#171 Kavita Ganesan's Guide to Leveraging AI for Business Success

 Hello, welcome to a new episode of the CT O Show with Mamed. Uh, today I'm very pleased to have with me Kavita, she's joining me live from the us. Kavita, thank you very much for being on the show today. I really appreciate it. Uh, can you just, uh, tell us a little bit more about yourself and what your is area of expertise?

Sure. Thank you for having me. And, um, So I'm basically an AI practitioner of over 15 years. I started doing AI when it was just a very research oriented topic. Mm-hmm. Um, so back then we didn't even have the tools that we have today. So even for tools like SVMs, you have to write all the code for data preparation.

So that's where I got my start. So I went from research, um, to doing more research in my PhD program. And then when I graduated I was doing, solving a lot of industry problems, um, solving AI for business type of problems like recommendation systems, uh, classification models. And this was across different industries, like working on code, working on healthcare data.

So a variety of, uh, domains. And today I'm a consultant, um, and I've been doing this for the past five years and I've also authored the book, the Business Case for ai. And my focus is very much, uh, applications of AI from a very. Practical perspective. So not so much research oriented, but more practical perspective.

That, 

that's fantastic. That's fantastic. And again, thank you very much for being here today. So just to get started, you know, um, companies that, you know, they are starting to think to explore the possibilities of ai, what would you say are the first step they should take to successfully implement AI in their organization?

Yeah. So the first step, especially for medium to enterprise large businesses, is trying to understand where the actual AI opportunities are. Because often what I see in practice is that, uh, people just come up with cool ideas and then they go after those ideas. Mm-hmm. And then it becomes a prototype, a tool, but then they don't have consumers.

Uh, because nobody really needs it. It doesn't really solve a real business problem. So I would say look for the really high impact AI opportunities, and that's what you should go after. So one of, one or two of them will be pretty easy to implement. So maybe things where the tasks have been done manually, and maybe with the use of AI you can now improve that whole workflow.

So look for those, classify those as into buckets, like this is all the problems in customer service. This is all the problems in hr. So you'll know where your competitive advantage really is going to be by just. Looking at those problems. 

Okay. So first they need of course, to do kind of an assessment of what are the areas that needs improvement.

And then try to, to apply, uh, AI to it, right? Yeah. 

And once they've identified the problems, next step is to like frame it. So actually go through each one mm-hmm. And figure out what pain point it solves. Um, Is it really worth using ai or can you do it manually or is it cheaper to just do it manually?

Because sometimes having one person to solve the problem versus implementing a whole AI system can be much. Cheaper. So framing the problem is a very critical step, and I talk a lot about this in my book, uh, because I think this is the key in either pursuing or eliminating an initiative. 

Yeah, that's.

That's great. And you know, I, I like this, like, if you really don't overcomplicate things, if you can do it with, you know, still manually cheaper, that's fine. Yeah. Now let's take the next step. So if I'm today a business and I figured out, as you said, like I frame it, I know that I need to apply, uh, here, but how do, how can I know as a company or as an organization that I am ready actually to adopt ai?

Like what are the common challenges that I might face? In the initial stages. Um, 

so I would say a few things to know whether you're ready to apply ai. You need to look at a few things. One is if you have a manual process that where you are doing a lot of repetitive work, and the data from that repetitive work is available in some form, and this process is painful.

By maybe introducing some level of software automation, it can significantly improve maybe the number of tasks you complete. So that's a good indicator that this might be a good AI problem because data seems to be there. Um, we could use some software automation maybe, or maybe not. It's ai, um, and it has been solved before.

So you understand the problem and you'll know how to measure success because you have the baseline, which is the manual approach of doing things. So when you put AI in the loop, then you can say, Hey, we are getting 60% accuracy doing this manually, but within AI system it's 90%. So there is a comparison mechanism right there.

Mm-hmm. So that's one way to look at it. Mm-hmm. The other approach is more high level and enterprise level approach, where you're creating a strategy to prepare the organization to become AI ready. Which means once you put AI into production, you should be able to repeat the process over and over again.

