Victor Cheng on How to Teach AI Your Company

Julian Castelli (00:01.79)
Hello, this is Julian Costelli and sp and welcome to a special episode of Growth Elevated AI Sherpa podcast, where we talk with leaders in the tech industry, and on this track we talk about people people who are innovating with AI or helping companies innovate with AI. Past guests have included CEOs and CXOs of great companies like Workfront, CHG Healthcare, Pathology Watch, In Moment, Canopy, the San Francisco 49ers, and many more.

This episode is brought to you by Growth Elevated. We are a community of tech founders, CEOs, and CXOs who are committed to working together to share best practices and learnings in an effort to help all of us become better leaders. We do that through educational programs like this podcast, as well as our blog, and of course our annual Ski and Tech Summit, where we bring tech leaders to beautiful Park City, Utah from all over the world to talk about tech and enjoy some fun on the slopes. So if that sounds interested to you, please check us out at Growthelevated.com.

Today I’m excited to welcome Victor Cheng to the podcast. Victor is an accomplished executive and the CEO of SASCEO.com. Victor works with tech companies to drive value creation and is current s currently working with a lot of them on how they’re working with AI. In fact, he’s even writing a book about it, which I’m really excited to explore in this podcast. The book is called Teach AI to Your Company. Victor, thank you for joining the podcast.

Victor Cheng (01:29.88)
Thanks, Julian. I’m excited to finally get this on the books and have it. Conversation.

Julian Castelli (01:34.16)
Yeah, me too. So tell us a little bit about about your your practice and how you got involved into helping companies. Like w w give us your your career journey in in a in a nutshell, please.

Victor Cheng (01:43.223)
Yeah, so I work with tech founders, CEOs, and boards to help them scale and prepare for institutional capital exit. been doing that sort of in various capacities the last 20 years. And then AI happened. so the goal is still the same, but the means for getting there has changed quite dramatically in the last year or two. so looking forward to discussing that further, because I know you have a big interest in that as well.

Julian Castelli (01:55.162)
Yeah.

Julian Castelli (02:00.604)
So I’m looking for

Julian Castelli (02:05.017)
And and how do you you know, w be let’s let’s let’s understand before AI, your typical engagement would be with what type of company and how how would you help them?

Victor Cheng (02:13.846)
Yeah. So most of my companies I’ve worked with historically, probably three to twenty-five million in ARR, maybe on the high end up to maybe sixty million. And most of them are at the stage of, you know, beyond sort of just founder-led businesses, they’re having starting aspects of an executive team, and they’re running into the friction points of getting beyond 10 million typically. So it’s run managing through others, it’s having protocols, having scalability. you start getting into balance sheet activities in terms of raising debt and equity. and the business gets more complicated. You have boards, you sometimes you have investors.

Julian Castelli (02:22.437)
Most of them at the stage of Alpions were just founder-led businesses that having starting aspects and executive team and then the first of game.

But if you want to manage the others that’s cabinet protocols and scalability, you start getting to balance your activities in terms of basic debt and equity, and the business is more complicated, you have to work so sometimes you have an investor. And anytime you prepare for institutional complexity, which has a different working just building a company with a few customers in a new product. that’s kind of a historically, and you know that’s the goal. just the way you get there is a different but the analysis. Yeah, and

Victor Cheng (02:43.76)
and really trying to be prepared for institutional capital exit, which has different requirements than just building a company with happy customers and a good product. So that’s kind of what’s been historically, and you know that’s still the goal. Just the way you get there is a bit different with AI in the mix.

Julian Castelli (03:03.16)
So how have your engagements been changing now with in the with the world of AI?

Victor Cheng (03:07.33)
Yeah, so a AI is both very, very exciting and very, very threatening. And so in the morning it’s exciting in the afternoon by the afternoon it’s like threatening, right? so it ends up being both. It changes every week, sometimes within the week. and I think both are very true. All right. If you ignore it, I think that’s a threat. If you get ahead of it and take advantage of it, it could be really, really exciting, what the potential could be. And I think both are possible outcomes.

Julian Castelli (03:09.604)
Well, yeah, everybody.

Yeah.

Julian Castelli (03:18.444)
And it changes it every week, right?

Julian Castelli (03:23.002)
I think both very true. I it could be more equitable. I think that’s threat. If you get ahead of it, if you take advantage of it, it could be really exciting what that potentially would be. And I think both are possible. Yeah, a hundred percent. And and so Victor, you you joined us at Growth Elevated and and that’s a an event where we bring together a lot of founders and and executive teams from growth stage companies. So that the you know, y your audience and and and my audience, the people we work with and our peers.

And you shared some great things in the in a presentation about AI. How many of your engagements these days are AI focused versus just just overall value creation?

Victor Cheng (04:05.664)
Yeah, i it’s it’s always value creation. I think AI probably comes up probably 80% of the time in some way, shape, or form. I would say very intensively, maybe twenty percent of the time. I’ve been trying to get more people to be more proactive around AI. And and surprisingly, I don’t know if you’ve had this experience, but I found the leaders of sort of tech startups surprisingly slow at adopting AI themselves as well as within their companies, beyond just using like like a chat Gibbia instead of Google.

Julian Castelli (04:17.506)
And try to get more people to be more active on AM thinking.

I’ve found things with sort of extra system. I think slow adopting it themselves well as our companies. I’m just using like a picture GP to consider Google, but like workflows, but in the case of content work, I find the adoption not so licensing slow. Yeah, I think it is, and I think part of the reason is that it’s so there’s so much to do, right? If you’re running a startup.

Victor Cheng (04:35.568)
But like really using automated workflows with in some cases autonomous work, I find the adoption surprisingly slow.

Victor Cheng (04:49.71)
Yes.

Julian Castelli (04:52.785)
And I you know, I I I talk to a lot of lot lot of companies that are you know, of course you should be using AI, but they also have to, you know, hit their numbers and they have to, you know, hire their head of sales or head of technology and and there’s so many different things they’ve gotta do. It’s it’s it’s challenging, right?

Victor Cheng (05:10.262)
Absolutely. And and I think it’s you know, I’ve I’ve been sort of heads down in AI for for quite some time and I think it is even more challenging than I sort of realized in in terms of getting it right. And the benefits are incredible after you go through the level of effort. So the costs are higher and the rewards are higher. It’s both, which I thought was sort of interesting and I wasn’t expecting kind of early on.

Julian Castelli (05:18.934)
Isn’t it?

Julian Castelli (05:23.19)
In terms of getting it right, the benefits are incredible after you go through the level So the costs are higher on higher cases, which I thought was sort of interesting people think exactly that really fine. So how do you how do you get the CEOs to prioritize AI and and actually start start having tangible efforts that that they can they can pr produce ROI? And how do you get them to get through that clutter and that challenge of, you know, I’ve got so many other things.

I I use the analogy of like, hey, you y you know, stop running with a bicycle, right? Because you’re you’re carrying this around like a burden and you you feel like you don’t have time to stop and r learn how to ride. But if you did, you’d go so much faster. How do how do you get them to overcome that challenge?

Victor Cheng (05:54.627)
Mm-hmm.

Victor Cheng (06:04.352)
Yeah, so I I to to be honest, I I’ve I’ve I’ve I have CEOs sort of in two groups. one, they’ve embraced it and they want to do it and they’re just not sure how. And the others they’re just overwhelmed with regular life, you know, sort of business as usual. and to be honest, I’ve not yet successfully convinced one in that second category to move over to the first, which was part of the motivation for writing the book, is to to help people understand what’s really involved.

