How to get started in AI for your Business, Expert AI Strategies - Episode 012

00:00:00: Hello and welcome to this week's episode of the AI Boardroom.

00:00:05: And today we have a really special expert, Svetlana.

00:00:11: You're the expert actually.

00:00:14: Tell the people about what we're doing.

00:00:17: Yeah, I think what we wanted to dive in just to give you guys a little glimpse as

00:00:22: to what AI strategy is and how to implement it, what it means to actually

00:00:27: bring solutions to

00:00:30: to kind of fruition, right, in your enterprise.

00:00:33: So, Edgar's gonna do some talking.

00:00:35: We'll kind of do a little bit of talking.

00:00:37: But the idea, and again, it's a limited amount of time, so we are gonna scratch

00:00:42: the surface.

00:00:42: And if there's any particular parts that you wanna dive deeper on, we're happy to

00:00:48: do a separate episode.

00:00:49: And I think if you're staying until the end, there's a little bit of a surprise

00:00:52: that I think Edgar is gonna offer.

00:00:54: So, I hope you guys stay till the end to find out what it is.

00:01:01: Yeah, thank you very much.

00:01:03: So Svetlana, why are you an expert in strategy?

00:01:06: Tell people a bit about your background.

00:01:09: Yeah, so I've been working in digital for over a decade.

00:01:16: So building products, supporting strategic initiatives across different

00:01:21: organizations.

00:01:22: And it's always been kind of in the innovation front.

00:01:25: So the types of products or projects I've been leading.

00:01:29: have been at new portfolio levels or new initiatives that, you know, the

00:01:34: organization was either trialing.

00:01:36: There's not a lot of kind of research or maturity in that space.

00:01:40: So I was kind of initially recruited as one of the first to kind of lead those

00:01:45: efforts and that hasn't changed even when I entered healthcare.

00:01:48: So I do have a full -time job where I continue to lead these innovative efforts

00:01:53: and the type of projects that I continue to work on and create strategies.

00:01:58: whether it's at the portfolio level or at the individual product level, they tend to

00:02:02: be very new.

00:02:04: And it's exciting to me because it's things that create new business value.

00:02:09: It's the things that drive additional revenue or drive additional value to the

00:02:12: organization in that new ways.

00:02:15: And I know that there's a lot of talk and we spoke with Edgar quite a lot about even

00:02:21: in our past episodes on the news covering...

00:02:26: You know, the problems with cheat sheets and, you know, rolling out and talking

00:02:31: about prompt guides and all kinds of things.

00:02:34: Yes, they streamline operations, but they're not creating that new value for

00:02:38: your business.

00:02:38: And so I think that's what I do.

00:02:41: And I, you know, have an understanding of the business.

00:02:45: You know, I'm a trained MBA.

00:02:47: I went to a top 15 school for my MBA, paid a lot of money for.

00:02:52: for my degree as we kind of discussed with Edgar.

00:02:57: And I'm trained in also workshops for opportunity mapping within, you know,

00:03:05: specifically in the AI space.

00:03:07: I have been a practitioner in developing AI solutions for healthcare.

00:03:13: And then I've been working with other companies to help ideate solutions within

00:03:17: their business, how to bring those and scale solutions across different

00:03:21: industries.

00:03:22: Construction, finance, really, I mean, healthcare is not limiting.

00:03:26: I think it's this innovative aspect of it that excites me most.

00:03:30: It's like, how do you bring more and new value to the organization and how do you

00:03:34: have the strategic vision into the future?

00:03:36: Because AI, to be frank, it does take a little bit of time to mature and to

00:03:40: actually get value once the solution matures and scales.

00:03:46: So how do you have that forward thinking?

00:03:49: to bring that new value to your organization.

00:03:52: What is the interpretation of a new value and more value?

00:03:59: More value, I think, is meant to elevate existing stuff.

00:04:05: And new value would be going new path.

00:04:10: How would we understand that?

00:04:12: Yeah, there's a few ways that you could define new value.

00:04:16: So one is, let's say you do have legacy products.

00:04:19: And for example, they have been built traditionally and they haven't been

00:04:23: improved over a long time just because of technology was delivering factor.

00:04:27: And so you can create new value by improving or rethinking that whole

00:04:32: workflow with, for example, AI.

00:04:35: So AI kind of opens up a new world of possibilities, new world of efficiencies.

00:04:40: So sometimes it's time to create or rethink that whole architecture.

00:04:45: And again, I have experienced rebuilding tools that were...

00:04:48: built in kind of legacy via legacy software, and that you're rethinking it

00:04:53: completely with AI using data and things that people were kind of seeing as like

00:04:58: impossible before.

00:05:00: And they were doing, you know, parts of that solution providing value to different

00:05:05: parts of the organization.

00:05:07: Now it's possible to create more new value for those legacy, like what used to be

00:05:13: legacy products, but now AI can take that to...

00:05:17: a whole net new scale and provide that value to the rest of the organization that

00:05:22: wasn't possible before.

00:05:24: You can also create new value in thinking about possibilities that you just didn't

00:05:31: think could be done before.

00:05:34: So when you create new value, and these are also some of the projects I'm very

00:05:39: excited about doing, is looking at white space opportunities.

00:05:43: So white space opportunities basically means that...

00:05:45: There maybe has been some research ideation happening, but there's not

00:05:49: solutions in the space that have tackled or successfully approached that solution.

00:05:55: So you create and it requires some level of strategic thinking in order to find

00:06:01: solutions for those spaces.

00:06:03: Because if it was easy, people would have figured it out and it would not be a white

00:06:08: space opportunity.

00:06:09: But you can create new value where no solutions exist and they're now made

00:06:14: possible with AI.

00:06:16: And they do tend to be a little bit more complex, complicated, then require a

00:06:19: little bit more thinking and maybe potentially budget.