So you, you have AI in different parts of your company, so that is more strategic and, um, that requires looking at your data infrastructure. So are you collecting data? Are you collecting the right type of data? The format? So all of the data related pieces. Mm, 

mm-hmm. Then yeah, yeah, please go ahead. 

Yeah, in my, in my book I, I discuss, um, five elements of AI preparation.

I call it B kids. So budget, culture, infrastructure, data, and skills. So these five elements, you, is something you have to look at for enterprise level AI strategy. 

So, Yeah, great. Actually, I was about to ask you about the book, cuz you mentioned the book a couple of times, the business case for ai, like only very high level.

What are some of the key takeaways that can guide businesses in, in building a strong foundation for AI integration? 

Yeah, so common takeaway is data has to be in place. You need to look at your infrastructure. So your IT infrastructure may not be ready to support AI systems. So look at, you'll have to look into how to expand that.

How do you conduct a pilot to see which platforms you want to start using? And um, yeah, so to evaluate different platform options. Mm-hmm. And also your cultural readiness, because some companies, as you know, they're very afraid of automation. Because of true losing jobs and a lot of risk factors like biases.

So they are very weary of implementing ai. So cultural elements is also critical. And that comes with education. Yeah. And skills. Yeah. The last one is skills. You need the data scientists. You need your AI experts. So are you willing to hire these people or do you wanna outsource? So you'll have to think about how you're gonna implement those solutions.

Yeah, 

I think it's common because I remember back in days when, you know, we started to talk about digital transformation and digitization, you know, the same, I would say, frictions, uh, business start to see because of the culture, which is, I think it's very important. And the skills, as you mentioned, is something really important.

Now, f experie, like, have you seen, you know, Examples of companies that really successfully started with AI and you know, they benefit from it. Like, if I want to reframe it another way, what would be the biggest benefits that, you know, organization would see when they adopt AI in their initiatives? 

Um, the biggest near term benefits would be, For some cases increase in revenues.

Mm-hmm. So if you're an e-commerce company, the use of AI has a huge potential for boosting revenues because of the discovery and recommendation capabilities that it can offer. So it depends on the industry. And then f like from manufacturing, the use of AI also limits revenue loss because you're doing things, um, more in real time versus stopping the production line for humans to do the work.

Like, um, Quality control pieces. So even if, uh, you're trying to detect like cracks and dents and problems in like bottle caps that you're manufacturing, an AI system can take an image of that and determine, um, if indeed there's an abnormality in the cap. So you don't need a human in the loop to do that.

So that immediately, uh, makes the production line more efficient, so you get a lot more throughput. So it does affect revenues, but at the same time, it also affects efficiency. It affects employee burnout, so it, it affects multiple things. So I would say it's very problem dependent. That's 

great. Now one thing, you know, that caught my, my attention when, when you.

Well, introducing yourself. And you said like, you know, you covered from business perspective, like how is AI in the business world different from, from the academy or academic world? I would say yeah, 

that's a very good question and uh, I am, I'm surprised that it rarely comes up. Um, so the research world is more, uh, predictable.

So sometimes you know what data you're testing on sometimes you know what problem exactly you're solving. So it's very, very. Narrow and those data sets that you evaluate on, um, over and over again, come from standardized, uh, benchmarks, benchmark data sets that researchers have published. So every time a new model comes up, like from chat g PT to something more recent, uh, GT four, they're using the same, they may be using the same benchmark data sets to test how well each task performs.

But in reality, our problems are not so well defined. We have company specific data, so maybe you're working in an agriculture domain, so your data may look very, very different from what's being benchmarked. So if you're gonna trust that benchmark, um, uh, evaluation numbers, than you might be surprised to see that, Hey, it works so poorly on my data.

Why is that? Also industry problems are more messy. So there is not one model that can solve the problem beautifully. You'll have to have like, um, one model to solve majority of the problems, and then a rules based to solve the age cases, and then humans for fallback. So it's often messier and it's more hybrid rather than a single model for one problem.

Yeah, that's great. So if I want to make an analogy, you know, like when, when we used to be in school and college, we used to like study math and everything, you know, it's like neat and you know, like the graphs look, but in real life we know that, you know, when we do the experiments, like things get little bit messy.

