met them. We'll use Einstein in this example. He comes in day one as a GTM analyst looking at all

6:58 the information in your CRM. He doesn't know what to make of it. You ask him what deals are at risk.

7:04 He has no idea. He doesn't know which ACME is the real ACME across all of your different systems.

7:09 You've got ACME Inc. in your marketing automation platform, ACME Co. in your CRM, and just ACME in

7:15 your product database. Are those the same client? Are they different? How was he supposed to infer

7:20 that on his first day? And that's what these models do. They just make up the middle and figure out the

7:26 path along the way. We're going to show you exactly how you steer them in the right direction. So when

7:31 you hire Einstein that first day, you bring on the new model, whatever rolls out next, it has all of

7:38 the information needed to be successful. And then lastly, I love this one as well. Anthony and I both

7:45 have toddlers at home. So this one, this one hits deeply. You could have the fastest runner in the

7:51 world in a two leg sprint. The first leg is run by Usain Bolt, whoever the fastest fan alive is

7:57 nowadays. You're not going to make incremental gains by coaching him when the second leg of the

8:03 race is being run by a toddler. What we see here is again, tripping, falling, stumbling over itself,

8:09 getting distracted by other things and running down a weird path where you're chasing them and

8:13 trying to pull them back to what you want them to focus on. Again, you need to coach that second

8:19 leg. You need to steer the toddler in the right direction. One more that I'll throw out here is

8:25 essentially the Einstein example. You're hiring these AI models every single day. Every single

8:31 day is their first day that you're onboarding them. So what we're going to show you is how you get

8:36 out of the onboarding constant cycle of informing the AI, telling them what's going on, bringing

8:44 them up to speed and really operating from a consistent foundation. So again, regardless of

8:49 the model of the tool that you're bringing on, everything is operating off of the same foundation.

8:55 Yeah, and I think just to reiterate what you've said, there's just so many people. I was kind of

8:59 in this camp too, where it's like, oh, the model is going to get better and do the work for me.

9:05 It's like, no, categorically, just by the sheer definition of what it is we're talking about,

9:12 more horsepower isn't going to build the map, more intelligence isn't going to give you the context,

9:17 and then making Usain Bolt run faster isn't going to solve the toddler leg. So I think just making

9:24 sure people know that, oh, this isn't something that the model will just solve for me is really

9:30 important. And then why is this important? Okay, we're all in go-to-market, we're in revenue

9:36 operations roles, CROs, CMOs. Why do we even need to be deploying in the first place? And I think

9:43 for us, it really all comes back to one thing, which is maximizing yield from your go-to-market

9:50 machine. A lot of these companies, I know, Jake, you've seen a ton of them, they're investing

9:57 millions, they're investing tens of millions of dollars into these massive machines and

10:03 organizations that are designed to go out and get revenue, make bookings, grow customers,

10:08 convert deals. And a lot of times, companies really aren't even measuring what the yield

10:13 of this investment is bringing. And the most effective way to get more production out of

10:21 the investment you've made in your go-to-market engine is to increase yield. So increase your

10:26 conversion rate, reduce your time and stage. If you do this really well, I know people understand

10:32 that the gains can be really massive. Sometimes I think people forget the magnitude. If you're

10:39 in a $20 million plan with a 20% conversion rate, increasing the conversion rate by one point

10:47 will give you another million dollars in the plan. And reducing the sales cycle by one day

10:54 can help cycle in a quarter of a million dollars of revenue, which could be the difference of making

10:59 your quarter or not. So I just wanted to share, I think this is people, they know it's important,

11:06 but sometimes I think forget the magnitude. And especially if you combine both of these at the

11:11 same time. So when we're talking about leveraging AI to get more out of your go-to-market and why

11:17 you should invest in a robust data model in order to do that, this is why. It's how can I increase

11:23 my conversion rate by a couple points? How can I reduce my sales cycle by a couple of days? And if

11:28 you can do both of those at the same time, the results are really going to compound. I think

11:33 unfortunately what people are missing right now is they're not getting everything out of the models.