Julian Castelli (06:10.661)
One category said they want to do it and they’re just not sure how. And the others are just overwhelmed away, or like yeah, sort of business as usual. and to be honest, I’ve not yet successfully convinced when that second category to move over the first, which was part of the motivation quite well, is to help people understand industry involved, umistic view of what it takes to just stack and do

Victor Cheng (06:30.306)
Get a realistic view of what it takes in terms of staffing and resources so that it’s demystified and you can make a well-informed choice as to whether you want to do it or more likely when do you want to do it. but I think right now the biggest constraint in AI adoption is not the AI models in themselves. it I would say right now, really it’s psychology. It’s human psychology around. Absolutely. 100%. Yeah.

Julian Castelli (06:37.588)
I think right now the biggest constraint and adoption is not the AI models that also it I would say right now really if psychology is Yeah, it’s change management, right? Yeah, that’s what I’m seeing as well. And and you know, you you walked us through a great framework with your AI playbook and I just pulled it up.

And so I’m going invite you to you know walk us through some of the the highlights of that. But before we do that, you know, the the the the change management is getting people to change the way they do business. And and it’s about education, it’s about change management, it’s about prioritization, it’s about, you know, a lot of things that are, you know, back to our McKinsey days, you know, are are are more just fundamental business leadership and and organizational behavior things than what model to use or or other things, right?

Victor Cheng (07:20.91)
Mm-hmm.

Victor Cheng (07:32.171)
Absolutely. Absolutely. Yeah. I it’s it’s it’s funny. It’s the more I learned about AI and the more I realized everything else I learned before AI is incredibly relevant. so it’s

Julian Castelli (07:44.838)
That’s kind of a relief though for for for guys like us, isn’t it? ‘Cause because I’m never gonna be the most I’m never gonna be the most

Victor Cheng (07:46.063)
It is, it is. But like I I look right back where I started, really? I I didn’t know that. I thought maybe my you know, I you know, I I’m sort of in my you know, at the end of my career, right, later stages. I’m not the twenty four year old wonder kid anymore. I’m now the the historian on the board. I can talk about tech errors that were that was that existed before people were born in some cases. but it all kind of comes back. Like the basics absolutely still apply, just with a twist.

Julian Castelli (08:08.199)
Right. Okay.

Julian Castelli (08:18.17)
Yeah. I I had the same reaction, which is like, okay, some of these basic skills of blocking and tackling are just as important or maybe more important even than what the latest model is or or how to vibe code, you know, on on six different platforms in parallel. those things are important, but I my I also get a sense that those are gonna keep changing. And so some of the more durable business business principles are are just as important.

Victor Cheng (08:37.069)
Yes.

Victor Cheng (08:43.606)
Yeah, and and I think it’s interesting when when I was writing the book I was trying I was being very deliberate to make sure the book was still relevant ten years later, you know, or even fifteen years later. So sort of first order principles as opposed to this model versus last model.

Julian Castelli (08:51.291)
Right.

Julian Castelli (08:56.719)
Yeah. And because you know how stream is like really important I I can think about. but I think that big principles are the same. And you know, one of the things I talked about in the sense of my clipbook and is the

Victor Cheng (08:57.254)
And as you know, models change like literally every four weeks. I I can’t even keep up. but I think the base principles are the same. And you know, one of the things I talk about in in the SaaS AI playbook and and then in the book is the just understanding how models really work, because I think there’s actually quite a lot of confusion around that. And if you don’t understand that foundational piece, then everything that follows just doesn’t make sense, or you end up making, you know, making the wrong

Julian Castelli (09:18.561)
If you don’t like the back down the HP, then everything that follows just doesn’t make sense to the end of being.

Victor Cheng (09:27.088)
Yeah.

Right. And and I think a a big key principle really is that, you know, LLM models specifically are are just prediction engines. That j it’s just math. Right. So if they’re basically taking all the publicly available knowledge that’s in the written form and other forms, think of it as the entire internet, if you would, which is pretty good proxy. And they have mathematically mapped every word that exists on the internet and how likely is it to appear next to a different word, given the words that have appeared before it. So

Julian Castelli (09:29.794)
I think a big key principle really is that you know L models specifically in our prediction engines that just have right so if you know basically taking all the public individuals not once it’s written down and other forms I think of it as connected, but which is pretty good proxy. And they have mathematically have to every word that exists on the internet, and how likely is it to appear next to a different word, even the word that appeared.

Victor Cheng (09:58.215)
They’ve looked at every every document in the internet around the word dog, right? And if you say hey, what percentage of time when someone says watch the dog, and then there’s a blank, right? And they have this math algorithm that says, well, when you say watch the dog, 60% of the time the word is run, right? maybe 10% of the time it’s jump, maybe 7% of the time it’s like eat. And

Julian Castelli (09:58.522)
So they look at every every document in the internet about the word dog. If you say what percent of the time when someone says watch the dog as they have to math algorithm says, well, when you say watch the dog, sixty percent of the time the word is run, right? ten percent of the time it’s jump, it’s a feature. And if we do that.

Victor Cheng (10:23.774)
Basically doing that and with every additional word that you either have it write or you provide an LM model like such as Chat GPT, every additional word it recapits all the math. So if I said, hey Julian, watch the dog blank, right?

Julian Castelli (10:25.679)
And with every additional word that you either have the right or you provide an alcoholic, such as ChatGPT, every additional word can recap into all the math. Wow. So I said, hey Julian, watch the dog playing. And maybe I’ve had a conversation about people. I know maybe you have dogs, I don’t know if you do or not. I do. Right? You know, that’s an internet thing, it would change the probability you have that next word.

Victor Cheng (10:40.198)
And maybe I’ve had a conversation about you. I know maybe you have dogs. I don’t know if you do or not. And we have a particular kind of dog, right? And if that’s in the internet, then it would change the probability of that next word. Watch the dog swim or watch the dog ski, right? Because you live in a city. So that’s why like there’s such massive computational need to do all those calculations. It’s insane. It’s cute.

Julian Castelli (10:52.695)
Watch the dogs swim, watch the dogs sleep, right? So that’s why like there’s such a massive computation. Yeah, that’s that is that is hugely computer intense or ca calculation intense.

Victor Cheng (11:06.88)
Yeah. And I think the big insight people miss is that the AI oftentimes is disappointing, not because the model doesn’t work, but because of the information you’ve provided to the model is insufficient for the model to do the kind of work that you would expect intuitively. And I think that’s the part where people make the misunderstand significantly.

Julian Castelli (11:14.19)
Because

Mm-hmm.

Yeah.

Julian Castelli (11:29.708)
I I think that’s my biggest conclusion at this stage in the AI evolution, which is that it’s so important about what you feed or what you enable the model with specific context. And I know that’s we’re changing to local models and and and you know, creating company brains. And I I spend my time, so much of my time downloading PDFs and putting them into the specific projects. And in some regards, that you know, first of all, that’s frustrating and tiring, but

Victor Cheng (11:53.186)
Yes.

Julian Castelli (11:57.781)
On the positive side, it kind of trains me as to how these things work, which is no different in the way that you educate humans, right? You know, you put them in places where they can they can get count a lot of context and then then they could they can start doing pattern matching that become more useful through through study and experience and that sort of thing. And so the mechanical nature of doing this, who knows? You know, I I’m sure in the future they’ll fix this and it’s gonna happen much more automated and and and I think scalable.

But maybe this this awkward teenage stage is useful because it’s teaching us a little bit about how the models work and that’s gonna be valuable as as we you know bring them on as partners, right?

Victor Cheng (12:37.923)
Absolutely, right. And then you know, sometimes founders will ask, Hey, what can I do today that’s sort of

within the level of resour level of resources I have, that sets me up as well for the future, right? So basically how do I crawl today in a in a way that helps me setting up setting up for walking and running and then flying with AI. And and probably like the number one thing, and and Julian, you already you’ve already talked about this sort of in what you do, but I would I want to sort of really formalize it is take all this what I call implicit knowledge or tacit knowledge that you have inside of a company inside of your head and just get it down in some kind of documented form.