Yeah. And I think because we, we deal with real data, it's as you said, like the real data would be like not as, The one we used in, in research, which is just great. Now, you mentioned a little bit about, you know, chat g pt and generative ai. Like of course, I, I'll discuss from two perspective, like first of all, from you with your, all your experience and you work with a lot of businesses, like where are you seeing it, like mostly used from industry perspective and you know what?

Do you expect from, you know, tools like Charge G pt, because we know like Genive AI is not only charge G pt, but what, where do you see, like, you know, the, the, the future is, is, is, you know, bringing us to, from business perspective? Yeah. The 

future's moving pretty fast. So as of now, I would say, um, idea generation is a good use of tools like Chat G p D because of its underlying problems like hallucination.

But that's also another problem cost. So a lot of the tax classification tasks or, uh, things that you would train a model to do, which now you can easily do with chat. G p D costs a lot of money. So each call to the API costs mm-hmm. The money. And this is not just, um, After deployment, it's also through experimentation that you're paying, then through testing and then deployment, and then after that for years to come, you are tied into this solution versus you're training a simple machine learning model, which solves the task maybe equally well.

But you have control over the data, what goes in and what comes out, and you can change it over time. So it's a one time. Costs, which you have to maintain. Mm-hmm. But there's a lot more control. So it depends on which way you wanna go. Are you okay with getting tied into this api, which could change over time?

The performance can change under the hood over time, so your company's performance also will become unpredictable. So you'll have to think through those risks and the costs. So, as you know, risks of. Tools like chat because it tries to generate, it's sometimes making up non-factual answers and it happens all the time.

Yeah. And also the answers are less predictable. So one time it may say A, B, and c, next time it may say something completely different. So it's less predictable than, uh, training your own models for a specific problem. Mm-hmm. Yeah. Mm-hmm. But I think will shine in, um, helping you generate data to train those.

Specific models. Uh, so you can create a lot of synthetic data using chat g pt. Uh, you can use it for things like paraphrasing. So if you need. Sentences that are paraphrase in different words or headline creation. So those things, it's safe to use a tool like CHA because it's always a human who's verifying is this, uh, something I should be using or not.

Mm-hmm. So I would say it has its users, but I wouldn't say that that's the only option right now. Like you should use generative ai. Um, so that's not the case. 

Okay. So, so I can understand from what you're saying, like Yeah, it's not, as some people mention, you know, a, a silver bullet for every, you know, use case, right?

Yeah. Okay. It has its use case and also it depends on the company's policies. They may not want to upload, uh, proprietary data to a third party api. So then what are you left with? Maybe problems where it's public data that you're using and you are allowed to send that to a third party api. Mm-hmm. So for those, for those other problems, you'll have to find different ways of solving it.

Maybe use chat PD to create that, uh, synthetic data for you, and then train your own model or, uh, just train your own model using the data that you have. 

Yeah, unless, I'm not sure if, if they would be able to come with something to run the model. Of course. Like I'm doing something very theoretical in a cheaper way locally where the data would be processed locally.

I know like it's, you know, you, you need too much compute power to do this. Um, but this is something I raised also previously was someone, you know, what she mentioned to me that it's, it's sometimes hard to do, right. 

Yeah,  

okay, great. Um, So you mentioned about the risks, but you know when people, you know, when they discuss about the risks, there's some ethical aspects. So what are the ethical, ethical consideration that businesses need to be mindful when, when they decide to utilize tools like chat, G P T?

Yeah, 

so one of the. Most common risk, like I said earlier, was the hallucination aspect. So if you're asking it to come up with answers to specific questions, the answers may look very accurate just by looking at it. But when a subject matter ex expert checks it, they may find faults in it. So factual accuracy is at stake with any generative AI type tools.

So that's the first risk. So maybe you don't want to use it to generate text, or even if you use it to generate text, there should be a way to validate that what is generated is factual. So having that check, final check in place is critical for generative AI type tools. The second is it may have underlying biases.