11:42 The ceiling of what value they could be extracting out of AI right now is completely capped and

11:48 limited by the data model they have in place. So I think all of those potential conversion rate

11:54 improvements, time and stage improvements, really are being lost until you get this data model

11:59 problem figured out. Okay. So now we know kind of why it's important and seen some numbers behind

12:05 it, which is pretty illuminating. When you think about those just small incremental gains and how

12:11 much they add up to over time, I think it's super important to show exactly how you do this and how

12:16 we see clients getting started with this. I think also there's a perception in the market that

12:23 we need to fine tune a model or we need data science to come in and build kind of a custom

12:28 distilled version of some of these models that are unique to our business. When in reality,

12:32 it's much, much simpler than that. So first and foremost is the semantic layer. And here at lean

12:39 scale internally, we call this our GTM brain. Essentially all this is to boil it all down and

12:45 make it nice and simple is all of your business logic, all of your context written down codified

12:52 in some sort of repo. We use GitHub internally. It's widely distributed. You can start for free,

12:59 totally easy to set up and it integrates with all of these different LLMs nice and cleanly. They

13:04 know exactly how to use it, but you can also use something like Google drive to house this.

13:08 Wherever makes the most sense for your team and your agents to be able to access this data

13:13 is where it should ultimately live. The main thing here, there's no real limits to what you put into

13:20 this GTM brain semantic layer. We have everything from our brand messaging, our design kits,

13:27 who we sell to and who we don't, what products we have on the shelf today,

13:31 what good looks like to us for each individual project that we run. And then we also have where

13:37 to find key business insights and really defining all of those at a super fine granular level.

13:43 So this starts just as simple as saying, what does win rate mean at our company? How do we classify

13:50 bookings and what value should we be looking at across different pipelines, across different

13:56 segments, whatever that looks like unique to your business should be documented and codified here.

14:02 The key thing that I'd really like to point out, and I'll talk about it a little bit later as well,

14:06 but having one owner for these metrics is massive. You need to have one person who's really controlling

14:13 the narrative on each of these. And that's to avoid many cooks in the kitchen and CS not buying

14:19 into the way that sales is operating or marketing is operating. Everybody should be working again out

14:24 of this consistent foundation. We've seen this across multiple clients where we ask a question,

14:32 what's my forecasted ARR for this quarter? Super simple, nice and clean. The AI goes out,

14:37 hits your back end to look at your product usage and start doing some forecasting.

14:42 And what it does is it runs some very simple SQL. The SQL responds with data and the AI says,

14:48 nice, I did it. I got some data. This looks good. These are actual numbers. I can return this to the

14:54 end user. What the AI doesn't understand is the nuance in your CRM or in your data warehouse,

15:00 where we're looking at opportunity products or product usage, and it's fanned out across

15:05 multiple line items. So now a $10,000 deal looks like a $30,000 deal in your forecast.

15:12 You're over reporting over estimating. And that's just one very small example of what we see day in

15:17 and day out. Now the context graph, again, to boil this all down and make it super simple for everyone.

15:24 All this is, is connecting all of the different dots that you have, all of the different data

15:28 points that you've gathered through enrichment, through events, through outreach, whatever that

15:33 may be. It's so important that you have this data model defined so the AI isn't guessing which

15:40 records are associated to which. We really think of this in a crawl, walk, run format. And we think

15:46 most should just start with making their CRM their true first context graph. All of the data

15:52 should be there already. If it's not, bring in key product data, key usage insights. Whatever is

15:59 missing to make it a complete picture of your organization, that should be tied back to your

16:04 source of truth, which we ultimately think should be the CRM. So it's just making sure your contracts

16:11 are associated to the right opportunities, which are associated to the right accounts,

16:15 which have the right contact roles defined so that the AI can go in and very quickly understand

16:21 what an account looks like, all the key players in the game, what a good renewal looks like, what a

16:26 bad renewal looks like. If it's disparate and disconnected, you're going to get overinflated

16:32 or underinflated numbers and not be able to address the root issue. There's also an aspect of this

16:38 that comes down to tokenomics. Again, going back to six months ago, we had a big token leaderboard

16:44 and all we wanted to see was that number keep climbing. And now what we want to see is the output

16:48 relative to the number of tokens. So we're really looking at this from an optimization perspective,

16:54 where your AI isn't going to have to index your entire CRM every time. If it has these connections,

17:00 it makes very easy hops through the data. Instead of saying, here's this opportunity, can you go look

17:05 at all the usage data behind this? Now it has to index this entire database and try to piece

17:11 together exactly which of the custom objects are tied back. So this is where we see extreme value

17:17 in getting your CRM dialed in. When cloud force rolls out, as we've seen with the native connector

17:22 and all the skills that they've baked in, that's not going to save you on the token side if your

17:26 CRM doesn't have the good quality associations to meetings booked to the tickets filed. All of that

17:34 is so, so important to again, building that holistic picture of what's going on inside your company.

