Leading Through Change: Adopting AI in Your Organization | Predict 2024

Hello, everybody. This is Chris Wide coming to you live from New York City. Um, if you are here to hear about leading through change, how to adopt AI in your organization, then you are in the right place.

Um, to get started, just a little bit about myself. Um, I lead engineering and data science at Lonely Planet and the Points Guy, um, focused on all things travel, rewards, redemptions, um, and really how to get those, those, uh, travel experiences that are all aspirational for the, for the lowest possible cost. Um, some work that we've been doing, uh, around AI has really been focused on leveraging, uh, generative ai both to, uh, basically extract from our lonely planet guidebooks and build, um, itineraries and other custom content assets, making our content work harder for us.

And also on the points guy side, we've built, uh, within our mobile app, the ability a chat experience powered by OpenAI, um, to help people understand when points versus cash redemptions are, which option is better, uh, as well as what next best card, uh, may be for your wallet. So, super excited to talk to you about the experiences that we've had as a team, some things that we've learned, um, and ultimately helping you get ready to get started on your AI journey and or reset the journey that you're already on. First, though, I wanted to talk a little bit about the current outlook and really why this is applicable.

I don't wanna rehash 2023, but I think there are some important level setting components to the current outlook around AI and generative AI that make this a really applicable topic for Predict 2024. First of all, 77% of companies are already investing in the use of ai. 35%, uh, of those companies pulled, um, are actually already integrating AI into their production systems.

Current research shows that AI has the power, generative AI has the power to contribute up to $15 billion to the global economy by 2030, and all of us wanna figure out how we're able to take a piece of that pie for our respective companies. Every industry from healthcare to travel will be impacted by this technology. The main focus of 2024 will really be how can organizations customize user experiences, uh, creating a slicker funnel, uh, for transactional outcomes with the customer while, uh, really bringing to the table a higher degree of value exchange.

Currently, issues are resolved 37% faster and 14% more efficiently, uh, when resolved by AI versus a human actor. So in all the research, um, most companies are already going after this massive economy, economic impacts, and definitely some early use case implementations that show the real power of this technology and what it can do. Why does it matter for you?

Why does it matter for your business? Re-skilling now will help you enable, ensure you future proof your workforce, your respective teams. As AI takes over lower cognitive tasks, um, companies that ignore AI are at risk of being disrupted across any industry.

com boom, uh, that nobody believed would ever be able to fail, but they, they chose not to continue to innovate, not to continue to do the right level of r and d, or they felt like their piece of the pie, um, that they had carved out for themselves was theirs in, in perpetuity. You know, I'm thinking the Motorolas, the research in motions, et cetera, uh, of the world, faster data retrieval and decision making will lead to a higher pace of play, um, and innovation and ultimately delivery, um, fail fast, deploy more experiments, and ultimately, uh, when you are deploying more of those experiments and placing more of those betts, you're more likely to win. Inaction has historically led to the demise of many companies when new technologies emerge.

I gave a few examples, but this is why it matters right now. So, you know, if you're one of those individuals that's been looking at this over the course of 2023 and as the technology leader and you're saying, you know, this isn't gonna impact our business, you know, we're highly regulated. I don't need to worry about that.

The fact is, you should be worrying about it. You should be thinking about how, uh, and where to apply this new technology, uh, within your respective organizations. So, how to get started, I think how not to get started is to go out and just start, uh, leveraging this technology, uh, in an unstructured sort of, uh, unthoughtful way, uh, just going out and starting to play with it or unleashing it to your teams, or setting a mandate to say AI at all costs.

Um, when I think about my experiences across three major companies, uh, uh, helping to support cloud migrations, you know, it, it often starts as a technology project and really, you know, a singular executive or senior leader kind of Cadillac, um, you know, really going out and pushing for, um, sort of cloud at all costs, right? And in some cases, I've heard many people talk about, you know, they're getting AI at all costs. Um, you know, if we look back at those cloud migrations, we often just, you know, did many of them were lift and shift and really didn't yield the value, and in some cases actually created more expense for, uh, uh, the company that was adopting it.

And AI is no different. It's absolutely a game changing technology, but before you just apply it, you need to assess the current state. There's really three elements of the current state that I think are so important to understand.

Um, first of all, are your people ready for the change? That's two faceted, both from a skillsets perspective, but also just from a, a, a mindset towards ai, right? There's a level of psychological safety that needs to be in place for individuals within your respective organizations to understand that this new technology is not being brought in with the intention of taking their jobs, lower cognitive tasks being performed by ai.