Julian Castelli (12:44.084)
Again, the melod resource level resources like that’s that’s the call for the future, right? So basically that way called today companies that want.

Julian Castelli (12:57.928)
You’ve already talked about this kind of sort of what you do, but I would I want to perfectly formalize it is take this what I call implicit knowledge or tacit knowledge that you have inside what happens inside of your head and just get it down as we kind of document. Yes. I do a lot of reportings as being recorded. Yep. Right. MP3 files or MP4 can be transcribed. Sony Intergrade or Microsoft those can be transcribed.

Victor Cheng (13:13.338)
So for you it’s PDF files. For me, it’s MD files. I do a lot of recordings. This is being recorded, right? MP3 files or MP4 can be transcribed. Zoom meetings are great, or Microsoft Team meetings, those can be transcribed. Like just getting, like what do you and your customers talk about, right? Like we you know that. I know that. Absolutely, absolutely. Yes.

Julian Castelli (13:27.359)
Like this gay. But what can your customers talk about? Yeah. Well, that’s the context. That’s the context, right? And it it and it mirrors how we get up to speed and how we’re able to help companies. But you know, you have to replicate that in painstakingly difficult and laborious E ETL activities, right? Extra extract, transport, load.

Victor Cheng (13:46.211)
Yeah.

Victor Cheng (13:50.849)
A absolutely. So I think if you just I did that maybe in two thousand twenty-four. My staff hated me. I’m like, hey, everything I’ve ever said to in a presentation, you know, on a webinar, like on a stage, and and like I want to get all those recordings in one place and transcribe them. And so they’re going back to compact discs, right? From like like two thousand like nine. All right. We had a couple CDs of some speeches, and so you know, my staff, they were like, you know, digitizing those off of C D, right? And then unfortunately.

Julian Castelli (14:04.415)
And like I’m like all those recordings in one place and transparent. So they’re going to have to compact this. Wow. Like like 2009. Wow. And so you know my scale would like you know digitize and develop a C D, right? Unfortunately, as case of say some people here listening don’t be put CD. Yeah. it’s a physical device.

Victor Cheng (14:20.792)
Unfortunately, I hate to say it, some people here listening don’t even know what a CD is. It’s a audio recording device that predated MP3 files, right? as a physical device. And that was just months of work. Get all that stuff from the early 2000s transcribed. And I think by the time we were done, probably four million words.

Julian Castelli (14:32.398)
And that was months of work. Sure. I think five million words like in PDF files, which is the weapon of one of like 130 paperback books of things I had in the sure.

Victor Cheng (14:42.954)
Like in PDF files, which is the equivalent of probably like 130 paperback books of just things I’ve said and speeches I’ve given and things I’ve written. And so that was, you know, I understood the math back then, I didn’t really appreciate the formulacations. But that’s been hugely impactful for at least our own AI efforts, because now we can do things with that context.

Julian Castelli (14:53.225)
And so that was, you know, at least in the math back then, I didn’t we have patients with phone with patients. but that’s been huge, be impactful for at least our own AI efforts, because now we can do things with that context. Right. I can give you my model, I can give you all my code, right? But if you don’t have contacts, it’s going to be useless to you. That’s right. Well, Victor, why don’t you take us through the AI playbook? I wanted to put it.

Victor Cheng (15:06.966)
Others can do. I can give you my model, I can give you all my code, right? But if you don’t have context, it’s completely useless to you. But if you do have the context, it’s hugely, hugely valuable.

Julian Castelli (15:20.904)
One of the things I wanted to do to to your point, you gave a great talk. I know everyone was excited about it afterwards and you know and then then life happens and you know you you can’t go back to that. One of the things I wanted to do was capture a summary of your presentation here on the podcast so people can come back and look at it and find it at the Growth Elevated website. So why don’t you walk us through the SAS playbook here? I’ll I’ll turn the pages at your direction and you know, walk us through you know, your your playbook, please.

Victor Cheng (15:48.131)
Yeah, no problem. Let’s let me start with the the agenda slide, slide two.

Just to give you kind of an overview of the playbook. There’s sort of like the big picture concept, and we’ve sort talked a little bit about that so far. I’ll fill in a couple more blanks. and then I really think about AI sort of in two ways. One is AI in the products a tech company offers to customers. and this also increasingly includes technology-enabled services, right? So a lot of services businesses are using a lot of AI, so this concept equally applies. So think law firms, right? A lot of AI-native law firms are, I would lump in this category. and then the third,

Julian Castelli (15:53.411)
There’s sort of like the big picture of Tom sort of talk a little bit.

Julian Castelli (16:01.46)
think about AI sort of in two ways. One is AI the products that tech company offers to customers. And it’s also the most needing to technology table services. A lot of services businesses are using a lot of AI so this concept equally applies. So think law firms, right? A lot of AI data law firms are highly category. And they’re not very fixable AI for under functional operations. So marketing tail

Victor Cheng (16:22.414)
Area is really around AI for running your functional operations. So marketing, sales, you know, human resources, the functional areas, and using AI to be more scalable in those areas. so if we jump ahead a little bit, we’ve talked a little bit about that big picture, which I won’t labor over. And if we can jump ahead to let’s go to slide

Julian Castelli (16:31.631)
function here is using the cable. but we don’t have a little bit we talked about the big picture which I won’t linger over. And let’s go to slide hang on a second, Victor. I I’m afraid

Victor Cheng (16:47.231)
yeah.

Julian Castelli (16:50.415)
I’m gonna we’re gonna have to I’m gonna clap. We’re gonna have to stop and restart because my studio is telling me it stopped recording. Give me give me one second. I’m so sorry. But we we we’ll we’ll edit this together because we had a great opening here.

Victor Cheng (16:58.018)
okay, no problem.

Victor Cheng (17:01.934)
Okay,

Victor Cheng (17:27.318)
Angel, while while we’re waiting for that to start up, how much time do we have? I I wasn’t sure if I should go through the whole slide presentation or I can kind of highlight

Julian Castelli (17:32.941)
I think we should kinda cover the cover the the the the high levels. We p we wanna keep the whole thing within forty five minutes if we can.

Victor Cheng (17:37.752)
Perfect. Okay. I’ll do that. Got it. Okay, great. I can definitely do that.

Julian Castelli (18:24.207)
Okay. Victor, I think we’re back. Can you hear me?

Victor Cheng (18:27.292)
I can.

Julian Castelli (18:29.157)
Okay, I apologize. But let okay, Victor, so let’s dive into the playbook. Why don’t you kinda hit the high levels? We’re gonna put this up on the Girl Televised website and then this would be hopefully be a a talk track that could give people a little bit of a guide as to how to get get through that and you and use it and and how to find you, of course. Nope, no section to start on slide number 18. just talking about the I inside of blocks. And I have this you know it’ll it’ll recognize.

Victor Cheng (18:44.288)
Nope, no problem. Yeah. So if we look at the playbook section two, which starts on slide number eighteen, it’s talking about AI inside of products. And I have this you know, you’ll you’ll recognize this, Julian, two by two matrix, right? on slide nineteen. really useful. how to think about using AI in your products and services increasingly that you offer to customers and clients. And you can think about it of in terms of how much autonomy you you have AI

Julian Castelli (18:57.948)
Of course. But I think about using AI in your products and sort of the CPSD that you offer to your customers in the client. And think about it in terms of interspersion, you can have AI performance terms of its work tasks on the supervision work with a lot of tech, for example. And then how much work we allow them to do sort of function. So on slide 20.