17:40 Yeah, I think a lot of people don't realize. So you're mentioning like, yeah, lean scale,

17:44 we were token maxing all day and a great time is fun. I think a lot of creativity, but

17:54 we definitely are going to have to get more efficient. You know, anthropic opening eye,

17:58 they're already starting to cut how, what your token limits are within some of their plans.

18:04 So understanding, okay, how do I make this a more efficient process? It's kind of like a nice to have

18:10 maybe six months ago in the future. This is going to be unbelievably important. And I think your

18:18 CFOs are going to be all over this. So laying the foundation early now. Okay. We got the semantic

18:24 layer in place, got the context graph set up. Um, there's different levels to this too, but

18:29 getting that done early is going to save people a lot of money.

18:33 A hundred percent. And as I always say, it's a matter of when not if we know that these providers

18:40 are going to pull the rug out on these gravy train max plans that, you know, for $200, you get

18:46 $5,000 worth of usage. And they know that they have all the insights on that usage and they are

18:53 going to start cranking up the dial. So to your point, if you address this now, while you're still

18:58 on the gravy train, it saves you from the fire drill, two, three, four, who knows how many months

19:04 from now, when you get that first 25, $30,000 a month bill and have to go explain to your CFO

19:10 exactly why you need to cut budget in other areas to support this overinflated AI usage.

19:17 If I were to guess, it's going to be one quarter post IPO from either open hair and drop pick or

19:24 whoever. So again, to tie it back and a little bit, I guess I'm a pessimist, so I used which deals are

19:31 at risk instead of which deals are going to expand, but just to tie it back to a core business question

19:37 that, you know, each and every one of us probably have questions about and have plugged into one of

19:41 our LLMs in the past. So before it really started to look at how many touch points are on this

19:49 account, right? And so a healthy account might have X number of touch points, but what it doesn't

19:55 understand is the sentiment behind some of those touch points. Tools like Gong are great at this.

20:01 They've been doing this for a long time. Sentiment analysis is so important, but when stuff comes in,

20:06 let's just say through email, through Slack, you're not necessarily going to get that insight

20:11 straight out of the box. So an account that has 150 touch points looks great until we start to

20:18 look and say, no, those are support tickets that haven't been answered, right? We need to start to

20:23 classify what good touch points look like, what bad touch points look like. And that's where the

20:28 semantic layer comes in is it really flattens this out to say, what are all the really good

20:33 opportunities looking like? What are all the really negative ones looking like? So we start to do this

20:37 cohort analysis and really break down the data into these identifiable buckets that can then be

20:43 modified over time. So this isn't something that's permanent set in stone. You can always modify this,

20:49 but without it, you're really operating from a very vanilla assumptions-based inferred model,

20:55 no confidence behind it. Now the context graph that comes in with the additional receipts.

21:02 So that's going to tell you exactly why certain things happened, why certain touch points were

21:07 negative, how many of those in the ecosystem there were, and then putting that all together,

21:12 the semantic layer in the context graph in the response is something that a human can actually

21:16 take action on. And that is the key. When a human can take action on the insights given from an LLM,

21:23 that's again, when we see the maximum value coming from those tokens. Now, a key here,

21:30 do not let AI author your business logic. It is not there yet. And I don't think it ever will be,

21:38 as we've alluded to, it can be the best, most performant model in the world. That is going

21:43 to take the fastest route from point A to point B, especially as it starts to optimize on tokens.

21:48 It's going to say, yep, what are standard industry practices? Boom, ship it. We are good to go.