Yes, but there need to be people who manage the AI systems, the underlying infrastructure, the prompt catalogs around prompt engineering, right? There are new roles, just like with the adoption of cloud. There were new roles that emerged out of that trend, out of that new technology that enterprises began to implement.

Uh, so that being said, it's mindset and skills. Um, secondly, how does data and technology stack up in your company's top priorities? Not every company looks at itself as a technology company, and I would even argue not every company fully values technology in the role that they play in the organization.

So are your executives and your senior leadership, are they bought in, in terms of continuing to invest in technology, data and ai? And, you know, if some of the interest in AI is coming from that group of individuals at your respective company, do they have a full understanding of the wherewithal it takes to actually bring AI in, uh, at scale, in a safe and compliant way? Um, and, and, and are they willing to sort of underwrite that investment?

And how have they treated other technology Investment historically can help to inform that element of the current state. Lastly, probably the most important. What is the state of your data and data engineering function?

Do you have, uh, structured data? Do you know where all of your data lives? Is it being maintained?

Um, do you have the right subject matter experts on the data that can help influence and guide as you begin to build, uh, systems around it? And do you have the right, uh, maturity levels and clear ownership and responsibility of data stewardship and governance within your organization or not? If the answer is not, then you should probably consider, uh, enabling some of the core tenets of a good data, uh, science and data engineering organization to make sure that you are starting from a place of good with regard to the foundational data that you're going to leverage for any use cases that you'll be applying AI to in your organization.

So, are your people ready? Is leadership bought in? And what is the state of your current data capability?

And the data itself? Once you understand, um, those three things and you, you are still feeling as though, okay, passing grade across sort of these pillars of readiness around the current state, the next thing you need to do is find your champion. That champion may very well be those of you out there today, uh, curious about how to get started with AI in your organization, but often, um, an evangelist needs to find a champion that's in a more senior role.

Um, you know, I think a couple of things that, that I think about when I'm trying to identify champions from respective businesses that I'm working with to implement new technologies like ai. First of all, um, the champion's probably gonna be the most critical part of the transformation around this new technology. Secondly, it's a marathon, not a sprint.

So you really need to find somebody with an authentic interest, uh, in this technology with influence in the organization, some degree of technical acumen and a broad insight of the organization, um, the TLDR and all of this, finding your champion incredibly important with regard to the early days of your transformation, that is going to be the individual that helps you get the right audience with the right key decision makers helps steer, um, you know, you through the, navigate you through the different, uh, situations that you're going to have to, uh, encounter with, uh, engaging with your CFO and, and, and, you know, figuring out how are we going to budget for this? Getting the right team members and, and essentially hiring, uh, some of the right team members and ultimately justifying the investment. Uh, this person is gonna be absolutely key, and it's not always clear who this individual is.

Um, so really important as you're putting your transformation roadmap together to, to really figure out who is the right person to champion this initiative within your organization. It's not always a technology leader, it's not always a CTO or CIO, um, and certainly, uh, needs to be, uh, somebody who has bought into, uh, both technology, uh, and it's, it's it's impact to the business, but also bought into the generally interested in in ai. Next up, you've got your champion.

Current state is a check. You feel good about where you're at. Um, you wanna identify your first use case.

A lot of companies that I talk to actually skip those first two steps go right to identifying a use case. Um, they've usually already, by the way, picked the technology that they think they wanna use, um, as their air quotes, AI tool for their respective company. Um, and, and you know, they've made some of these other decisions and then they start to think about what are the actual business problems that we're trying to solve?

Or we is in the customer, you know, maybe we're trying to build something for HR or finance. Um, you know, I've seen it time and time again where, um, you know, technology is selected. Then we, based on the limitations of that technology, we determine what problem it is we wanna solve.

We're really identifying your use case should be an early phase before we picked up any technology or started training or working, uh, with any models, you need to understand like what is the first year use case that we're going to sort of identify within our company, show value and, and move on from there. So knowing your use case before you select a model or partner is paramount. Not all models are created equally, and I'll get into that in a little bit later.

Um, next you, your use case should be specific, a business problem that needs to be solving, I want to call out that it usually just making up a problem doesn't suffice. It really hits and it's much punchier if you are able to identify. We we're, this report takes a really long time to generate, or in my case, at Lonely Planet, we have 50 years worth of guidebooks with just rich content from local experts that we could make that content work so much harder for us.