So let's say you're using chat to make predictions on who should get a loan. Mm-hmm. You'll say, okay, this person has this, this, this background. Should they get a. Loan or not like an education loan or not. So it may have its underlying biases. Maybe things that specific groups should not get loans or younger students should not get loans.

So whatever the biases is, and that can perpetuate to the model and you will not know of these biases because you have no control over the data that it's trained on. So you, you are less aware. Another thing that with, uh, tools like generative AI that people miss is the evaluation component. So when in when we do traditional ml, we are forced to evaluate the model in different ways.

But with generative ai, people are not just building applications over the model without evaluating how well it's doing on the task. So I don't see much of like accuracy numbers being. Generated on a specific task, for example, like question answering. So you need to have that evaluation piece regardless of whether using gen AI or just pure machine learning.

Yeah. Like this is, you know, just brought to my mind like how companies that own these large language models, they have a huge. Responsibility, I think with, with the data that they own actually. Um, and you know, like if, if as you said, like they have a bias data that might give a really bad decisions and, you know, even like, sometimes it's, it would give even wrong uh, uh, assumptions.

I don't know. Like really it's, it's something that. You know, business should be aware of when deciding to adopt. Now, like of course now everyone talks about generative ai, but is there anything like, because you, you are also close to, to, to a lot of businesses and also like you, you have your experience and you did the research part as well.

Everyone is talking about generative ai, but is there anything, you know, like you, you predict that also to come to the surface anytime soon. That also would be, you know, interesting for, for us to, to adopt. 

Yeah. I think there is a lot of research going on in common sense reasoning. Where we actually want to reason.

So right now with, um, ity and generative ai, it's not really reasoning, it's generating based on what it understands based on historical data. So it's not reasoning like us humans, so, But what would be nice is if you can apply some common sense knowledge to answering questions. So that will give, um, more factual accuracy to responses.

So that is currently in the works. Um, so you may see more and more. Common sense reasoning in the next three to five years. And also I think generative AI itself, they're gonna have some way to maybe check factual accuracies. There's research going on, on fact checking these types of tools. So that will add a layer of, I think, um, safety for us to use these tools more reliably.

Mm-hmm. Yeah. And do you see any intersection between AI and other emerging technologies that businesses will benefit from? 

Um, that could be, um, uh, I mean, I think there already is like AI and blockchain, but I, I, I think AI and cybersecurity has the biggest impact. Mm-hmm. Uh, today because we have a lot of problem with cybersecurity, high volume problems, very hard to detect.

But with AI, we can solve a lot of those, um, detection issues. So I think that's where there's a lot of potential and I think a lot of governments are focusing and investing in that. 

Yeah, yeah. Um, yeah, like, you know, the cybersecurity part and actually multiple guests, they mentioned that a lot of people who are in cybersecurity now, they want to quit within the next six to 12 months.

And now the hope is, yeah. So the hope is that AI can take little bit of their, I would say, stress and you know, all the things that they have to deal with. But again, the challenge is, Bad actors might be using also ai. Yeah. On the other side. So now it's like we have AI fighting another ai, which is an interesting thing that we need to see how it will, will, will end in the future.

Yeah. Now, um, now I talk sometimes with, with d businesses from different sides and they tell me, you know, like, we don't think this AI is for us, like we, AI is only for enterprise. Is that true? 

Um, for small businesses, AI is for them, but maybe not building something from scratch. Mm-hmm. It can be tools that are already published, like, um, we have a lot of um, text to speech, so to read articles.

I use a lot of text to speech cuz sometimes I don't want to, I wanna just rest my eyes and I want something to read to me. So, which is pretty good. Which, if you use, um, Microsoft Edge or any, there, there are lots of tools, apps. They'll do text to speech, um, and maybe speech to text, like audio from meetings to text transcribed.

Mm-hmm. So that you can then analyze what you discussed and pick out specific information to put in your proposal. So there's a lot of use cases for us, um, small businesses to use AI also and chat gpt to generate ideas. Um, Just content ideas. You don't have to use it as the final thing. But even if you ask a few questions, it can spark ideas in your brain that you can then, uh, explore further.