21:53 Unless you are running the most vanilla basic lemonade stand SaaS company in the world,

21:59 that is not going to work for you. What you need to do is just start writing it down in plain

22:04 English. Just start putting pen on paper. Start typing it out in a Google doc where AI can help

22:11 is putting it into a format that's good for GitHub and good for agents to access. It can turn it from

22:16 a text document into a YAML file, get all of that aligned, and then push it to GitHub. But what you

22:22 don't want to do is ask it, hey, go look at my CRM and tell me what a good renewal looks like. No,

22:26 no, no, no, no. That all needs to come from your insights, your historicals, and your team that are

22:32 working these accounts day in and day out. So as we've talked about, I mean, setting this up is

22:38 pretty cheap. You can basically get started with your current stack. If you have an LLM license

22:43 and you have a CRM, you're pretty much ready to go. Just start with the basics. Don't try to get

22:49 every possible metric that you've ever calculated done at the same time. Focus on the ones that

22:55 matter the most. And I always tie this back to the toughest meetings, the board meetings.

23:01 That's what I always try and tie it back to is focus on those metrics that you know you're going

23:06 to have to speak to in this quarter. Then we can start to build the ones that add additional juice,

23:11 help us move the needle on certain conversion metrics, or move things out of a stage faster.

23:17 Those are all great. But the ones that you really want to make sure you have confidence

23:21 on are the ones you have to defend in the board room. So I always try and emphasize getting that

23:27 first set of metrics done codified and with a set owner is so crucial going forward.

23:34 Then we start to look at ways to optimize the spend. And I think that's where the tokenomics

23:38 come in and getting really clean data. So you're not going through and indexing your entire CRM

23:45 every single time. And there's a lot of tools that can help in this space as well.

23:49 ETL tools are now coming up with these unique ways of surfacing data in a blended fashion,

23:56 where, again, you can pull in your campaigns from HubSpot, your gong calls, your clay data,

24:01 and your CRM data to form a unified model that you're not bringing into your CRM.

24:07 Or it's a distilled version of that. So only the key fields that matter to your company

24:11 when you're doing forecasting. And I'm not having my AI, again, dig through 500 custom opportunity

24:18 fields that we see. It's like a very distilled set to get the most juice out of those tokens

24:23 that you're spending. Yeah. And I know there's quite a few tools that are coming out. One we're

24:30 partnered with closely, Polyatomic. I think that's a really good option with their data harbors

24:35 model so they can support this. I know their founder CEO is going to be giving a talk

24:43 at OpStar coming up. So I think that's a good option. And I'm sure we're going to see more

24:48 of these too as time goes on. I think this is going to be a huge need for a lot of companies

24:54 that are trying to maximize the value of their AI and data model. 100% and have some sort of

25:00 middleman because we see that all the time of the fear of connecting an LLM directly to your CRM.

25:06 How do you control the data that it has access to? How do you make sure that reps aren't stealing

25:11 each other's accounts and overriding the ownership? There's all of these dilemmas that need to be

25:17 solved. And you alluded to one, Polyatomic is a great tool. We use it internally as well. And

25:23 it allows us to really bifurcate systems and only allow the agent to access the data that we've

25:29 approved. So again, we're distilling it down into this harbor. And then the harbor actually has its

25:34 own context layer so you can put skills in there. So then anytime anyone connects to that, they're

25:40 running the same consistent skills, the same consistent output on the same consistent data.

25:46 I think consistency is kind of the name of the game here. Yeah, another platform too. And it'll

25:52 kind of fit this maturity curve. I know you'll go through it a second. But Vasco. Vasco is a really

25:57 good go-to-market platform that pretty much forces the function of building your semantic layer and

26:04 will give you the context graph all in one. So as you go through the motions of setting up Vasco,

26:09 you get two for one and giving that setup. And then you can point your AI just to Vasco and ask

26:14 the questions you need to ask. So it's I think a couple of good options there. And I think there's

26:21 more to come. And I think hopefully it'll make this process a little bit easier in the future too.

26:28 Totally, totally. Yeah, and tracking to go is so important. There's just so much functionality

26:34 out there that I think most teams haven't quite started exploring yet that's going to be really

26:38 exciting over this next quarter. So on the maturity curve side, I think this one's interesting.