Um, you know, by leveraging AI to reimagine the different content assets that that could be shaped into, um, your use case should be quantifiable, preferably based on revenue money talks. Um, and, and this gets into the, you know, don't make up a problem. Don't go find something that is an easy solve with generative AI or machine learning.

Uh, take a real business problem that is manifested that you know, you understand who the stakeholders are, um, and bring that in, uh, and really work back from that pressure. Test your use case with, with stakeholders. Make sure it matters to the organization, right?

So if you're trying to keep yourself accountable to make sure, hey, this is the problem I wanna go solve with this new technology to really be the bedrock on which, uh, our transformation, uh, and its influences is built upon. You need to make sure that the stakeholders are, are bought in. So identifying your first use case, extremely important.

It's going to dictate so many other parts of adopting AI in your organization. Next up, I like to talk a lot about your AI transformation should be a whole company transformation. I think most companies got this wrong.

When we adopted cloud for the first time, cloud often became a tech, was a technology project, usually started within one department in technology, and eventually, um, sort of caught fire from there. And, but what wasn't happening is the privacy people weren't included. The finance people weren't included.

And, you know, you get into the 11th hour after millions of dollars are invested and your finance team realizes that you are now, all of your infrastructure costs are on the OPEX line item and not on the CapEx line item. By leaving different disciplines of your organization out of the mix, you're putting your adoption of AI and the transformation that it can bring to your organization. Uh, you're bringing more risk into the equation.

What I recommend here is create an AI work group. This is what we've done, two reasons. It creates just the, the needed transparency across all disciplines within your organization.

For example, on Lonely Planet, we created an AI work group that was representative of our content teams operations, data, data privacy technology and data engineering, for example. And because we had really every facet of the company represented in the work group, every every, every, every team was really able to participate, add value, uh, represent their challenges and where they thought AI could help them. And it really helped us validate our use cases, and in fact, brought a lot of use cases to bear that we actually didn't, uh, have in our, in our, in our first go around with, with defining that.

So cross-functional group of people should re represent as much of your company as possible. Um, and they should be sort of the team that is evangelizing, uh, providing transparent progress into, uh, what are the use cases? What are we experimenting with?

What tools are we using? Um, how can that be helpful? Uh, what ideas do the rest of the company have can really be centralized and funneled through this work group represented and really assessed for merit and prioritized transparency is the key to all of this, right?

I mean, so many team members, going back to my comment earlier on, psychological safety. Uh, when you aren't transparent with what you're trying, why you're trying it, and how you're going about it, that begins to foster fear. Um, and, you know, quite frankly, a cross-functional approach will just ensure you have more evangelists across more of your company, representing the impact AI can have, uh, within your team, but also make sure you don't have any blinders on.

Make sure you're not missing anything or missing a use case that could actually be 10 x the value of the use case that you're sort of currently working through. So, cross company, uh, rep representation in the AI work group. Incredibly important.

One place for new ideas around how AI can change different parts of your business. By centralizing all of that and creating sort of a group of accountable people to lead the transformation forward, you're helping insurance success, also known as a whole company approach to transformation. Alright, I said this earlier when we were talking about use cases.

I promised a little bit about tool selection. Uh, this is probably what I get asked about the most. Um, Chris, I know you've been working with AI and machine learning for a while, uh, doing a lot of research on it, you know, help me understand like what tools would be good for, for me in my role, or what should I be advocating for at my company?

Um, you know, when we talk about tools, there's a few things I wanna say. Uh, in terms of implementing AI at your company, first of all, not all models are created equally. Um, I'll get into that in a minute.

Uh, platform versus AI as a service. What this means is, are you, are you looking to build a platform that is enterprise grade that will allow your team members to continue to add new models, new tools, new use cases to your capability sets? Um, you know, whilst keeping all of your data, all of your pi, everything within your VPC, your virtual private cloud, or within your ringfenced ecosystem, um, or you're looking for simply like, I want AI as a service.

I wanna be able to go out to a chat bot and ask it some questions. I wanna be able to call an API, um, and, and, and really just get, get that data back from that, that API, uh, and, and then sort of leverage it in some downstream processing in, in whatever way I need to. Um, you know, are you building for a long game or just a proof of concept that matters?

Um, you know, some examples of this bedrock hugging face are, are great examples of platforms, hugging faces, uh, I would argue the leading platform for, for open source model development, uh, and deployment. And then of course, Amazon's new bedrock product that, uh, really allows you to work within your VPC. You can leverage different models like cohere, anthropics, Claude models, stability, AI models, et cetera.