00:06:22: So the types of solutions that are released in that space tend to be like

00:06:25: larger organizations and enterprises.

00:06:28: But that's really, you know, a huge competitive mode for the organization when

00:06:32: you're putting your data to use in more unique and innovative ways, because that's

00:06:36: IP that you could potentially sell too.

00:06:39: So again, those types of solutions and products really excite me.

00:06:44: because again, the potential is limitless.

00:06:47: Interesting.

00:06:48: And so you said you like do it for a long time.

00:06:51: I think it wasn't always AI involved back in the days.

00:06:56: What makes AI special or AI projects in specific?

00:07:03: So traditional products, digital products go through a maturity cycle.

00:07:09: So...

00:07:10: just think about not even digital, but flash disks, right?

00:07:14: So flash drives, you barely see those around anymore, right?

00:07:20: Because they do follow this kind of bell curve type of a maturity, they raise to

00:07:28: the top and then they reach the peak of like the highest demand and then they

00:07:33: decline in demand just because there's not much revamp that is happening and not much

00:07:39: improvement that you can do.

00:07:40: beyond what that product can actually handle.

00:07:43: And that's true of digital products.

00:07:46: Google, you could even argue, is going through that phase right now where they've

00:07:51: improved that and optimized it to a certain phase.

00:07:54: And now a new emerging solution came in, which is doing solutioning for that space

00:07:59: in a completely different way.

00:08:01: And it's being disrupted.

00:08:02: So you can also position it as kind of going down the decline.

00:08:07: For AI...

00:08:09: products, what's great about those solutions is that user feedback kind of

00:08:17: drives the improvement.

00:08:18: So they call this the flywheel effect.

00:08:21: So as you're putting in more data, you're providing a better user experience, better

00:08:27: recommendations.

00:08:28: So think Netflix.

00:08:30: And why people continue to come back to it is because of those really...

00:08:35: attention grabbing headlines or attention grabbing thumbnails, people want to see

00:08:39: more because Netflix gets me.

00:08:42: Netflix understands what I want.

00:08:45: And because users experience such a great user experience, they recruit more users,

00:08:50: which drives more data and it kind of continues going up in that flywheel effect

00:08:55: again, until that product potentially gets disrupted.

00:08:58: But the...

00:09:01: great idea of the solution.

00:09:03: There's really no limit for how well that system, for how much that flywheel can get

00:09:10: better and better because the more data, the more users and the scalability of AI

00:09:16: solutions is limitless.

00:09:17: Again, the only limitation is the compute behind it.

00:09:21: I think at some point, I think improvements get so marginal that you kind

00:09:29: of...

00:09:30: You might be limitless theory, but the improvements are at some point mitigating

00:09:37: or also small, they are not really benefiting the whole process anymore.

00:09:43: But if you think about also the scale, so when the traditional products kind of

00:09:49: provide a generic experience, right?

00:09:52: So you can't do the same level of personalization that you can with these AI

00:09:58: systems.

00:09:59: So...

00:09:59: The, I would argue that the improvements are done at a user level, right?

00:10:03: So like you can get that experience fine tuned to that exact user over time and our

00:10:08: preferences change, right?

00:10:09: So, and then the beauty of it is that the systems are reactive and then they're

00:10:13: going to adapt to those changes as the user adapts in their kind of cycle.

00:10:18: So that's another difference is that AI can handle, yes, more data, more users and

00:10:24: things like that.

00:10:24: So generally the product itself will continue to improve.

00:10:28: But the level of personalization and how customized that experience for each

00:10:32: individual user is, is a huge differentiating factor for what AI systems

00:10:39: can do at that scale.

00:10:41: So you can actually personalize every experience.

00:10:44: That's what Netflix does.

00:10:45: Your experience would be completely different from mine and the types of

00:10:49: thumbnails I see would be different from yours.

00:10:52: I've seen an interview of the CEO of

00:10:59: of Nvidia.

00:11:00: He was at Stanford because I think he graduated there back in the days and then

00:11:06: he built Nvidia and up to this day he is CEO there.

00:11:11: So, and he was talking about generative AI and also like huge performance

00:11:16: improvements that are going to come into in the next like four, five to six years.

00:11:21: And he was telling me like, he was saying something really interesting.

00:11:26: He said like everything,

00:11:27: we had up to this point was kind of pre -recorded, even if you're like video,

00:11:32: audio, of course, but also if you have an app, like the UX is also pre -recorded,

00:11:42: kind of, because like you put in the ways to go through the app beforehand and

00:11:50: they're kind of fixed.

00:11:51: You might have branches, but they are limited.

00:11:55: And rightfully so.

00:11:57: And in the future, we will have basically everything generated on the fly, which I

00:12:04: found interesting.

00:12:06: I mean, the dynamic experiences, right?

00:12:08: So like, it doesn't have to be static.

00:12:10: And I think, again, I think, I don't remember what the user experience that I

00:12:17: was seeing, but like, why not?

00:12:19: Why not, you know, adjust the experience of the app based on what your...

00:12:26: kind of experiencing.

00:12:27: So maybe the future even of iPhones, like, why do you have to see a home screen full

00:12:30: of like these busy apps?

00:12:32: Why can't anticipate based on your schedule or what you have going on in that

00:12:36: day and, you know, bring on the platter to you, like, here are some options of what

00:12:41: you can do now and it becomes much more efficient.

00:12:44: So maybe the future of like these operating systems becomes much more

00:12:48: contextualized to what you have going on.

00:12:50: I think we are close to the first step of this.

00:12:54: I'm pretty sure.

00:12:55: that this year's developments will make a huge step on our mobile devices to become

00:13:02: personal assistants.

00:13:05: We'll see.

00:13:06: Apple is releasing their new operating systems and maybe a new version of the

00:13:10: iOS.

00:13:12: Yeah, yeah, like the new Android is coming out, it's like right around the corner.