26:45 There's definitely a shift happening right now. I know a lot of the companies that we used to work

26:49 with, you know, we're evaluating a data warehouse, they might have like one little cluster that

26:55 they're tracking their product usage in, but no CRM data flowing between the two. And we're seeing

27:01 that shift earlier and earlier in the growth stages to where companies are investing pretty

27:06 heavily in this data infrastructure early on, because they know the value down the line of

27:11 tying this all together. So really what we're seeing is a lot of companies are at this centralized

27:16 level one zone where they have the BI layer, they have the data warehouse, they've got all

27:21 the infrastructure in place, but we still haven't defined where everything lives throughout that

27:28 ecosystem to build that confidence with the AI. So that's when you move to level two, resolve those

27:34 identities, really flatten out your duplicate problem. I know, I know it's a problem for

27:39 everybody. It's a problem for us. We've actually developed agents and tools around this because

27:43 it's just such a prevalent thing of duplicates, misaligned data, right? I'll go back to the ACME

27:49 example, but even something like Amazon, you've got Amazon, you've got AWS, you've got Amazon

27:56 logistics. There's so many different options in that one given account that it gets really hard

28:02 to tie that all together or the AI is looking at AWS when it should be looking at Amazon because

28:07 the domains are the same, really flattening that out, getting to that canonical identity model is

28:13 huge. And I know I'm using a lot of fancy words there, but all that really means is deduping your

28:18 CRM, getting one consistent key to identify an account, whether that's the domain, whether that's

28:23 the record ID and the CRM, just making sure you have that across all your systems. So that flows

28:29 all the way to the data warehouse. That's where we get to the definition metric. That's where you go

28:34 through and say, if you're looking for win rate, here's the column you should look at and here's

28:39 the filters you should apply. We're going to talk a little bit more about what that looks like,

28:42 but I think this is where we're really trying to push our clients to, is getting them to this point

28:48 where they have all their 20 to 30 key metrics defined, agreed upon, and version controlled.

28:56 Now the agreed upon part is big and Anthony's going to talk a little bit about that as well,

28:59 I'm sure, but getting that alignment is really the core problem. Sometimes difficult discussions

29:05 need to be had. We work with a ton of different companies. I think all of them recognize revenue

29:11 just a little bit differently. So it's really important that you come in and just define those

29:16 owners, define what these mean to you. And going back to the Amazon example, one of their values

29:22 is disagree and commit. And I think that's so important to bring into these meetings,

29:27 trust the owner, even if you say, no, no, no, no, we've been using this spreadsheet over here and

29:31 our calculations look much better. Well, that's great, but let's try this. Let's commit to this

29:37 and build on this foundation and bring your insights in. Let's get to an ultimate foundation

29:42 that we all agree upon. Level four, we move into the retrievable. So that's where a lot of the

29:48 polyatomic harbors we talked about getting really good, clean associations in your CRM. So you're

29:53 not indexing the entire database every single time is so important. This really comes down to

30:01 highly retrievable, quick insights that don't cost a lot of spend. That's the next evolution. And

30:07 again, you can get really advanced in this with the extra credit mode, going with a vectorized

30:12 database that sits alongside your product warehouse. You can even do this in a cloud

30:16 database, something like super base, using like a Postgres. But really, we think most should just

30:22 start cleaning and associating in their CRM. That's where you're running your business today.

30:28 That's where you're going to get the most juice out of all of the teams are working out of that

30:31 platform. They don't need to connect to an additional database. And ultimately, you'd

30:36 probably write a lot of this info back from that vectorized database to your CRM. So we really try

30:42 to focus on using the tools you have today and getting the optimal usage out of them when it

30:46 comes to AI. And this is a lot of work. It's not super easy to get all this set up. And I think

30:54 it's really the hard part. If you think about, okay, there's two different camps. There's get

30:58 the data model set up. And then there's build the agents, build the workflows. I think building the

31:04 agents and workflows is so easy. That's the cheap part. If you have a great robust data model, all

31:10 you have to do is ask Claude, hey, what type of agent can you build me for XYZ? Yeah, point to

31:14 this data set. So when you do have the data models set up and everything ready to go, that's when,

31:23 okay, yeah, we can line up all the sales agents, the marketing agents, CS agents,

31:28 all the rev ops agents that we need to spin up. You can do things like forecasting. You can count