Um, and also storytelling. So some of us are not good in storytelling. So let's say you wanna write a bio of yourself, so you have a rough bio and you want it to rewrite your bio. Mm-hmm. So using the rewritten bio, you can then just add it and make it look, uh, more accurate. So lots of ways to use ai. Yeah.

Great. Uh, just I want to give a hint for people who would be listening to us, um, Try to, to, you know, instead of giving orders to chat, g p t try to ask questions because just a, a personal experience, um, I get always better results and as you said, try to ask it to make it in a storytelling way or try to put some, um, you know, like, uh, put some.

How they say sentiment, you know, like tell it. Okay. Tell it in a funny way or tell it in a professional way. So, uh, when you give this, uh, command, I would say in the prompt, I, I get better, uh, uh, results usually. Now we talked the Kavita a lot about Chad, G p T, but what are some of the other tools that you use yourself and you find them useful?

Yeah, I use a lot of the text to speech, speech to text for my own work. Um, uh, Microsoft has a PowerPoint designer feature within the, mm-hmm. I don't know if it's AI generated, but it's, it's really good. Like I use a lot of visuals in my slides and now instead of hiring somebody to go make those slides look better, I used Microsoft's visual designer and he does a really good job.

Um, and also, um, for blogging, I use like, title suggestions from tools like ity. Um, it, it gives me decent results, which I tweak and then make it my own. So, yeah, so this is, these are some of the ways that I do use AI in my personal work. 

Okay. Nice. Yeah. Just to, to, to add to what you said, I use the designer as well, but if you are, uh, into using some tools like Canva for example, so Canva and if you need to be on the paid version, and I'm not, um, affiliated with Canva in any way.

I paid them actually, so they had now a. They call it AI assistant, kind of mm-hmm. So if you want to design a slide, you put what you are trying to do, and then it generally is like five to six slides for you, and then you can complete from there, which is excellent. I use, um, and again, I'm not affiliated with any of these tools guys, so I use Jasper ai.

Jasper ai also same as Avita mentioned. It can give, you know, some blog posts and on, you know, titles and these kinds of things. So, uh, I use it as a tool now. One of the things, and I'm asking a professional because people, and maybe we repeat it multiple times on the show, but really, do you think that AI would replace and eliminate a lot of jobs and, you know, what should we do in, in, in order to, to save the our jobs?

Like what, what's, what's your thought about this topic? 

Yeah. I think if your job is so specific that you're doing one specific task, like maybe you're just inspecting bottle cap, uh, cracks and dents, then it's highly likely those jobs are going to be reduced. So if the, maybe you have 500 jobs now, you may have just 10 of those jobs, but your role then becomes something different.

So instead of being that QA engineer who sits and does that work? You are going to become maybe the data generator for those systems because AI systems, they can't learn on their own. They need good quality data to make, um, high quality predictions. So you can become the data, uh, generators for the systems.

You can become the quality assurance managers for AI systems. So let's say it makes a lot of mistakes. So you have to go in and debug. Why is it making those mistakes, right? Then feed it back into the model and then the model gets retrained. So, So you're going to be working closely with this AI systems rather than doing the work yourselves, which is probably a good thing because humans tend to get very distracted, um, and bored on repetitive tasks.

So if you have a software mechanism helping you, that's already a good thing. So even when I do my research, I always wish there is a tool that can do some of the groundwork for me, and I just go through that results because that saves me a lot of time. So I think our workflows will be more ar augmented and we will become, uh, more the people who are working in those roles will become the quality Azure managers and, and they'll have roles around these AI systems.

Mm-hmm. Do you think we will see AI in, in a, let's call it like decision making role or managerial role? Can, can we rely on AI to, to have, um, you know, decision making roles? 

Decision making on specific tasks like, does this, um, item have a defect or not? So like specific tasks, uh, but I don't think it's going to do the work that a manager does because managers not just doing specific decisions.

They are, uh, solving problems between team members. They are corresponding with the upper level management. So there's a lot that humans bring into the picture that AI systems cannot just replace. Mm-hmm. So all that collaboration, the emotional aspect, uh, conflict resolution, so all of that, I don't think an AI system can handle well.