00:13:16: And in June, we have the Apple Developer Conference.

00:13:20: So yeah, everything's close.

00:13:22: I'm really curious what this will bring.

00:13:25: Well, yeah, I'm a new company.

00:13:30: I'm an old company.

00:13:32: That's better.

00:13:34: I have my processes, but I'm seeing my competitors, which like we've seen in the

00:13:40: last news episode is 79 % of managers think that they have to integrate AI to

00:13:49: stay competitive.

00:13:51: And I highly agree.

00:13:52: because the efficiency gains you might or can get are so high that your value output

00:14:02: grows really, not exponentially, but at a really high rate after implementing AI.

00:14:12: So either you save costs and you can be more price competitive or you save or you

00:14:17: create more value and can just take higher prices.

00:14:21: or you can just create more, like handle more load to sell even more.

00:14:26: So there is a lot of opportunity in there.

00:14:31: And some, like I always say, like someone out there will do it.

00:14:34: So, and go ahead and actually do it.

00:14:40: But now I'm completely non -technical, but I know I need at least to evaluate how I

00:14:48: could use AI to elevate my business.

00:14:51: Where do I start?

00:14:53: What are typical use cases in businesses that might kind of be a blueprint to start

00:15:01: from?

00:15:02: I would say maybe three different steps I would recommend as starting points.

00:15:07: One is understanding what AI can do.

00:15:12: So I'm not talking about the history of AI or how all of these AI ML...

00:15:19: you know, deep learning techniques actually work.

00:15:21: Just trying to understand at a high level, elevate yourself to understand what can AI

00:15:27: do to be able to, again, understand like what translate that into your business?

00:15:33: Where do those capabilities, what could you utilize across your different business

00:15:40: functions in order to progress your business?

00:15:42: So at that first step, you're just trying to understand the fundamentals of AI, what

00:15:46: it can do.

00:15:47: and to set yourself up to understand like realistically what value it can deliver

00:15:51: into your business.

00:15:52: And that's why I think, you know, trying to maybe partner with a mentor, you know,

00:15:57: I do also coaching on AI, but, or take some introductory courses, again, with the

00:16:03: intent to understand what AI can do from a capability standpoint and not the

00:16:07: technology.

00:16:09: Two is, you know, as it...

00:16:11: respective to your industry, because you are going to see different utilization of

00:16:16: those capabilities across different businesses.

00:16:18: Let's say healthcare.

00:16:19: When you do legal banking, you're going to see how some of those technologies get

00:16:24: leveraged in completely different ways.

00:16:27: So getting an idea for how some of these AI capabilities being leveraged across

00:16:34: those industries.

00:16:36: And sometimes I think it makes sense to even look at parallel industries to just

00:16:39: get ideas for

00:16:41: how you could even potentially create new value in your business with how maybe

00:16:45: parallel industries are utilizing it.

00:16:48: So first you build the foundation, but in foundational understanding of

00:16:51: capabilities, you look at applications of those in the industry just to get your

00:16:56: ideas flowing.

00:16:56: And then you look in your own business internally as step number three, and you

00:17:00: kind of look objectively, well, what objectives do I have set for my business?

00:17:05: And is that uncommon for organizations to have one year, two year, five year

00:17:09: objectives?

00:17:10: to kind of understand like, how do I actually operate my business and what

00:17:15: matters to my business to focus on this year?

00:17:18: Are you focusing on minimizing customer churn?

00:17:22: Are you looking to increase profits for this year?

00:17:25: What are some of your high level objectives?

00:17:29: And then you look at your either product mix or services mix.

00:17:32: You look at your workflows, what kind of services you are providing.

00:17:36: And you're looking to see, okay, well, based on my objectives,

00:17:40: which one of these functions or product lines or service lines can actually

00:17:44: improve with AI.

00:17:45: And then you would layer, you know, there's workshops.

00:17:48: One of the other things that I do also is AI design thinking workshops or design

00:17:53: sprint workshops where you kind of ideate and you align the objectives to the

00:17:58: different products mixes or service lines and the capabilities and you ideate

00:18:03: different use cases within that.

00:18:06: So you're kind of coming in there with the lens of like not designing the end

00:18:09: solution, but

00:18:10: If I have this product, how can I boost it with AI?

00:18:15: How can I do that work better?

00:18:17: Or are there things that we have been thinking about doing, revamping this and

00:18:21: it wasn't possible before?

00:18:22: Because before AI, we just thought it was just the amount of business rules that we

00:18:27: had to implement was just unfeasible.

00:18:29: Can we revisit that improvement for that product?

00:18:33: Is there a process that's inefficient in the business that could be improved again?

00:18:39: in marketing and sales?

00:18:41: Are there things that are taking like the to get an approved marketing copy?

00:18:46: Is it currently take a week, two weeks, five weeks?

00:18:49: I don't know to get from a concept into the market.

00:18:54: Could we cut down on that time and can we get a few more turns of that marketing ads

00:19:02: into the market?

00:19:05: Like you look more objectively into your business and align it to your objectives.

00:19:10: And see, okay, what are some opportunities?

00:19:12: And what that enables you to do is to focus on not chasing the AI technology for

00:19:18: the sake of chasing technology.

00:19:19: You're looking more strategically at where can I invest money that matters to my

00:19:27: business?

00:19:28: So you make sure that that investment can actually give you a positive ROI and that

00:19:33: you're not investing it into like just chasing a technology and saying like, hey,

00:19:36: we actually do AI.

00:19:38: but it's not an impactful product that's actually moving the needle on any of your

00:19:44: strategic objectives.

00:19:45: So you kind of have to start top down.

00:19:47: And I think we talked about it yesterday too, when our news episode where the

00:19:52: dangers of not driving strategies top down is that sometimes employees, because of

00:19:58: how much value AI has, they're going to go and hunt those solutions themselves

00:20:02: because they want to do their job.