31:34 on the forecasting analysis it's providing. One of the biggest things, so Jake and I, before we

31:41 found our way back at LeanScale, we worked at a company called Emailage. And I, my first half at

31:49 that company, I actually ran customer success. And I had 40 CSMs all around the world. We had

31:57 hundreds and hundreds of customers. And it was almost impossible to get what every head of CS

32:05 wants customer health. And it's because you're dealing with so much qualitative data, you're

32:10 dealing with so much data, and there's so much analysis and sentiment to try to read that is

32:14 nearly impossible. You'd need like a full data science team just to do it. But now if you get

32:19 your data model set up, you can get a very robust customer health model and something you can rely

32:25 on. It can help identify problems early on. It can help tee up companies that are likely to expand

32:31 so you can maximize those opportunities. But there's just so much you could do. And a lot of

32:35 these just won't take very long once you have a data model that can support it. If you're trying

32:40 to start with the agent first, inevitably, you're just going to have to go reverse engineer the data

32:45 model it needs to run, rather than, okay, why don't we just build this whole thing out that we can

32:51 then run any agent on. You'll be able to deploy agents in half a day, ones that are production

32:59 grade, you can share with your entire enterprise, they're scalable, reliable, trustworthy, where if

33:05 again, if you just keep going to let's build the agent first, it's going to have to manually

33:10 connect to all of the different sources of truth, you're going to be building hundreds of many data

33:14 models for every agent you want to spin up. And just to double down on that point, just real quick.

33:19 Yeah, I think it also democratizes the build. And for me as a CTO, that's so important, because for

33:26 a long time, it was like, Jake is the vibe coder of this company, right? Like, if we need to do

33:31 anything, it's got to run through Jake. But after we've built this out, and we've gotten everybody

33:36 connected to GitHub, and we have these design kits, Anthony's thrown in a ton of PRs. I mean,

33:42 he is cranking stuff out, like, if you would have told me six months ago, even that I'd be approving

33:46 Anthony's PRs, I would have told you that you were crazy. But here we are. And it's amazing.

33:52 It's awesome. And I know I can trust the output. I don't need to go back and reinvent the wheel

33:57 every time. And so for us, it's really democratize the build of these, where each individual person

34:02 can build their own agents or build their own automations with the confidence that it's

34:07 operating again on that same foundation. So just to double down on how important it's been to us,

34:13 me personally, if I'm being honest, but it can be so impactful across organizations.

34:19 It really has been impactful. And it has unlocked a lot of builders at LeanScale to be able to

34:26 build something, like you said, that's built on data that you can trust and the results it's

34:33 giving you you can trust and the way it's working. But also, then it's very easy to say, okay,

34:38 that's awesome. Let's now get that exact same thing into the hands of the entire company

34:42 so they can use it too. I think once you make the investment, you'll definitely see the value.

34:49 So I want to go through how to deploy it. Okay, how are you going to actually get there?

34:53 And Jake alluded to this earlier, literally half of this is an alignment problem. So the engineering

35:02 here is not that difficult. It's something that I'll walk through the types of roles you'd likely

35:08 want to lean on to get it done. There's a little bit of technical work to do, of course. But most

35:14 of this is, okay, can we get on the same page? Can we all agree on what the definitions mean? Can we

35:19 agree on what the source of truth is for all of these definitions? And then can we agree on the

35:24 person who owns what this definition is? And if you can do that, then okay, now let's codify this

35:31 and house this somewhere. And then let's make sure our data is connected meaningfully and clean it

35:36 up, deduplication, all that stuff. Then you can go have fun building agents. But I don't want to

35:42 diminish the fact that the hard part of this, like most things in life in general, is going to be

35:48 alignment and communication between the teams. So starting with the semantic layer, the way

35:54 you might want to think about it is, okay, every metric, every definition has to have an owner.

36:04 And on the sales side, we need definitions on what a qualified opportunity might be.

36:09 On the marketing side, oh my goodness, I can't tell you how many different definitions of a

36:14 marketing qualified lead I have seen. You need to figure it out for your company. So write it down,

36:20 agree on it, and then just stay consistent with what an MQL is in your company. Customer success.