Victor Cheng (19:14.312)
perform in terms of its work tasks, how much supervision it is, or does it work completely on autopilot, for example? And then how much work do you allow it to do sort of in one shot? And so on slide 20, you have this sort of progression of AI autonomy, right? So I would say most people are on this first category, like my mom’s in this first category, AI assist, right? It helps a human being do the work. So it is using AI like a chat GPT or a chat bot.

Julian Castelli (19:27.403)
have the sort of progression of GI atomic like so i would say most people on this first category have mom to this first category okay it helps the human being do the work so it is using the i like a chat or a chatbot instead of asking questions what has the best one open how do we get in those kinds of things and the key thing is this is that the human being is still responsible for making the code

Victor Cheng (19:44.372)
instead of Google asking its questions, what time is the restaurant own open? How do I get from A to B? Those kinds of things. and the key thing in assist is that the human being is still responsible for doing the work.

Julian Castelli (19:57.177)
Yeah, it’s really basically basically accelerating retrieval, information retrieval, right?

Victor Cheng (20:00.747)
Absolutely. Yeah. Which is useful, right? It saves time. It’s a it’s a better version of Google. automate sort of takes a step up where the AI does some very task-level work, five seconds, 10 seconds, 15 seconds level amount of work. And then the human being’s role be doesn’t go from being the doer of the work to the approver. So best example is human in the loop.

Julian Castelli (20:05.283)
Yep.

Julian Castelli (20:16.123)
Yes. So human in the loop, right?

Victor Cheng (20:25.09)
human human human at the task level loop specifically. So it auto-suggests a reply to an email, right? So a customer service agent, AI drafts the reply to the customer service ticket. You AI the human being looks at go, yeah, sounds right, you know, click send, right? That’s kind of classic autonomy. And I would say most AI products are kind of, you know, in that sort of zone.

Julian Castelli (20:30.254)
Okay.

service agent, AI draft the replies that comes from a service ticket, you can AI that looks and go, yeah, sounds right. No, it’s that’s kind of classic automatic. Yes. AI products are kind of you know in that sort of zone. Right. Then we get into orchestrate where it gets much more advanced. It’s been one five second task. It’s doing three, four, five, six, seven, or eight, five to ten, fifteen second tasks.

Victor Cheng (20:48.142)
then we get into orchestrate, where it gets much more advanced, where the AI is isn’t doing one five-second task. It’s doing three, four, five, six, seven, or eight, five to ten, fifteen-second tasks all combined together and supervising a chain of activities.

Julian Castelli (21:03.44)
all combined together and supervise the HTTP activity. And are those different agents doing that typically?

Victor Cheng (21:08.722)
It it can be. It can be s orchestrated a couple different ways. and this will probably change over time technically. so it can be one, usually it is one main agent that’s sort of overseeing the parts.

And the subparts could be a computer code, like give me a I want a sales tax calculator, right? You don’t need AI to do that, right? You you have a computer program that can do that. But then I want to put the sales tax onto a proposal for a bid, right? That’s where you might be doing AI. So the human being becomes more of the supervisor. They’re either reviewing every step of the work along the way, right? All six steps, let’s say, or they’re reviewing the the output at the end.

Julian Castelli (21:22.649)
So the he becomes more of a supervisor, they’re I get the reviewing every step of the work along the way, right? All six steps, I’d say, or they’re reviewing the output. So uncontrol.

Victor Cheng (21:49.601)
So still full in control, just they’re not hands on keyboard anymore. they’re just watching the work being done. you know, my my kids call it, Dad, are you babysitting AI again? Right. I’m just sitting there and drinking my coffee. I’m working, honey. I’m kids, I’m working. Don’t bother me, right? They don’t believe me. because it doesn’t look like work. But but it is because you need that expertise to go, that’s right, that’s right. Ooh.

Julian Castelli (21:52.36)
They’re not hands on keyboard anymore. They’re just watching the work being done. my my kids call it are you babysitting AI again? Right? Yeah, right. ‘Cause that’s that’s a yeah. You’re kinda waiting. Yeah.

Yeah. You have to be following along what’s going on to be the you know, the the quality control.

Victor Cheng (22:15.114)
Absolutely, you know, and you bring up a great point, Julian. I think one of the key skills, which we’ll talk more about this later, is the quality control skill. And so AI is the first technology, I think, where people like me and you with a lot of domain expertise are more capable of using this new technology than someone that’s maybe born with it. So when social media came out in in you know

Julian Castelli (22:17.152)
I think one of the key skills, which we’ll talk more about this later, is the quality control skill. So AI is the first technology, I think, where people like with a lot of jumping expertise are more capable of using the SP technology and some of the SDP1 is X. So

Victor Cheng (22:41.45)
TikTok, my kids were all over that like years before I got in there, right? And they’re teaching me all those things. And I’m outpacing my kids and everyone, you know, sort of their age on AI because I have all this domain expertise not related to AI at all. And suddenly so I can orchestrate and go, Yeah, that’s right. that’s wrong. that’s that’s not bad, but it has some adjustments and you can do the same thing. So

Julian Castelli (22:42.007)
My kids were all over now like years before got there. And they’re teaching me all those things. And I am outpacing my kids, and everyone sort of their age, I can get eyeballs for me expertise. Not physically at all. Yeah. And suddenly so I can work and straight and go, Yeah, that’s right. Ooh, that’s wrong. That’s that’s not bad, but that’s the key. Yes. It’s so critical, and it gets scary when you watch, you know, my son in college. He’ll start just coding something he doesn’t know the top the basics about. Like, you really gotta learn.

how the f how the process is performed today before you can do it with AI, otherwise you can’t quality control it.

Victor Cheng (23:17.728)
Absolutely. So that’s true with software development, it’s true with financial models, it’s true across the board. And so you don’t people won’t sometimes wanna hire AI experts and and that and you do need some AI expertise, but you need like process expertise, right? How do you build a business plan, right? How do you do a board deck? how do you do a proposal that that works? That kinda

Julian Castelli (23:23.638)
Yeah.

Julian Castelli (23:28.95)
That’s right. How do you do the board to act like that? How do you look a proposal to that? Yeah, what’s a market map?

Victor Cheng (23:41.206)
I’m sorry, say again? Yeah, yes.

Julian Castelli (23:42.494)
What’s a market map? I you know, he was my my son was trying to have AI build the market map, but you know, and he he he he was halfway through the project. He said, Dad, what’s a market map? I said, it it might be important to understand that before you just take whatever AI gives you.

Victor Cheng (23:50.863)
Yes. Start with that first, right? Absolutely. Absolutely. yeah, then as you sort of progress, this is much more advanced. You have AI operate, right? Where AI oper AI is sort of doing 80% of the work on its own. And 20% of the time when it’s unsure, it kind of kicks it out as sort of exception handling. and the human being’s role becomes when the AI is unsure, I’ll take it away. Still on the axis of autonomy. Yeah.

Julian Castelli (24:06.097)
Yeah, on its own yeah.

kicks it out has to provide anything. and then counts. So this is still on the axis of autonomy, right? Going from low to high.

Victor Cheng (24:20.288)
Yeah, so that’s operate and then optimize is AI’s running the whole show. It’s self-healing any errors, it self-corrects, it’s self-optimizing, right? That’s kind of like the you know the the future, you know, the the future vision people see and get excited about. And then the humans’ role is really governing, like guide guardrails, policies, the rules, right? That’s kind of where the human role is. So that’s kind of the autonomy spectrum. and that’s one way to grow. The the other way to sort of make decisions within products is sort of on next slide is is scope.