00:20:04: more efficiently and effectively.

00:20:06: And I think the report that we saw from Microsoft yesterday, that 75 % of people

00:20:10: are using some AI tools to support their workflows.

00:20:13: And at that point, it becomes really hard to measure the investment or the ROI in

00:20:19: those initiatives.

00:20:20: So you want to make sure that you are democratizing access, but you're doing it

00:20:25: strategically.

00:20:26: You're investing in the right things in the business that you can measure the ROI.

00:20:30: You can measure the investment that you've put in and how much value you're getting

00:20:33: out.

00:20:33: you're actually assigning metrics to that if you kind of do it in those steps.

00:20:38: So not like just handing it out and hit and hope.

00:20:44: But yeah, one thing I would love to add is if you have existing processes and you try

00:20:52: to evaluate them, look at all the parts that are actually not creating value.

00:20:59: For example, in sales,

00:21:02: managing the context or putting them into the CRM or writing a protocol, stuff like

00:21:07: that.

00:21:09: It has value with the information that are inside, but the process itself is just

00:21:16: tedious.

00:21:17: And that's something where something like a large language model with its language

00:21:23: understanding capabilities might be a way to streamline the process.

00:21:30: And I'm...

00:21:31: That's just an example.

00:21:32: So the value lies in like getting rid of the non -value creating parts of value

00:21:38: creation.

00:21:42: So yeah.

00:21:43: And another thing, what you can look at if you don't want to touch your existing

00:21:49: processes is like, what stuff did we want to do, but we weren't able to because it

00:21:55: wasn't cost effective enough to do it.

00:21:58: It might be like, it might AI, may AI help me to do it cost effectively or at a cost

00:22:09: rate where it actually makes now business sense where it hasn't before.

00:22:15: So definitely looking at marketing stuff you didn't do because you didn't have

00:22:24: the...

00:22:25: spare time or didn't have enough ROI to justify it.

00:22:30: And now you can maybe pump up the volume, stuff like that.

00:22:35: Yeah, and I think just to add to that, I think historically, I think natural

00:22:40: language processing as a field or kind of a capability on its own has been

00:22:45: historically been something that big corporations did because of compute.

00:22:50: or the technology that was required in order to develop those systems or

00:22:54: customize them to your industry.

00:22:56: It was just so cost prohibitive that people would not even just go there

00:23:00: because it's like, oh yeah, it's for the big corporations to play around with.

00:23:04: But now if you look at, and we talk about large language models right now, so if you

00:23:08: think about the amount of value and the use cases that that opens, like you could

00:23:13: use this for summarization, but you can use it to curate data.

00:23:16: You could look at...

00:23:18: using large language models for some analysis tasks and things like that.

00:23:23: So things that you were potentially looking to leverage natural language

00:23:27: processing for, but it was so cost prohibitive before.

00:23:29: You could totally revisit now because the costs of some of this compute and

00:23:35: utilizing these large language models in your infrastructure have gone down

00:23:39: significantly.

00:23:39: So for the benefit that you're gaining out of it...

00:23:42: weekly basically.

00:23:44: Yeah.

00:23:45: Yeah.

00:23:45: Yeah, so I think, I mean, definitely revisit that because I think, yeah, the

00:23:49: ROI is much easier to demonstrate now than it probably was in the past.

00:23:55: So you've done this for some time now.

00:23:57: What do you have experienced, like maybe as common pitfalls during AI projects or

00:24:05: during strategizing AI projects?

00:24:09: There are a few, I think.

00:24:10: Um, inflated expectations are unrealistic expectations of what AI can do and what

00:24:16: timeframes.

00:24:16: Cause I think depending on the complexity of the projects, it does take a little bit

00:24:20: of time to implement.

00:24:22: Um, there's also some experimentation that happens with AI systems.

00:24:26: Um, unlike again, traditional software development techniques where everything's

00:24:31: kind of, you know, rule -based, you determine the rules and then the outcome

00:24:34: is pretty certain with AI.

00:24:37: It really depends on the right technique.

00:24:39: to use sometimes, it's very unique to the use case, the industry and things like

00:24:45: that.

00:24:46: So the types of techniques or models you're using.

00:24:48: So for example, in healthcare space that I have my full -time role, we are looking

00:24:54: for equivalents of general models, but that are trained specifically on medical

00:24:59: tech.

00:24:59: So there's again, nuances to the types of models, some experimentation, just the

00:25:04: fact that you're building something.

00:25:07: means that it doesn't exist elsewhere.

00:25:08: And that's another thing that we may need to talk about, but there's a build versus

00:25:12: buy assessment that you have to do.

00:25:13: So you shouldn't just build these tools because it's quite expensive, but because

00:25:18: it is new, it means that it doesn't exist elsewhere.

00:25:21: So there's a level of experimentation that you have to do.

00:25:23: So it's not as deterministic.

00:25:26: And also look in the open source space, there's a lot of movement there, a lot of

00:25:31: things you can just integrate a fork and...

00:25:36: Yeah, adapt to your use case.

00:25:38: So that might give you a great head start.

00:25:41: What I want also to add is you don't have to...

00:25:51: how to say it.

00:25:54: Okay, we cut that.

00:25:57: I just want to say, but one other thing that I'll just add that with AI system,

00:26:08: because there are probabilistic systems, it's hard to predict the accuracy or the

00:26:13: outcome, right?

00:26:14: So it's not as like, you know, you put this in and you're going to get a

00:26:18: deterministic output.

00:26:19: Sometimes, you know, as you're kind of getting the systems up to a certain

00:26:23: accuracy,

00:26:24: you might get there with the first two 80 % or 90 % accuracy by third iteration,

00:26:31: which may take a little bit of time.

00:26:32: But even then, you're only promising 80 % accuracy of the promised output.