36:29 I've run into this so many times too, which would probably be shocking to most people. What is a

36:33 customer? And how do I know? Is it this field that says customer on Salesforce? Is it they have an

36:40 account and our ERP? Is it that they've had an invoice? Do we count pilot accounts as customers?

36:46 Do we count customers that have put in a churn notice still as customers? Or do we wait until

36:51 the churn time comes? What is a customer here at this company? You'd be surprised at how many

36:56 different definitions you can get on that. And then there's all kinds of company-wide ones like

37:00 what's annual recurring revenue? For those who are listening, most of your company is valued on ARR.

37:07 That's just the way people value companies, especially early on. So what is the definition?

37:15 How do you articulate what your ARR is? Especially if you have usage-based products and a usage-based

37:20 service too. So anyway, getting this all on paper, codified, and then stored in one spot,

37:26 super important. Now I'll kick it to Jake to go through quickly, just technically, like okay,

37:34 where does this live? But again, most of it's going to be that alignment. Totally, totally.

37:41 And just to double down there, this isn't a new problem. This isn't something that AI invented

37:47 with this lack of definitions around an organization. As you alluded to earlier at emailage,

37:51 having 40 CSMs around the world, if we didn't onboard them properly and we just threw them

37:56 into the wild, do you think they'd be successful? Probably not. So this is just really highlighting

38:02 the importance of getting this stuff on paper because again, you don't want to be Adam Sandler

38:07 in 51st states playing the same video every morning for your AI and trying to tell them exactly what

38:13 bookings means to you. So not a new problem, just AI has definitely accentuated the need for it

38:18 across all these different business units. So the semantic layer, I promise I'm not going to make

38:24 you look at too much code here. Sorry if you just threw up in your mouth a little bit. This is

38:29 YAML on the right and then just a simple folder structure on the left. So as I talked about

38:33 earlier, we use GitHub for this. It's very native to all these LLMs. There's nice clean connectors.

38:39 It's free to set up, free to use. You can house things privately. You can share them amongst the

38:44 team. Very easy to configure, very easy to add automations around as well. So we're big fans of

38:49 using GitHub for this, but you could even use Google Drive with the correct folder structure

38:54 in just a shared drive format. So on the left is just that exact folder structure. It's the

38:59 semantic layer first, the different models that we have, all of the different entries and metrics

39:04 that we would want to capture. And one example of a metric is on the right there. So this is

39:09 the YAML file. You'll see the metric at the very top. In this case, we're looking at win rate.

39:13 The next line is the owner. And this is so important. We got a great question when we

39:19 presented this live of how we think about who should be the owner. Ultimately, we're going to

39:24 talk about some of the roles that are involved in this as well. But RevOps is kind of becoming that

39:30 pioneer of change when it comes to AI and being kind of the first AI enabled team,

39:36 making sure that we're rolling it out responsibly. But ultimately, your owner per metric, and I go

39:41 back to this time and time again, is who's going to have to speak to it in a board meeting? Who

39:48 is the owner of what that metric means? Because if you've ever walked into a meeting and tried to

39:52 defend a number that you don't have context on, that is not a fun razzle dazzle game. Let me

39:57 tell you, I've done it a couple times. I'll never do it again. It is not fun. You start sweating.

40:03 So whoever is going to be speaking to that number should have full confidence in exactly how we

40:07 defined it, how we got there, what it means to us, and be able to back that up. They also approve

40:13 any changes. So like I said, and Anthony alluded to, there's tons of different ways to qualify MQLs.

40:19 You can change that quarter over quarter. If you say, "Hey, we were over-qualifying a little bit.

40:23 We want to tweak it down." Perfect. Submit a PR. Because the owner is in that metric defined in

40:29 the code, they're going to get a notification to let them know, "Hey, so-and-so just submitted a

40:34 new PR. They want to update the win rate metric. What do you think of this?" They can pick and choose

40:39 and say, "Yeah, yeah, yeah. I really like that new lead scoring model." Or, "No, no, no. I don't want

40:43 to change. We always want to MQL hand raisers." Whatever that looks like, they have a chance to

40:49 interject and then stay in the loop as well. Because as we've alluded to, they're going to

40:53 be speaking to these metrics. So super important that you just start writing these down. I know

40:58 the code makes it look a little bit aggressive, but really you just need to start documenting this.