Julian Castelli (24:28.166)
It’s self-optimizing. That’s kind of like the you know the the future, you know, the future vision you will see. Right. speed covering my extra fun easy extra. Okay. that’s one way to grow. The other way is to make decisions with products is on the next slide is is you know, how much work you like gate to whether you’re involved or not. So

Victor Cheng (24:50.533)
So how much work do you let it do whether you’re involved or not? So at the bottom of scope, the easiest task, right? Hey, what time does that restaurant close? Look it up for me. All right.

Julian Castelli (24:57.756)
I need to write notes, I gotta summarize my action items, I need to send a thing for Gmail. That’s the workflow. Yeah.

Victor Cheng (25:01.974)
Workflow would be, hey, I’m on a sales call. Let’s take, I need to write meeting notes. I got to summarize my action items. I need to send a thank you email. That’s a workflow. That’s an example in sales. And AI can do that if you want it to. It can run a process, which might be a multi-stage, multi-step piece of work, multiple workflows, from maybe managing a call to a discovery meeting, doing a demo, all the way through to contracting, right? Managing that whole process. And operating cadences, they’re kind of like running the business.

Julian Castelli (25:27.607)
Right managing the whole process of an operating case that are like running business in a lot of ways. We probably have weak staffing that’s I was my IT and AI stabbing money on case because it does not like any basis because it does a weekly order. and it is making self-direct loops in that process. And then the last level that I guess really far off there is it just manages the results. Yeah.

Victor Cheng (25:31.881)
In a lot of ways, right? Where, you know, you probably have weekly staff meetings as I do with my my team. And AI is having running on cadences. There’s things that it does on a daily basis, things it does on a weekly, the end of month, end of quarter. and it is making it has self-correcting loops in that process. And then the last level is really far off there, is it’s just managing to the results. Hey, I want sales to grow up to twenty three million, you all figure it out. I’ll I’m gonna go on vacation, right?

Julian Castelli (25:56.691)
Have you seen that in in in work is is that still theoretical or have you seen that in in in examples yet that are out there? I think

Victor Cheng (26:04.162)
That I think is still pretty theoretical. I definitely see lots and workflow in process. The my my better clients are definitely in process. Operating cadence really early stages. at least the companies I see. and and so I think the last two are a little bit further away.

Julian Castelli (26:12.124)
Yeah. it’s I see I think the last two are a little bit further away. Yeah. Yeah, I I I call it human in the loop and then full autonomy. And I think that you know, I I’ve got a a a six fra stage framework, and you know, e even Tesla’s been working for fifteen years now to try and get that there. It’s not quite there, you know, with self driving, but but it but they’re they’re they’re progressing, you know, steadily towards the goal.

Victor Cheng (26:33.996)
Yes. It’s not easy, right?

Victor Cheng (26:40.972)
Yeah, so I think of it there’s human in the loop and then there’s human on the loop and then fully autonomous is how I think about it. All right. So so in the loop is I I approve every decision sort of in the process. I’m quality control. On the loop is the AI approves it eighty ninety percent of time. On the high risk cases, I I jump in or it or it calls me in, right? So I’m still involved. Yeah.

Julian Castelli (26:47.569)
What’s what’s your what’s the difference in the loop versus on the loop?

Julian Castelli (26:53.497)
Okay. Home loop is the AI operators can’t eight nine high risk case of I jound the more so specific. So not not every time, but but when it’s when when the AI’s unsure. Yeah. I I I I agree with that.

Victor Cheng (27:08.951)
Yeah, yeah.

So one way to think about sort of you know product roadmaps, particularly in the world of AI, if you sort of skip ahead a couple of slides, I I think there’s sort of four key decisions, which is on slide 24, and regarding one’s AI product strategy.

Julian Castelli (27:21.456)
Yeah.

Julian Castelli (27:25.648)
So what’s the specific measurable result that you want to tell the curtain to customers with that you’re contemplating using? Yeah. Well, as the scope of AI, right? Can we have some digital right is it’s you know one of the competitors are getting right where to do that to step around level of small what you’re talking about and then timing and looking at this order versus next year versus hereafter.

Victor Cheng (27:27.882)
Right. So one is what’s the specific measurable result that you want to deliver to customers with AI that you’re contemplating using? What is the scope of AI right that you’re gonna allow it to do? Right. Is it sort of fully is it you know running the whole company to an EBADA target or I or is it doing a couple of tasks? There’s a range. the level of autonomy which you talked about, and then timing, like what do we do this quarter versus next year versus the year after? And so I think of this sort of chessboard on slide twenty-five.

Julian Castelli (27:53.343)
Well, I think of this sort of chapter on fly twenty five.

Victor Cheng (27:57.419)
of mapping out when do you do where on the chessboard do you jump, you know, in in any particular year. So an example of that would be on slide twenty six. this is what we want to do today, this we want to do in a couple quarters, where we want to be three years from now. This is our roadmap forward. And so I think it’s a useful way to think about

Julian Castelli (27:58.732)
So an example of that would be in the 526. This is what we want to do today. But this would be for a specific part of your part of your your product here. yeah.

Victor Cheng (28:22.548)
yeah. So like for your for a company’s portfolio of products or suite of products, maybe to this is something you might wan wanna talk about with your board or your investors and then eventually with your customers. Hey, this is our this is our AI strategy, right? We we

Julian Castelli (28:27.587)
Maybe Q this yeah, but not with a partier that.

This is our this is the RBI strategy. And how you’re working from bottom left to top right eventually, right?

Victor Cheng (28:38.166)
Eventually, yeah. Or you sort of go bottom left to like mid up but all the way to the right. Right. Maybe that’s what makes more sense. Right. You know, or it’s more to the left but higher, right? So there’s not like a right approach, but there it should be an intentional approach based on knowing your customers and market and

Julian Castelli (28:44.265)
Okay. Like we have in this example.

Yeah,

a light approach. But there it should be an intentional approach. And it’s so important to have the end in mind here, even if they’re you’re not there yet, so that you can start building, you know, getting the data, getting the the processes mapped. Otherwise if you just try to go to the end, you’re you’re gonna realize you you know you don’t you don’t have all the supporting infrastructure, correct?

Victor Cheng (29:01.847)
Absolutely.

Victor Cheng (29:14.56)
Absolutely. There is lead time on all this in personnel, in I would call implicit data, qualitative knowledge.

Julian Castelli (29:22.68)
data quality of knowledge, quantitative data, stable databases, API access, things like that. Or things like you you know the end money work backwards to say, you know, that thing you want to do a year from now, we don’t have that data. Right. It’s time to turn my head clean We don’t have a person else. Yeah. They gotta go clean it up, we gotta do it validated. That’s gonna take a year.

Victor Cheng (29:24.142)
In quantitative data, SQL databases, you know, API access, things like that. And so you when you know the end in mind, you work backwards and say, you know, that thing we want to do a year from now, we don’t have that data. Like it’s fragmented, it’s dirty, right? It’s we need to get someone to clean that up and we don’t have that person in the house. Great. Well, we’re gonna go hire somebody, then we got to go onboard them, they gotta go clean it up, we gotta validate it. That’s gonna take a year. shoot, we’re like two weeks behind. We’d better get that person on it now.

Julian Castelli (29:51.789)
And I can see why this this this framework forces you to think through those things and you might put a later date on that given that that operational requirement that you just described.

Victor Cheng (30:05.088)
Absolutely. It’s thinking through all the steps needed to get to the outcome that you’re intending and and seeing where can things fall apart. And so I think if you work that way, it’s easier to spot the problems and then and not have problems be a surprise, right? Everyone hates negative surprises. So but if you know it’s gonna happen, then you sort of you can be proactive about it. Yeah. yeah, let’s jump ahead a little bit. I got a

Julian Castelli (30:07.161)
All the stuff needed to get the document that you’re intending and taking direction things fall apart. So I think if you work that way, buttons and then have problems based advice, right? I don’t think negative surprises of like what’s gonna happen that be five. Makes sense. Okay, so we got the chessboard. Where do we go next?