00:26:38: So you can't also calculate your ROI the same way you are doing it for standard

00:26:44: processes.

00:26:45: So the way that you could actually do it is incremental.

00:26:48: So for every percent of improvement that you gain with each iteration,

00:26:53: you promise an X amount of value gained.

00:26:56: And so that way you're not anchoring yourself to a hundred percent accuracy,

00:27:00: which very hard to achieve with some of these AI systems.

00:27:03: If you're lucky, you will get into the nineties.

00:27:05: But you are also kind of self -proving or safeguarding yourself so that if your

00:27:13: solution is built to 80, you demonstrate, okay, the ROI and that investment, how

00:27:17: much you put in to get to that 80 % is X.

00:27:20: So now for every 5 % improvement,

00:27:22: we are looking to project why, and that's what would make sense.

00:27:25: And when you reach that time where the invested amount of money is just not worth

00:27:33: the ROI at that point, that's when you say like, this is probably as good as we're

00:27:36: going to get.

00:27:37: And any change beyond a certain point does kind of start to level off and the

00:27:42: improvement tends to be very incremental.

00:27:44: So sometimes it's just not worth putting in huge amounts of dollars into an

00:27:49: initiative if you're only getting a small increment of value back.

00:27:52: And also, I would love to add the value is not only calculated by what percentage

00:28:01: helps me to like, for every percent, this is the value you add, but it's also you

00:28:06: have to weigh against what does it mean if it's not right?

00:28:12: Like if the answer or the outcome of your AI solution is not right, what does...

00:28:17: the consequence of it is like, is the consequence just after we start the

00:28:22: rephrase the question.

00:28:23: So what is the consequence that someone has to like cancel some bookings might be

00:28:31: might be a different beast.

00:28:33: So that and what I would love to add is also really think on how you sell the

00:28:40: stuff to your people and to the people using it.

00:28:42: Because if you sell 100 % solution to them and you

00:28:48: may not even have 70 right now, it's fatal because people will just stop accepting it

00:28:58: and saying like, does work like every third time and dismiss it.

00:29:05: So you have to sell it properly to have it implemented successfully in the end.

00:29:11: I would say just like, oh, maybe to add one more that you'll.

00:29:16: probably frequently encountered, especially if you work in larger

00:29:19: organization enterprises, but when you roll out the product, you have to set the

00:29:22: expectations.

00:29:23: Yeah, that's what I mean.

00:29:24: For the product as to what you can expect as output, as far as, you know, because

00:29:32: the output may not be perfect in every scenario.

00:29:37: And so we preface that, then we go on road shows before we ever deploy anything

00:29:41: organization wide.

00:29:43: is to educate the organization on what the solution is, how it works and the

00:29:49: importance of feedback.

00:29:50: So the solution becomes better with them engaging with the tool, with them

00:29:57: providing feedback and telling us what's wrong with a specific piece of output.

00:30:01: So it comes to us.

00:30:03: There's also a phase that we do controlled rollouts, especially for the tools that

00:30:07: are internally facing.

00:30:09: where we are more proactive and there's different phases that we do for rollouts.

00:30:14: But we seek that feedback more actively initially.

00:30:17: And then once we roll out, let's say a beta, we have other more automated

00:30:22: mechanisms like implicit feedback or explicit feedback that we collect

00:30:25: additional data points to improve the product.

00:30:28: And then we proactively seek so that when we roll it out into the practice, for

00:30:34: example, or into the enterprise,

00:30:37: people are educated as to what to expect of that product.

00:30:40: So we don't hear, although it does come sometimes like, oh, this is really wrong

00:30:45: or, but once you go back and you educate or you revisit kind of some of the things

00:30:52: that how this product works, I actually appreciate it.

00:30:55: But I don't want to underestimate the amount of education and change management

00:30:59: that needs to happen to educate folks within the organization for...

00:31:04: what's different about this product and why it's not built the same way that the

00:31:09: traditional products that they're used to.

00:31:11: From our talks, I remember that you also, like oftentimes, you have an emphasis on

00:31:21: scalable solutions and that's important for you.

00:31:25: What does this mean?

00:31:27: So, yeah, I say this a lot, but quality and scale are the two elements of AI

00:31:32: solutions that you want to

00:31:33: Make sure you have, so it depends on the use case.

00:31:36: So of course, if you're in a small organization, maybe scale is going to be

00:31:41: like scaling to five users or maybe 10 users.

00:31:44: But when you're talking about enterprises and scalability, what that means is, so

00:31:49: quality ensures that you're building something and then you have a pathway to

00:31:54: improve the solution, you know, up to a hundred percent over time.

00:31:58: So there's a roadmap for how you continue to improve, whether it's feedback or other

00:32:03: means.

00:32:04: And then the scalability aspect of it, you ensure that you're not creating a small

00:32:10: use case for a small population that could be as good as just writing a few rules,

00:32:15: like business rules.

00:32:16: So sometimes if it's a use case that you invest a lot of money, but it is so small

00:32:22: as far as impact to the organization, then it may not be worth the investment.

00:32:27: So you want to make sure that you scale it, but you also do it in a way that...

00:32:32: you can deliver value not only to one business unit, but you do it, you develop

00:32:37: it as a platform or you get a solution to a certain part where you let other teams

00:32:46: leverage the technology or kind of, you know, let them use the modular components

00:32:51: of the architecture in order to customize it to their unique use cases.

00:32:55: So,

00:32:55: It is highly emphasized and like how large organizations and enterprises actually do

00:33:00: this and how they're able to scale, not scale, but also let organization, you

00:33:08: know, deploy hundreds and thousands of models a year is because they're building

00:33:11: these platforms.

00:33:12: And then each individual business unit is then customizing to their, their needs.

00:33:17: But then also, you know, we have, I have experienced use cases where it is.