And even if it does try to handle, uh, it's going to take you down some rabbit holes. It's gonna use some therapy tools and say you need therapy or. Something that it's learned from data and, yeah, 

yeah, exactly. So, so this is what I always tell, you know, anyone who asks me, um, like, don't be scared because AI needs always actually someone.

To give the prompt, right? So now mm-hmm. Because people are now familiar with this term, prompt, uh, chat g pt. Yeah. So, but it applies to, to AI engineer and, you know, better than me, of course, like AI is like, think about as a black box and you have an input and you have an output, so, mm-hmm. Always the input needs to come from a human, um, right.

Like, and, and we need to give the orders to get the results. And, you know, this. Completely. Even though autonomous systems, and correct me if I'm wrong, Kavita, even an autonomous AI system actually needs also, you know, the spark or the ignite of of it. It comes from a human right? 

Yeah. The initial data generation has to come from some human process, so which has already been done.

The data is. Still very essential to these AI systems, and we generate that data. 

Yeah. , are we living a hype or is it a moment, you know, the, the AI moment? Like what, what do you think? 

I think it's both. So we have a lot of hype. That hype is good thing because it has caused a lot of companies and leaders wanting awareness.

So what I've seen is that, um, a lot of people who read my book, they started off with generative AI chat, g pt, and then they wanted to learn more about this whole space of ai. And then they read my book. Then they learn, oh, it's not just that simple. We need to think about all this risk, we need to think about data preparation.

So it got them into ai and thinking about AI and now thinking through strategies, uh, whether it's enterprise or small businesses, they are thinking through strategies. So I think, um, it's a mix of good. And a little bit of that. 

Yeah, that's good. That's good. Per personally, I'm, I'm, I'm with you. Like, it's, it's both because AI is not something new.

It's something that has been mm-hmm. You know, in the, from last century actually, the research has started, but yeah, I think what chat g PT mainly did it, brought it to, to the. You know, to the, to the scene more to the forefront. Yeah. Yeah, exactly. And, and this is where everyone now start to talk about ai, but yeah, I, I, I, I agree a hundred percent with you.

Now, Kavita, I have a very, you know, famous last question. Is there anything that you wished I asked you? 

Um, yes, actually, I was thinking about the last question. Um, so I was talking about the hype, right? So, mm-hmm. With the hype call comes a lot of dangers with AI and the dangers. Not that AI systems are gonna become super smart and take over all humans.

Mm-hmm. But it's in the misuse of this technology. So people who don't quite understand how it works, they start building applications on it without understanding the risks. And that has a lot of downstream problems. It is gonna have a lot of downstream problems. And we've already seen this in a case where, A, a tool like Chad PT took somebody on a rabbit, like I said, in a, on a rabbit hole through therapy based, uh, talk.

And that person ended up taking their lives. Um, so that kind of danger, um, will become more. Prominent. So I, what I, what I suggest is that people think through the risks and think through the evaluation aspects of ai. 

Okay. That's really informative. Yeah. Uh, well, Kavita, thank you very much for your time today, uh, on this episode.

I really appreciate the time and, you know, the, um, information that you provided where they can find more about you and the book. 

Yeah, so my website, so www.kavitaesen.com and uh, I, I guess you'll add that to the show notes, right? Yeah, mm-hmm. And also my company website is um, open-analytics.com. 

Okay, that's great.

So I will make sure that this will be in the episode, uh, description As you mentioned, thank you very much for the time today, and as usual, this is how I, uh, end my episodes to the audience. If you have any questions, any feed feedback, you know, you, as Kavita mentioned, you can reach out to her. Like you can learn about the book and, you know, uh, more about her work.

If you have questions about, you know, this topic or. The show in general, you can reach out to me by email, LinkedIn or Twitter where I'm most active. If you are interested to be a guest like Avita was today, you can also reach out to me and we can arrange for that. And, um, as usual, usual, I hope you enjoyed the episode today and until we meet and then another episode.

Thank you very much. Bye-bye. 

  📍 hit that subscribe button, share the show with your tech savvy friends and fellow entrepreneurs, and leave us a review on your favorite podcast app.

Your support means the world to us.