Victor Cheng (30:34.04)
Let’s talk a little bit about AI and operations. That makes sense. So let’s go to slide, where’d that go?

Julian Castelli (30:36.818)
Okay. Yeah.

Victor Cheng (30:43.16)
Slide forty eight.

Julian Castelli (30:44.824)
So so just the the transition here. We’ve been talking about AI in your product now. Now we’re gonna go AI internally on the operations, right? Forty eight? Yeah, forty eight. Okay.

Victor Cheng (30:51.18)
Yes, yes, absolutely. yeah, 48, and then sort of the the key ideas here on 49 for AI operations. these are a couple key topics that are really important to think about for AI operations, right? We’re gonna go through each one of these one one at a time. so the the first one is maturity for internal operations, right? And

Julian Castelli (31:04.914)
go for each one of these one at a time. so the first one is maturity for internal operations.

Victor Cheng (31:13.214)
I I think that AI can be so powerful for internal operations. It’s probably underappreciated because the the sexy products for it in demo sort of like to have that wow factor kind of kind of captures headlines. but like there’s a lot, there’s a lot of nitty-gritty daily task.

Julian Castelli (31:14.431)
that AI can be so powerful for internal operations. Yeah. Kind of kind of captures that line. But like there’s a lot, a lot of maybe well and it’s more accessible because it’s your own sandbox, right? You so you can you can you can iterate, you can you get a lot more flexibility what you’re doing in s within your four walls than when you start creating a commercial product, right?

Victor Cheng (31:44.62)
Absolutely because you you you have access to the end user, which is like Joe down the hall, right? and and so you can get feedback. You have

Julian Castelli (31:48.906)
That’s right. Yeah. And you can tolerate, you know, mid you know, th the the the process where it’s not, you know, it’s not perfect yet.

Victor Cheng (31:59.383)
And it doesn’t need to be immediately either. Right? So it’s much more forgiving. You already have more of the resources needed, maybe not all of it, but more of it. you know the domain, right? Your staff has been struggling with this for years or decades. So you got domain.

Julian Castelli (32:01.384)
Yeah, right.

Julian Castelli (32:10.087)
That’s right. That’s right. You have all the advantages here, theoretically.

Victor Cheng (32:16.876)
Sorry, say again? Absolutely. Absolutely. And it applies beyond sort of pure tech companies, right? Service businesses, other kinds of businesses. Yeah, absolutely.

Julian Castelli (32:18.013)
You have all the advantages here. you know.

Julian Castelli (32:25.275)
100%. Yeah. Yeah, absolutely. so same thing, and then a couple of slides in terms of AI automatically, like how much of how do you want to give it. the one thing on slide digital I have, which is really underappreciated a lot for is specifically in operation, which is a bit different than external, is you can use the gathering. Like AI is really good for.

Victor Cheng (32:28.402)
so same thing in the next couple of slides in terms of AI autonomy, like how much autonomy do you want to give it? the one thing on slide 52 I’ll add, which is really underappreciated a lot with with using A specifically for internal operations, which is a bit different than external, is you can use AI to do planning. Like AI is really good for planning.

and I think most people don’t even realize that’s a possibility. They don’t use it very often. So everything I’ve learned about AI, I learned having AI tutor me.

Julian Castelli (32:53.101)
I think most people don’t even realize that’s enough literally often. So everything I learned about AI, I learned having AI tutoring. Hey, there’s just new technologies that I’m not familiar with. Like, what does the stack do? What does it analysis to in terms of highway understanding? I didn’t lay out all the technical trade-offs, give you an introduction. One of my favorite fonts that I have as sort of vacant references is.

Victor Cheng (33:02.734)
Right. Like, hey, I there’s this new technology stack I’m not familiar with. Yeah, can you explain it to me? Like what does the stack do? What is it analogous to in terms of I already understand? And it can lay out all the technical trade-offs, give you an introduction. one of my favorite prompts I have sort of baked in there my preferences is explain things to me like I’m an idiot, right? Explain it in lay person terms on everything. That’s one of my prompts, that’s sort of automatic. And the other is tell me why I’m wrong, right?

Julian Castelli (33:21.146)
Explain things to me like an Indian person on everything that’s on my class. that’s sort of automatic. And the other is tell me why I’m wrong. Don’t tell me, don’t, you know, don’t just sort of tell me what you think I want to hear. Yeah.

Victor Cheng (33:32.493)
Don’t tell me, don’t, you know, don’t just sort of tell me what you think I want to hear. Tell me why this is wrong, why it’s flawed, what are its weaknesses and on on everything. So now when I say, hey, I want to go, what are the hours for the burrito place down the street? It’s like, well, the weaknesses of burritos are right, it’s not as, you know, whatever. so so it kind of takes things quite literally, but but planning is really helpful. So if you say, hey, I want to have this grand AI vision of fully automated workflow in my finance department, where do I start?

Julian Castelli (33:39.569)
Now when I say, hey, I want to go to the hours for the weekend down the street, it’s like, well, we can still redo start, right? So I kind of fix things back a little bit about that. So making this really helpful. So you say, hey, I want to have this AI vision of a workflow and I find this department. Where do I start? But you don’t tell you like what’s data. Yeah. Okay, I took the five phase. They look pretty good at that, right? Right.

Victor Cheng (34:01.207)
Like it’ll tell you like what’s needed. And it’s like, okay, break up into five phases. It’ll it’s pretty good at that, right? And getting the phases right. you’ll see a lot of

Julian Castelli (34:08.951)
That and that’s so important because so many people have prompt paralysis because you know the the task, you know, you think you have to architect the whole thing and that’s overwhelming. And what I’ve heard people say, what I’ve used myself is, you know, no, no, should start with the goal and then and then just just iterate, right? You know, ask ask me questions. Let’s go, let’s build build this up one at a time and and and start there. And so that’s

Victor Cheng (34:25.55)
Yeah.

Julian Castelli (34:35.791)
One of the reasons I’ve found that just diving in and doing it with the goal and starting with the goal teaches you how you can you can you can iterate and and extract strategy or even planning steps with the AI, even if you didn’t know how to get there in the first place.

Victor Cheng (34:50.346)
Absolutely. And you I used to do this manually, sort of in my McKinsey days and since then. Like, what are the key dependencies in the plan? Like, if I don’t have A, B won’t work. Tell me those things and then help me assess whether A is good enough. Like if you break it down to that level, these are task level things A is quite good at, but you break up into pieces. And so it can really help de-risk a project plan. It can really help educate you, board members and and staff as well.

Julian Castelli (34:59.086)
Mm-hmm. I don’t have a it won’t work. Those things. And then help me assess whether A’s good enough. You can break down that level. These are just level things not good at. But you break up the people and so it can really help T Risk a project plan and really help educate you what numbers against that as well. Makes a ton of sense. Yep.

Victor Cheng (35:21.998)
that’s good. Yeah, let’s see. let’s cover just the highlights here. So let’s go to slide fifty-five, top five AI insights.

Julian Castelli (35:22.393)
What what ne what’s next page?

Victor Cheng (35:34.581)
These are things I think maybe might be counterintuitive to some listening. So insight number one is I I I firmly believe, and Julian, you absolutely epitomize this, CEOs have to use AI personally, at least at a moderate to advanced level. You can’t have a CEO that doesn’t know how to use AI, you know, other than very basically lead an organization that is very advanced. Like I’ve just I’ve never seen that happen. Because

Julian Castelli (35:43.791)
They miclist the

Julian Castelli (35:52.305)
doesn’t have to use a hit I other than very basically an organization that is very advanced. I I’ve I’ve never seen that happen. Hey Victor, can you can you repeat that? I think I’m just gonna we’re gonna edit that out. I lost you there on the internet connection. Let’s let’s start with take me through insight number one, please.