00:33:23: you know, supporting a specific workflow, but historically, because of rule -based

00:33:28: solutions that were kind of tried in the past, they were only able to support a few

00:33:36: use cases, right?

00:33:37: So like for this situation, the situation, the situation, this is how you use it.

00:33:42: So it's, it's how do you actually do it, you know, challenge the system to do these

00:33:49: things, but then a hundred other things, right?

00:33:51: So.

00:33:52: How do you actually create this engine to take the data and let the data do the hard

00:33:57: work to provide recommendations for not just to use this use case, but then how do

00:34:01: you actually create this model?

00:34:03: It's really challenging, I'll tell you that, but we have experience doing this.

00:34:06: But you have to make sure that it is something that the whole enterprise can

00:34:09: take advantage of, not just a specific exclusive group that their use case is so

00:34:14: small that it becomes a shiny object for their organization and everyone else wants

00:34:18: it.

00:34:19: So you need to build a path again, whether you democratize the architecture or

00:34:23: modularize it so other teams can use it, or you build the solution itself in a

00:34:27: scalable way so that the whole enterprise can take advantage of that use case.

00:34:31: Again, it's a case by case basis, but there's two different ways to think about

00:34:35: it.

00:34:36: Interesting.

00:34:38: Like every project has to be done by people.

00:34:42: How do you, like, what is the team one would need?

00:34:48: if you would love to do it internally, like for example, I have my own IT

00:34:53: department and I have my own people and they have to do the project.

00:35:02: What are they?

00:35:03: Like I think you have internal teams and like, augmented with some external

00:35:08: contractors, right?

00:35:11: Yeah, so I would look internally first to see what skill sets exist.

00:35:17: So,

00:35:17: Not every IT team has a data science function or data analytics that have the

00:35:25: right skillset to work on AI solutions.

00:35:28: So I would evaluate your internal team and structure and skillsets to see if you even

00:35:32: have the right skillsets to deploy AI solutions.

00:35:36: So people have MLOps or just working in ML solutions, data sciences, data engineers.

00:35:43: Do you have those skillsets and are they appropriate for the project?

00:35:46: If you do,

00:35:48: I would also, you know, the technical piece of it is just one part of the story.

00:35:52: You do have to have a business representation, right?

00:35:56: Someone that drives the vision of that product, of that function, some, the

00:36:00: metric KPIs for what, if this product is developed or this initiative is developed,

00:36:07: what's considered the success?

00:36:09: You also need some, I want to say leadership champions, someone who really

00:36:14: believes in the effort.

00:36:16: and can champion this across the organization and has influence across the

00:36:22: organization.

00:36:22: So this does not become like the silo effort of someone experimenting with AI

00:36:27: and the rest of our organization is not aligned.

00:36:31: Right.

00:36:32: And they're just kind of independently working on this own and they don't have a

00:36:35: voice to like really get the organization excited about this effort.

00:36:38: Then it could quickly get shut down as an experiment and like something gets sweep

00:36:43: them to the rug that no one cares about.

00:36:45: So you do want to make sure that you structure this team, someone that's,

00:36:50: again, a strong leader.

00:36:52: You know, you need the product manager, you know, some leadership representation

00:36:55: in this, and you do need to have a technical team in order to kind of

00:36:59: implement that and see it through, you know, proof of concept, beta deployment,

00:37:05: like operationalizing the tool and whatever else have you.

00:37:09: But I think that the key here is that...

00:37:12: You know, not everyone is very fit for innovative stuff.

00:37:15: So there are going to be things, especially with the use cases of

00:37:19: automation.

00:37:20: There are, I haven't experienced this, but I have recently read a book about a

00:37:26: specific use case where, you know, this business leader was trying to deploy an

00:37:30: automation tool or automating certain aspects of the process.

00:37:34: And, you know, she was delegating.

00:37:37: I want to say like process mapping to the target team.

00:37:40: So basically the SME is on the ground and what might actually happen is people like,

00:37:45: and what actually happened in that use case, people started ghosting that effort.

00:37:49: So people started taking vacations unplanned, calling out sick, um, and just

00:37:55: like not documenting.

00:37:56: So a month went by and there's not much movement that happens.

00:38:00: So again, it takes a certain level of leadership, then also stamina and the

00:38:05: mindset.

00:38:06: also to like lead the initiative, but also the teams you surround yourself with kind

00:38:10: of have to be in that same mindset of like innovation and not be stagnant to their

00:38:15: old ways, reserving to their old ways of traditional kind of problem thinking.

00:38:21: And then there's some, some telltale signs for when things are not going in the right

00:38:26: direction.

00:38:27: And you can tell again, by the velocity of the team, where the initiative is going,

00:38:31: how long it's taking.

00:38:34: Yeah.

00:38:34: And I have some, some of my own kind of targets for how long things like a proof

00:38:38: of concept should take, right?

00:38:40: For you to even build out an initial version of the solution shouldn't take

00:38:43: more than like three to six months, like max, even for the most complex project,

00:38:47: you need to demonstrate value.

00:38:49: So you don't want to invest years.

00:38:52: Six weeks.

00:38:54: Like, yeah, depends on the complexity.

00:38:57: Yeah, just get something out.

00:38:59: And I think, um, just prove that it works.

00:39:02: Right?

00:39:02: So like you just need an indication, some signal that, Hey, this can actually

00:39:07: deliver value and then put it in front of stakeholders and say like, Hey, do you

00:39:11: want to scale this?

00:39:12: I'm actually going through some, some of those, those initiatives right now.

00:39:16: It's like, do you want to scale this?

00:39:17: Do you want to staff?

00:39:18: Do you, do you want to see, realize the full value and the full potential of this

00:39:21: product?

00:39:22: And that's when you make that decision and then calculate the ROI for each increment.

00:39:26: And when does it make sense for you to improve it to, um, in order to get.

00:39:31: the value for, for that product.