41:06 Okay. The context graph. I'm not going to spend a ton of time talking about this one.

41:11 Really, what I want you to understand is this is just connecting the dots in your CRM. But as

41:17 important as it is to connect the dots, it's also just as important to say, "What are we not going

41:22 to throw into this mix? What are we not going to tie together because we don't want that insight

41:26 or it doesn't belong in this individual context graph?" So I think taking some time to really look

41:32 at those associations, understand where you need additional, where the gaps exist today,

41:37 and then where maybe things are overinflated as you click into some of those calls and you see

41:42 internal calls being associated to a company record because they mentioned their name,

41:47 those types of things that need to be flattened out and fixed

41:50 before you have that full confidence in your CRM as your first context graph.

41:58 So I can hop in here. There's a handful of roles. The ones I really want to focus on,

42:03 you have to have an executive sponsor to push this through. There needs to be some budget

42:07 put here. You're going to need some tools. You're going to have to carve out people's time to make

42:10 sure people know that this is important. So whoever that is, typically it's going to be your CRO,

42:14 maybe even your CEO. Like Jake mentioned, RevOps is going to be a huge person to lean on. They are

42:20 already owning a lot of the business context and I think they are a perfect linchpin between the

42:28 teams to pull all this together. Now, if you have an analytics engineer and platform engineer on the

42:34 team, I think that would be fantastic as well. And then whoever's running your CRM. So if somebody's

42:38 running your CRM, that's a person that you're going to want in the room when you're making

42:42 these decisions. If we were to space it out, everything's going to look a little different.

42:47 What does your data environment look like, et cetera. But in general, you should be able to get

42:51 this done in about 10 weeks. And if you notice, more than half of it is mainly alignment on

42:58 definition, semantic layer tools you're going to use. And then it moves into the engineering side

43:03 and the agent building side. But this is a 10-week roadmap that you can copy for your organization.

43:08 And I think Taylor for your needs. A couple of things, ways this could go wrong. I don't

43:13 want to harp on all these too much, but I think the biggest one is not having a single owner for

43:18 definition. That's really where the wheels fall off on this. So, hey, your VP of sales

43:22 owns what a qualified opportunity is. Your VP of marketing owns what an MQL is. I wouldn't have

43:28 RevOps own many of these. I think most of these will live with a functional leader. And then one

43:36 that we highlighted starting with the agent building. It's like, yes, you can try to build

43:40 all the agents you want, but at the end of the day, they all need data to access. So,

43:44 sequencing this is really important. And then just a reminder, hey, if you do this right,

43:50 if you invest in the data model, you build the agents and you start to shave off a few days of

43:55 the sales cycle, improve conversion by a couple points. These results continue to compound. And

44:02 this is where you make the investment back. Just to strip the jargon and summarize everything.

44:07 Semantic layer, write down your definitions, find a good place to store it that you can point your

44:12 AI to. Context graph, connect the dots of your data between systems, even within systems.

44:19 Just make sure the AI knows when you're asking a question and it goes to a particular person

44:23 or account, it knows what other data is connected to that. And if you think about it that way and

44:30 make it pretty simple, I think you'll go pretty far. And then the last thing, this giveaway,

44:34 this is all for you. So in here, you're going to get this presentation so you can download it,

44:38 PDF PowerPoint, convert it to Markdown and throw it into cloud, whatever you want to do,

44:43 you can have it. You'll also get the metric definition worksheet that you can work with

44:49 your team. This should give you a head start. We have 10 pre-built agents that you can download.

44:54 Now, none of them are going to be that effective unless you invest some time in your data model,

44:58 but you're going to have them. And then also a way to stay connected with LeanScale. We actually

45:06 put everything into our MCP. So if you go to leanscale.team, you can get connected to that.

45:12 We have our playbooks, we have benchmarks, we have studies, we have podcast transcripts,

45:19 if you want to know what CROs and CMOs are actually talking about. And you'll have all

45:24 of the materials from here. And then if you're interested, we can also offer up a diagnostic

45:31 to take a look under your hood and see where we can help. So I hope this was helpful. Jake,

45:37 thanks for being my wingman as always. And I hope this helps whoever's listening,

45:42 invest what they need to into getting their data model ready to get the most out of their AI.