Or you can directly call different model providers like Anthropic or OpenAI as part of your solution. So it's really dependent on how highly regulated are you, do you wanna make sure you have consistency? Are you a large company where you really wanna establish that platform from the get go?

So adopting something like a bedrock or a hugging face would make the most sense because you can govern and apply a certain degree of, um, you know, consistency where you're looking just to build AI into an existing, uh, platform or service that you have via an API call what you need specifically on the foundational models that you choose. Now, getting back to my first point, it depends on your use case. And the beauty of it is you can have use cases that use 3, 4, 5 different models for different elements of them.

For example, if you're trying to produce new content, you can produce that new content with a language model, uh, and then you can, let's just say using GPT-4 turbo, then you can, uh, use, uh, stable diffusion to generate some accompanying, uh, images to go along with that content. Again, you can bring those two different types of foundational models together to drive an outcome. So if you need images, that's certain models specialized in images, translations, sometimes you need to provide massive amounts of context.

For example, if you want, you know, an entire book to be consumed, you're gonna need a large, larger context window for your large language model. You know, those are getting bigger and bigger right now. 1 has the leading context window at 200,000 tokens.

Um, you know, other things that matter in terms of your foundational model selection accuracy, right? Certain models are trained on mo, different models have different parameters, and different models have different sort of, uh, training data that they have been trained on. And depending on the model is the sort of completeness or recency, uh, of some of that data safety also very important.

For example, philanthropic, uh, really their, their constitutional AI approach is sort of a leading, uh, product with regard to safe ai. Um, and then of course, data security, right? It comes back to do you want, um, your data leaving your ring-fenced environment or your infrastructure, for example, as part of, uh, the ultimate sort of end product that you're building with ai.

These are incredibly important things to know as you're picking the tools, because they all have different approaches to this resiliency. Do you wanna be reliant on a single API or a single model, or do you wanna have optionality that matters? Uh, if you want optionality and you don't wanna be reliant on an on, on an API, for example, you're gonna wanna look to some of these platforms, um, and just giving yourself a lot of, uh, I'll say, uh, ability to pivot from model to model.

And in, in some cases, even, you know, from a business continuity perspective, maybe you wanna have the ability to sort of run, uh, certain processes with, with both cohere and Andro and Claude, for example. Um, do you want your data to be used to train models? Really important.

Um, I wanna talk about this for a minute because there's clear action you need to take if you've got content on the web and you don't want it, use scraped to, uh, train most large language models and, and certainly, um, happy offline to share some content on what you need to do, uh, to prevent that. But it is an optout situation, uh, not an opt-in. Um, and then, you know, if you do want your, if you do want to you, if you all right with your data being used to train, uh, models, then you know, you don't really need to think too much about this.

But depending on the model, the mode in which you're interacting with it, if you're putting data into free chat, GBT, for example, it is training the model. Um, so just, just be aware, there are certainly patterns and enterprise, uh, additions of many of these models that you can implement and pay for that allow you to opt out of, uh, a variety of, uh, different security and data privacy components, including the training of models with any of your respective data. And again, I'm gonna say it, it was in the current site assessment, what tools you bring in, you need to do another, you need to do another inventory of your data engineering, uh, organization, what their skill sets are, what their limitations are.

You have a very mature team in this space, hugging face and their, their selection of open source models, for example, which I'll get to the open source, closed source conversation in a bit, um, are, are great, great options. So, um, lastly on this I'll say tools will vary, um, based on your people, uh, and your use case. So as part of your AI work group, it's very important once you've selected your use case to really dig in to that quickly.

On partners, there are many of them. Partners can be your cloud providers, like AWS, Microsoft, GCP, uh, OpenAI. Uh, it could be SAS tool providers that are integrating AI into their tools.

Integr, I think different partners are all at different phases. It depends on the problem you're trying to solve, the maturity of your organization, uh, you know, and, and ultimately the solution that you're trying to build. And are you building a POC?

Are you trying to build a long-term platform? These are all things that play into what partner you go with. An important thing here though, two call outs.

One is everybody is saying they're integrating AI into their solution. Make sure you understand the what and the how behind that. It's not always valuable for you to pay a premium for an AI component and somebody else's solution when you can bill it yourself.

So make sure you understand exactly how they're doing that, um, as they're marketing that to you. And a good partner will bring resources. This is new to everybody, so don't be afraid to ask for help.