00:39:33: That's also something that's especially, um, especially useful or not, not only

00:39:40: useful, but especially if you are at a point where you have an AI project.

00:39:46: Um, and I've said it like 10 minutes ago that you have to weigh the successful

00:39:52: implementation against the failures.

00:39:54: Um, or like the successful runs against the failures.

00:39:59: that makes it even more important to get a proof of concept running, to even get the

00:40:06: chance to evaluate how far might we come.

00:40:10: There are certain things you can do in really setting up such an AI solution

00:40:18: where you know, okay, if we improve the problem, if we improve the architecture,

00:40:24: we can go even further, but that's a baseline we can work from.

00:40:29: And I, for myself, built for our solution, I built the first prototype in a week or

00:40:38: so.

00:40:39: Yeah, that's a good point.

00:40:40: Yeah.

00:40:41: And just made like one or two things I wanted to have happen.

00:40:47: And then I said, okay, the rest I think we can figure out.

00:40:52: Let's go.

00:40:54: Did it work completely as planned?

00:40:56: No.

00:40:57: But do we have a proper working solution now?

00:40:59: Yes, we have.

00:41:01: And that's the thing.

00:41:04: For that, it's also not bad to have someone, at least for a short period of

00:41:10: time, helping evaluate stuff who has actually experience in doing AI projects

00:41:16: besides typing stuff into ChachiPT.

00:41:19: And I think that's a really great point.

00:41:22: So when I say it shouldn't take longer than three months, it's a red flag that

00:41:26: when...

00:41:26: people like teams start taking longer than that.

00:41:29: But what I was actually, cause I'm doing quite a complex project right now that

00:41:35: we're on the timeline to deliver a proof of concept in five weeks.

00:41:38: And we're talking about knowledge graphs, large language models, natural language

00:41:42: processing, and machine learning recommendation system, all orchestrated

00:41:46: under one solution.

00:41:47: And that's being delivered in five weeks.

00:41:49: So it is possible to do it.

00:41:51: And again, you have to have the right people actually doing this, but...

00:41:55: It starts to become again, it should become a red flag when things are taking

00:41:59: longer than anticipated.

00:42:00: And then you should really evaluate are the right people or the right skillsets at

00:42:04: the table kind of ideating these solutions.

00:42:06: And if not, I think you should augment your teams with contractors or experts or

00:42:12: people who have experience actually building these solutions or educate teams

00:42:17: for actually how to strategize, how to put these solutions together.

00:42:20: So I know Edgar, you do some of this, I do some of that as well.

00:42:24: How do you actually build these solutions for proof of concept, but do it in a way

00:42:28: that you could scale those solutions even further and later.

00:42:32: So pick, you don't have to boil the ocean with the initial version.

00:42:36: So you do have to define the scope for those five weeks of what you're trying to

00:42:40: show as an indication.

00:42:42: So it can be small, but you want to make sure that that effort is not throw away.

00:42:45: They can just build on it.

00:42:47: If you do get leadership buy -in to continue.

00:42:52: I would say a little note to that.

00:42:59: It can be throw away if you just want, if it helps you, if one -way camera is good

00:43:08: to take the shot right now to see if it works and then go get the expensive camera

00:43:15: afterwards.

00:43:16: So it's totally fine.

00:43:18: And for AI solutions, the prototyping doesn't...

00:43:22: have to be in the system you will implement it like the real solution later.

00:43:30: I for myself oftentimes just go ahead, mock some prompt with some mock data and

00:43:38: put it into CHPT for just to see how AI reacts to what I want from it.

00:43:45: Or use a low code solution or something like that.

00:43:50: to just get some results because results and evaluating results is like really key

00:43:57: to creating successful AI solutions.

00:44:04: Yeah, totally agree.

00:44:05: So that can be throw away, but only if it really speeds up the process.

00:44:12: Otherwise, go ahead.

00:44:13: Of course.

00:44:14: And I think, again, I'll just kind of emphasize that AI is experimental.

00:44:18: So do even data, like the most experienced data scientists or ML engineers go into it

00:44:23: knowing exactly how it's going to work?

00:44:25: Not always.

00:44:27: So there are things that they are going to be trying different algorithms, different

00:44:30: approaches.

00:44:30: So exactly.

00:44:33: So by definition, anything that they tried that didn't work is throw away.

00:44:37: But I'm just saying that you shouldn't just build.

00:44:40: a solution for the sake of building it, something that should be using machine

00:44:44: learning, but you decided to just do the proof of concept with large language

00:44:47: models because that was just something that you could get done in like two weeks,

00:44:51: but that's not the technology you end up using at scale.

00:44:55: That's what I mean.

00:44:56: And also, I would love to add something to that.

00:45:02: The jump from a prototype, which already gets a lot of the way,

00:45:09: to an actual product, don't underestimate that.

00:45:13: The jump from an actual usable product with all the concerns about privacy and

00:45:21: data security and everything.

00:45:26: Like really, really, really, really take into consideration that just because you

00:45:34: have built a proof of concept in two weeks,

00:45:36: the solution won't be ready in four weeks time.

00:45:39: So there is a big gap from, okay, barely usable to production ready.

00:45:49: And there is also, and that was, for example, my learning building our solution

00:45:55: at one company.

00:45:58: We went ahead and I did the AI algorithm in November.

00:46:06: and attached it next time now in March.

00:46:10: So that's how long it took for me to get everything else ready.

00:46:15: All the systems, all the integrations, all the UI for administrating stuff, concepts

00:46:22: for interactions, authentication, authorization, all that stuff.

00:46:29: That's a lot of work.

00:46:31: That's a lot of work just for the surroundings.

00:46:36: to then give AI the proper environment to work in, which is like the most important

00:46:42: thing, because otherwise you won't ever get a good result if you don't have the

00:46:46: right context.