Victor Cheng (36:06.242)
Okay, no problem.

Victor Cheng (36:09.772)
Yeah. So insight number one is CEOs have to personally use AI at a moderate to advanced level. And it just doesn’t seem to work because it is such a different way of doing work, unlike anything I’ve experienced in my multi-decade career, that if that leader doesn’t know, have some personal grasp, I just don’t see them being successful in leading that change at organization wide. So that’s sort of insight number one.

Julian Castelli (36:16.707)
Yeah, I agree.

Julian Castelli (36:20.524)
Because it’s a different different maybe.

Yeah, if that paper doesn’t now have some personal privacy, I just don’t think it’d be successful that you can actually have sort of things like I I completely agree. It’s you know, this is not something you can just assign to someone to to run. I I I I tell people you do need to have a have a leader, you know, drive a full program, but you have to be leading from the front as well. It g you cannot be checking in monthly to see how things are going or it won’t work.

Victor Cheng (36:47.64)
Absolutely.

Victor Cheng (36:54.082)
Yeah, and one of my, you know, one of my principles, I I cover this in in the book, Teach AI your company, is one of the key roles needed to make be successful with AI is you need a a a leader that will is obstinate and stubborn and annoying, right?

Julian Castelli (36:59.079)
AI, you’re not the goals we need to maybe be successful at AI is we need a aerodynamical constant engineer to force through the the the the traction, the the the the what do you call it, the resistance to change.

Victor Cheng (37:16.334)
All that inertia. Absolutely. Absolutely. There is so much. And and that is not an AI skill, right? That is a stubbornness trait. Absolutely.

Julian Castelli (37:23.968)
No, that’s a s that’s a s that’s a trait. Resilience.

Victor Cheng (37:29.888)
Yes. And I think when I see companies doing it well, it is the sheer abstinence of the CEO and determination overriding dozens and dozens. Yeah, yeah. So super critical role. But it’s hard to have that conviction if you don’t use it personally, right? Because it’s so different. It’s so different. the other thing I’ll mention too, as a as an aside, you know, every other technology, Julian, you and I have seen in our lifetimes prior to AI.

Julian Castelli (37:35.158)
Yeah. That’s a great point. We gotta that that that that I I couldn’t agree more.

Julian Castelli (37:47.094)
Right. yeah. Different. the other thing I’ll mention too is a side, you know, every other technology and I have seen in our lifetimes, prior to AI, there was some in the log version of it before, right? So there was the big video television. And television was video pictures. Yeah

Victor Cheng (37:57.091)
There was some analog version of it before, right? So there was land-based radio and then there was television. And television was radio with pictures, right? Then came the cell phone. The cell phone was like a landline that you and I grew up using, except you there’s no cord. It’s a phone without a cord, and you gotta charge it. Right. So you know how to you have some intuition for how to use these new devices. AI is totally different. There’s no analog that’s easily accessible. That’s why it’s so important that the CEOs themselves have to use it.

Julian Castelli (38:13.191)
get a charge. You know how to you have some intuition from how the US is new devices. Yeah it’s totally different. There’s no analog that’s PC accessible. That’s why Joseph wanted that as DOS themselves on PC. I agree. All right, what’s inside number two?

Victor Cheng (38:28.398)
And side number two is AI today and even half a year ago is good enough to be practical and useful in every functional area, no exceptions whatsoever. Right? It is good enough. Now the context, the data hygiene, your leadership, change management, that might not be, but the AI itself is already capable enough and is and is only getting better every couple of weeks.

Julian Castelli (38:33.493)
Then impactful is

Yes. Yes. But the data itself is already capable of that because and it’s only getting better at before. How do you decide where to start? If it could be done everywhere. Is that a little paralyzing?

Victor Cheng (38:51.478)
Yeah. So yeah. So I it helps to use AI so you get some sense of what it is good at and what it is not. And there are certain things I would never have AI do. So anything super high stakes, right? You want more eye more human eyes involved. where AI is super really good is very simple tasks to start. Right. So an in you know an email comes in.

Julian Castelli (39:02.204)
I would never have been able to do. So anything super high stakes, I think you want more humanized. Right. But is it reasonable gas discharge? Okay. So it comes in. If it’s a complaint email with their using other words cursing us, or maybe happy. Yeah. Like yeah, I think that that like 89% of the time. Right. the classifications are pretty clean.

Victor Cheng (39:17.452)
Hey, is this a complaint email where they’re using four-letter words cursing us? Or are they really happy? Right. Like AI is pretty good at that, like 99% of the time, right? so classifications are pretty easy. summarization, right? We’re all sort of recording meetings and getting transcription of notes. That’s pretty good, right? Those kinds of things, it’s like very, very reliable for. and there’s a lot of that, right? Just at that level, there’s a lot of it going on in the organization. Okay. So I like things that are simple.

Julian Castelli (39:30.295)
summarization, right? We’re all sort of recording EP and getting transfers in books that we have. Okay, those kinds of things that’s a very reliable firm. there’s a lot of that, right? Just at that level, there’s a lot of things going on in the organization. Yeah. So I like things that are simple, but annoying these kinds of things that nobody likes doing. And there’s all there’s a lot of applause, right? So it’s easy because it’s a lot of effort and for lots of

Victor Cheng (39:46.867)
but annoyingly time consuming that nobody likes doing that and there’s a there’s a lot of volume of it. Right. So it’s each instance is a lot of effort and there are lots of instances. Those are good good candidates for using with AI.

Julian Castelli (39:58.383)
Those are good good hands for using the data. Got it. Okay.

Victor Cheng (40:02.542)
insight number three, I you know, I think I I mentioned this earlier that the biggest constraint to AI adoption is not technical. I think it’s psychological. So it’s the change management, the education, the fear, is this gonna take my job? Like all this stuff, all the emotions, ironically enough, that sort of you know are are creating this kind of resistance.

Julian Castelli (40:05.955)
I think I mentioned this earlier that the biggest constraints of the idea is not technical. I think it’s psychological. Yeah. So it’s the change management and education, the fear that’s going to take my job, like all this stuff, all the that’s sort of you know our areas kind of resistance. Yeah, I I agree. I think it’s you know, it it gets back to that that bicycle analogy. It’s just it’s just you know, you have to change your the way you work. And yes, it’s gonna be

unproductive and distracting for a little while, but the return is so strong, right? Yes, I think it makes a much bigger

Victor Cheng (40:35.342)
It is, yeah. So I think it’s a much bigger distraction than I think even most people expect, and it’s a much bigger impact and benefit too. So it is a yes and it is yes and yes. Yeah. And so that’s what leads us to insight four, which is I think AI is now primarily a CEO leadership or change management issue. It’s not a technical constraint. It is a leadership issue first and foremost. The second issue becomes technical, you know, pretty quickly here.

Julian Castelli (40:43.206)
bigger than benefit too. That’s right. The the cost is higher and the benefits higher. I agree. I agree.

Julian Castelli (40:52.488)
AI is not the CEO or change. It’s not a technical constraint. It should become technical. Yeah, the technical it’s it, you know, it doesn’t it’s does not auto-magic. It it it requires some some skills and processes, but it’s yeah, it’s never gonna happen if the CEO is not, like you said, stubbornly driving change. I I I agree with that.