I think you'd be surprised what a lot of these large companies, the programs that they have, um, to offer to help with organizations who are trying to get off the ground with, with ai. Should you train models? Should you not train models?

Honestly, the answer is almost always to begin with. You don't wanna train your own model. It's costly.

It requires a lot of precision, a lot of processing power from a compute point of view, a lot of resources. Um, you know, I will say with the emergence of smaller open source models, um, on hugging face like a mistral, uh, seven B for example, it certainly, um, is getting to a point where some of, some of those if, if high precision based purely on your content matters. The barrier to entry on this is, is, is getting lower and lower, uh, as, as days and and weeks go by.

So again, it comes down to the problem you're trying to solve in your specific use case. But don't jump out of the gate wanting to train a model on your stuff. Think about Bloomberg who is, you know, their Bloomberg, uh, custom model is now being outperformed by a number of much smaller, um, open source models, and they spent millions to do that.

So proof of thought, open source versus closed source. Um, another really important question. Um, you know, open source, some of the positives there.

You're able to work directly back to, you know, every bit of code and what data was used to train the model. Closed source, much less visibility. Um, open source is obviously just like open source software.

You're kind of free to use it in different ecosystems, um, under the open source license, whereas closed source, you know, you'll be under terms and conditions of that company like an anthropic, um, uh, Google or, or open AI respectively. The decision depends on your use case and your company's current state. As I had said a few minutes ago, uh, if you've got a really mature data science and data engineering organization, uh, open source absolutely can be an area where you could dabble in and and consider.

It's, it's got some commercial benefit and certainly a transparency benefit. Uh, if you're looking for, you know, maybe you don't have some of those things or you're looking for a faster time to market these closed source models, a lot of the work, a lot of the research has already been done for you by these organizations. That's why you're paying a premium for it.

The good news is if you build your systems right, you should have, and, and you're really looking at this from a platform perspective, you should have the ability to inter to interchange and exchange models that are in use for your, uh, platform or your, your product, um, as you're developing it, as you're maturing it. Because again, these, these things are all moving at such a clip that, um, you know, you wanna make sure you don't get any single like model lock-in as well. So it's a good point to really make that part of your non-functional, uh, requirements.

This is just a cut of data, it's probably now outdated. The whole goal of this is just to show you not all models are created equal. The amount of parameters that all these models are trained upon are massively different.

What you need is based on your UK use case. If you need highly accurate sort of generalist data, you want to go after some of the larger models. Um, and so, you know, definitely take some time to do the research on this because bigger isn't always better.

Sometimes smaller and super precise or super customized will be better for the outcome that you're trying to achieve. Data and privacy, the big thing that I want to call out here, back to my whole company approach, uh, commentary, it's really important to bring your data, your security, and your privacy teams along with you on this journey. You do not wanna surprise them in the 11th hour when you're trying to roll a new product powered by AI or integrated with AI out into production and only have, uh, your value realization delayed further because of, um, you know, issues that arise as a result of needing to implement additional controls.

Once you've validated your whole company approach, testing into it, the earlier you can identify issues, the easier it will be to correct the issue. Uh, make sure that you're sharing progress transparently and really anchoring in those, those hard KPIs. How are we impacting revenue?

Are we making people's jobs easier work-life balance better? Is engagement improving? Are we getting more done?

Are we, are we bringing in more revenue? Create advocacy and conviction through a scorecard, right? So you want to agree on your KPIs, track the scorecard, and deliver on results.

And that's really what this is all about. So just in wrapping this up, I wanna say I covered a lot of content that was on purpose. I wanted to show all of the different things that should be thought about really before we start getting hands-on keyboard with implementing a solution.

Once you have a well-defined total cost of ownership, you can really get down to those KPIs. It should be focused on making people's lives better, driving revenue, increasing organizational efficiency, and then establishing that ownership within your organization, um, and creating a culture of ongoing inquiry testing and transparency around AI in your organization. But again, it's all around anchoring those early days, early days, use cases on business problems that can have the most impact to your organization.

And then working through a lot of sequenced, uh, questions to ask yourself, what tools, open source, closed source, current state assessment, use case identification, and who the team is that's gonna be responsible. So I want to thank you all for your time. I really appreciate it.

If you wanna stay in touch, I would love that. Here's my email and LinkedIn, I think it's on the Techstrong website as well. Thank you for sitting through the talk.

Appreciate you and have a great rest of your day. Thanks.