00:46:49: And yeah, that's something I'd love to emphasize because I know also like from my

00:46:56: partners and for myself even like, I thought, okay, I built this in a week,

00:47:02: I'll be done in two months.

00:47:05: I wasn't, I wasn't even close to doing that in two months.

00:47:08: So keep that in mind.

00:47:10: The gap and from like, I do softener for 15 years professionally or close to 15

00:47:18: years.

00:47:20: I never had anything that was like so capable in its prototype phase, but took

00:47:28: so long to bring to production really.

00:47:31: So yeah, keep that in mind.

00:47:34: Yeah, I totally agree.

00:47:36: I think don't underestimate the security emphasis.

00:47:39: I think I would even layer that depending on whether this is B2C, meaning that it's

00:47:46: outward facing towards customers type of solutions.

00:47:49: You probably want to even build in additional time to do the due diligence to

00:47:52: make sure, especially for large language models, that it's not going to

00:47:55: hallucinate.

00:47:56: You put in the right guardrails in place, system prompts, and you anticipate all

00:48:01: kinds of scenarios.

00:48:02: So there's probably an additional...

00:48:05: level of Q &A that you want to do with smaller groups to make sure you iron out

00:48:14: all of those bells and whistles before you put a customer -facing product like

00:48:19: Chatbot into the market.

00:48:21: And again, depending on the industry as well, but things that are internally

00:48:25: facing, again, depends.

00:48:27: There's a potentially different approach that you can do with change management

00:48:30: where you would do still...

00:48:33: your due diligence of rolling it out to smaller set of users, but then the types

00:48:36: of people you roll it out initially with.

00:48:38: And then the subsequent phases, there's also an approach to that.

00:48:42: But I would also kind of add that in there as well, because again, these systems are

00:48:48: not meant to be perfect.

00:48:49: And in order to get that last level of feedback, accuracy built in is actually

00:48:56: better driven not by fine tuning these techniques, but it's actually through

00:49:01: getting additional feedback.

00:49:03: So, yeah, so it really depends, but I totally agree with you that the time from

00:49:09: POC to a production level product does take a little bit of time that could be

00:49:13: disproportionate to the amount of time it took you to get to that proof of product.

00:49:16: Yeah, definitely.

00:49:17: Okay.

00:49:18: I think this was a good closing.

00:49:25: Should we talk about what we'd love to do maybe during the next episode?

00:49:29: And if people want to take us up on it.

00:49:33: Yeah, a little surprise.

00:49:34: I was completely forgot about it, to be honest.

00:49:40: But it's in the middle of the night here where I am.

00:49:43: So have some, how do you say, pity with me?

00:49:49: I don't know.

00:49:52: Yeah.

00:49:52: What we love to offer you is you can hand in through the comments on YouTube or DM

00:50:01: us on LinkedIn.

00:50:02: hand over some ideas, stuff that you would love to try out or love to know if it's

00:50:11: working or not considering AI integration.

00:50:16: And we would offer up maybe one time and then see how it goes, but take one of the

00:50:24: suggested solutions or implementations and just do it live with you.

00:50:28: Just building live POC.

00:50:33: which you can take with you, do what you want with it.

00:50:36: I don't care.

00:50:38: It's about showing how to approach that.

00:50:42: So I'm doing that with some of my own projects live on my channel.

00:50:47: And we would love to do it in the AI boardroom context as well.

00:50:52: So if you have any interesting topics you wanna tackle, just approach us down in the

00:51:00: comments or DM us.

00:51:02: Yeah, just to let us know what maybe industry you're in and what use case or

00:51:08: what problem are you trying to solve with AI.

00:51:13: And so, and then we'll be in touch with you further to kind of get your additional

00:51:18: information, maybe documents or whatever you're hoping to achieve through this, but

00:51:22: drop your comments below.

00:51:23: We'll select someone from the comments, we'll be in touch and we'll be building

00:51:29: out the solution live with you.

00:51:30: to show you exactly how it works behind the scenes and just kind of let Edgar's

00:51:36: skillset shine live.

00:51:38: So for those, hopefully we won't do it at this late hour so he doesn't fall asleep

00:51:43: while building.

00:51:44: Honestly, I stream from 10 PM every time I stream, so it will be late at night, but

00:51:50: it's fine.

00:51:52: Okay.

00:51:52: Just make sure you drink some coffee so you don't fall asleep on that.

00:51:54: Yeah, I didn't have any coffee today.

00:51:56: That's my problem at the moment, I think.

00:52:00: Okay, yeah, thank you very much, Svelana.

00:52:04: That was really, really insightful.

00:52:07: Thank you for opening up to us about all the experiences you made.

00:52:14: And yeah, to you all, thanks for listening.

00:52:17: It was a pleasure, as always, talking with Svelana about our most favorite topics.

00:52:25: Yeah, I hope you get the value from it.

00:52:29: I've myself founded Value Pact.

00:52:31: I think you have to listen several times to get everything out of it.

00:52:36: But yeah, that's what we want to deliver value until it comes out of your ears.

00:52:43: I mean, the invitation still stands.

00:52:44: So if there's anything that you guys want us to dive deeper, maybe do some

00:52:48: diagramming live, I think we can talk and deep dive on a specific topic that you

00:52:53: want to understand more about.

00:52:56: So again,

00:52:57: Drop it in the comments.

00:52:58: We read your comments.

00:52:59: Thank you so much for again, engaging us and taking the time to listen.

00:53:03: But yeah, I hope you guys enjoyed this and we'll see you, I guess, on the next one.

00:53:09: And think, yeah, if you're not subscribed to the channel, please subscribe.

00:53:13: Oh, yes.

00:53:13: Leave us a like.

00:53:18: Yeah.

00:53:19: And comment down below.

00:53:20: We are always curious to hear from you.

00:53:23: And that's it.

00:53:26: See you.