Victor Cheng (41:15.726)
Absolutely. And and I think the last insight here on number five is that using AI as an AI mentor for CEOs, I think it’s great. Like it is very good at teaching you how to use AI better and explaining concepts. You can take you know copy and paste news articles, right? That may not be over your head a little bit, and it’ll explain to you. And

Julian Castelli (41:27.429)
Is very good at teaching you how to use AI better and explain to the absolute copy paste new variables, right? That yeah, over your demo and it’ll take you if you want it to if you say, Hey, I think I want to use this thing in accounting, like how would I use this big use or at least different use cases? Which one of these should I start? How would I decide? Like an equal conversation there, I’m thinking not to clarify things a lot. So you do I I I I agree.

Victor Cheng (41:38.484)
And if you want it if you say, Hey, I think I want to use this thing in accounting, like how would I use it? Is is this a good use case? Or I have these four different use cases. Which one of these should I start and how would I decide? Like you can engage in a in a meaningful conversation there and it can help clarify things a lot. So it’s really good.

Julian Castelli (41:56.944)
Fantastic. Okay. Well, anything else you want to cover in this deck?

Victor Cheng (42:01.89)
That yeah, just maybe one last slide, just sort of end on that. yeah, I I think the training versus using AM models, right? So that’s slide sixty-one.

Julian Castelli (42:03.483)
Yeah, I think it’s

Julian Castelli (42:07.043)
I think the bank versus using the slide

Victor Cheng (42:14.798)
And I’ll end on this note, right? So there’s using AI models like ChatGPT, ask a question, and then there’s training in this may in this or really tr not training in the sense of building your own model, but providing a context, right? So teaching AI about your own business, you know, which is the subject of my book. That’s so much of the expert.

Julian Castelli (42:16.529)
all the smell, right? So there’s using AM models like JBT ask your question. And then there’s training this maintenance maintaining the sense of building your own model, but like in context. Right. So teaching AI about your own business. Right. Yeah, but you’re you’re that teach AI your business. I I actually misspoke earlier, but is it it’s is deliberately like that, because you’re teaching the AI about your business is what you’re what you’re saying, right?

Victor Cheng (42:43.384)
Correct. Yeah. Most people think I should teach my company about w about AI. I’m like, no, no, no. I mean, yeah, you probably should but

Julian Castelli (42:47.737)
No, I missed that and now I now I see it. So so that yeah, let go let’s go deep there because that that’s that’s worth the whole book. I’m looking forward to reading it.

Victor Cheng (42:55.126)
Yeah, yeah. So, you know, the book’s title is this principle that most people are missing, which is why I kind of motivated to write it. It’s Teach AI Your Company, which is basically the framework for you gotta provide AI models with context about every little thing that you and I do as human beings and with our staff, right? There’s so many things, like my staff just knows certain things about me, right?

Julian Castelli (43:01.75)
is teach AI your company, which is basically the framework for you gotta provide AI models with context for every little thing that you guys do as human beings. That’s right. I there’s so many things like I just certainly got me not written down here because now it used to be how you do proposals, how we do invoices, what do we do them? I remember I was using AI for the carrier

Victor Cheng (43:19.008)
It’s not written down anywhere. I mean it is now, but it didn’t used to be, right? How we do proposals, how we do invoices, when we do them. I remember I was using AI for an entire year, and then it finally dawned on me like, you know, I don’t think AI knows the name of my company. Like, you know, I I went through my search, like, I don’t think I ever told it. It knows I have one. I never mentioned the name. And I was, no, I don’t think I told it what we sell. And it’d been a year. It had been a year.

Julian Castelli (43:31.458)
and mind on the mic. Well, I don’t think AI knows that my company.

I don’t think I told it one at all. We talk about it, we think about it all day long. I didn’t know. Right. He doesn’t know who my customers are. He doesn’t know who my best customer is. So I made a you know it’s a one-page document. This is the company name. This is what we do. I took all my conversations with my best clients in the past 20 years. This is our phototypical.

Victor Cheng (43:48.771)
My staff knows, I know, we talk about it, we think about it all day long, but it didn’t know. And then I realized, you know, it doesn’t know who my customers are. It doesn’t know who my best customer is. Right? So I read it, you know, it was a one-page document. This is the company name. This is what we do. And I took all my conversations with my best clients in the past 20 years. This is our prototypical ideal customer profile, right? And it was just two pages. It was a PDF document. That’s it. I put it in there with every single chat afterwards, two to three times better.

Julian Castelli (44:09.601)
Ideal custom profile, right? And there’s just two pages. It’s a big gift document. That’s it. I put it in there with every single chat afterwards, students becomes better. Yeah. Yeah. Every answer gets

Victor Cheng (44:18.712)
Just those two pages alone, every answer gets two to three times better just with two pages. And so

Julian Castelli (44:25.005)
That’s a great insight. So so that that that that that insight, you know, sp the sounds like it was the inspiration for your book, because that’s what you were doing. You were teaching AI about your company.

Victor Cheng (44:33.652)
Absolutely. And now we have maybe four million pages, right? Or four million words rather. And and it’s increasingly useful. We’re having it do things like write blog posts. We’re having it do internal work. it’s really quite amazing what it can do. We have self self-learning, AI writing code and fixing its own bugs. a lot you can do, but it was a lot of effort, and it starts with teaching the very basics of your company to AI, and then you build from there. That’s kind of the the problem.

Julian Castelli (44:35.947)
A million pages to make words rather.

Julian Castelli (44:50.534)
And starts with teaching the very basics of your company to AI and build there. I love that. Perfect. Well that makes sense. Victor, where where can people when can people expect to see your book and like where can they find it?

Victor Cheng (45:10.902)
Yeah, so the book called Teach AI or company will be available on Amazon and digital bookstores everywhere starting September first. and I’m reachable online at sasceo.com.

Julian Castelli (45:22.113)
Fantastic. And and your presentation here, we’re gonna put it up on the Growth Elevated website so people can take a look at it and they could follow up with with you directly at saceo dot com. And you know, thank you for joining us at at Growth Elevated in January and now here on the podcast. This is really valuable and relevant information for our audience and I look forward to continuing the dialogue with you in person.

Victor Cheng (45:45.154)
Thanks, Julian. Appreciate it.

Julian Castelli (45:47.105)
Thanks, Victor.

Introduction to Growth Elevated & Victor Cheng (00:00:00)
Julian welcomes Victor Cheng, CEO of SaaSCEO.com, to discuss how AI is changing the way growth-stage companies create value, scale operations, and prepare for the future.

Why AI Is a Leadership Challenge, Not a Technology Challenge (00:03:07)
Victor explains why the biggest barrier to AI adoption isn’t the models—it’s human psychology, change management, and getting leaders to prioritize implementation.

How CEOs Should Think About AI Adoption (00:05:10)
Why many founders delay AI initiatives, what separates proactive leaders from overwhelmed ones, and why CEOs must personally lead AI transformation.

How Large Language Models Actually Work (00:08:57)
Victor breaks down LLMs as prediction engines and explains why context matters far more than most people realize when using AI effectively.

Teaching AI About Your Business (00:12:37)
Why documenting your company’s knowledge, conversations, and expertise today becomes the foundation for powerful AI capabilities tomorrow.

Building an AI Product Strategy (00:18:29)
Victor introduces his AI Playbook, covering AI autonomy, product roadmaps, implementation priorities, and how companies should think about long-term AI strategy.

Using AI to Improve Internal Operations (00:30:43)
How leaders can apply AI across finance, sales, marketing, HR, and planning to improve execution before building customer-facing AI products.

Five AI Lessons Every CEO Should Know (00:35:34)
Victor shares his biggest insights, including why CEOs must use AI themselves, why leadership matters more than technology, and how AI can become an executive mentor.

Why Every Company Must Teach AI Their Business (00:42:14)
Victor explains the core idea behind his book Teach AI Your Company: AI becomes exponentially more valuable when it’s given the context, knowledge, and documentation unique to your business.