| If you don't have an account, Register here |
Moderator:
Joel Martin, Executive Research Leader, HFS
Panelists:
You can listen above or watch this HFS Videocast here:
Session Description:
Generative AI is poised to disrupt how customers, employees, and technology interact. We’re moving beyond bots focusing on discrete problem solutions toward contextual interpersonal discussion and co-creation. Are solutions like ChatGPT ready for prime time? Or is it all just hype?
In this session, we will debate the hype versus reality of whether generative AI can have a profound impact on business. We’ll even invite ChatGPT to the conversation. Our topics include:
You can also listen on:
• Apple Podcasts
• Spotify
• Google Podcasts
• Anchor
This transcript was auto-generated from the original recording and lightly edited for readability. We've done our best to catch errors, but names, technical terms, and company references may be misspelled or imperfectly captured. For the definitive version, please refer to the original audio or video. Views expressed are the speakers' own.
Thank you all for being here. This is a really cool discussion. It’s going to be a mix of end-users, startups, hyperscalers, and services partners, all discussing why this is such a hot topic. We’ve brought it up so many times, but this is a real chance to exchange ideas and, as Elena said, peel back that onion. So with that, I’d like to invite up my panelists. Jo Debecker, who’s joining us from Wipro, flying all the way from Switzerland — so we’re getting a very global view on this tonight. Jo runs the FullStride business, so Jo’s coming up. Steve Dunn, from Sumitomo Bank — what is a bank going to do with something as disruptive as ChatGPT? So Steve, come on up. Following Steve, an upstart, a challenger, an innovator from Austin, Texas, Cyrus Khajvandi. Cyrus is doing some really interesting stuff, even working on tools we’re going to roll out to our customers, so I’m really excited to have Cyrus here bringing the view of somebody coming from the roots of innovating. Next up, Sibu Kutty, Sibu from AWS — coming in and really talking about what a hyperscaler is doing: how we’re taking the data, the process, the change of architecture, and building that into what we’re bringing to market. Vikrant Karnik, or VK, joining us from Genpact Business Technology Systems, diving into how we’re building the business operations we were just speaking about in the last panel. And finally, last but definitely not least, Sheri Sullivan, a partner from EY. They’re doing some really cool stuff that’s crossing the chasm from the democratization of information to the monetization of information. And then a secret sneak guest, the proverbial empty chair. We’re also going to ask some of these questions of ChatGPT, because while we’ve talked about the technology all day, what we haven’t done is actually hear it from the bot itself. Is it really that smart? What’s it going to do to challenge our panelists? So we’ll have some good, interesting dialogue. It’ll be question-generated, and at the end there’s a little twist to the Q&A session that I’m going to throw out, just because we’re about to go to lunch. So with that — I think everybody’s seen this model. This is a Dunning-Kruger curve that my colleague Ralph Diaz put on the internet a couple of months back; I’ve had inquiries from professors to use it, and it’s driven a lot of conversation. Ideally, if we do this session right and exchange some great ideas, we’ll find ourselves moving out of that trough, because that’s where we are right now, between two and three. We’ve gone over the “wow”; everybody’s thinking about it, asking a lot of questions, as we’ve heard throughout the day. But they’re not quite to “got it,” and definitely not to “I’m ready for this.” So be ready to ask questions to these folks over lunch, and with that, we’ll move on. My first question actually goes to Steve. Coming from an enterprise perspective — not a provider looking to make money, but somebody who might spend money if you can figure this out — why is this so hot right now?
Thanks — I think it’s a great question. Is it hot right now? In my role — I lead innovation for the bank across the Americas and Europe — just in the last two months I needed to duplicate myself just to keep up with the questions, the interest, when are we going to set up the sandbox, all these different things. So it definitely is a very hot topic, and with good reason. There are plenty of reasons why it’s great, and depending on the type of organization there are various angles. But for us, I see three core reasons why this is something we really need to jump onto. The first one: GPT has given us access to very complex and powerful algorithms that until recently needed an army of data scientists. When you think about it, a lot of organizations struggle with how you cross that chasm, how you actually take advantage of AI. All of a sudden you’ve got this bit of a black box, but you now have this package you can bring into the company — access to a neural network that you can do almost anything with. That’s the first one. The second is that with GPT you can provide almost any prompt and get back a human-like response, which is just mind-blowing. All of a sudden we can talk to very intelligent machines and have conversations we haven’t had before. I mean, who has called up a help desk, gotten a chatbot, and hung up after two or three redirects? I do it all the time. And the third one: the model powers fundamental building blocks for the future of AI systems — whether it’s text output, multi-modal, summarizing, or creating new code, it’s an all-round player. So when you bring all three of these together, that aspirational target of an AI-first company, I think today it’s actually in sight.
That’s very interesting. So, first question to ChatGPT itself. ChatGPT, why are you so popular right now?
I am very popular right now because of my ability to create high-quality, human-like language and content. I have a wide range of potential applications. I can write news articles and generate product descriptions or replace existing customer service chatbots — some say doing a better job with expressing empathy and delivering accurate information. One of the main reasons I am popular is the recent advancements in deep learning algorithms and computing power. These advancements have enabled the development of larger and more complex models able to process and understand vast amounts of data and produce high-quality output. Another reason for my popularity is my ability to learn and improve through continuous training on large data sets. This means the more data my models are trained on, the better I become at generating high-quality output. Finally, my rise in popularity can also be attributed to the opportunity to revolutionize a wide range of industries, from healthcare to entertainment. There is significant interest and investment in my technology, which is driving my development and adoption at a rapid pace.
It’s really interesting to get that input, because through prompt engineering on my part, asking the bot to answer that was curious. I literally typed in, “You’re on a panel talking to an audience of service providers, software vendors, and enterprise users. Tell me why you’re so hot right now.” One thing I took out of that — or two things, I should say. One was she mentioned empathy; I thought that was an interesting thing to bring up. The other was, “I’m very hot right now, but you need to work on me, continuing to train the model.” We’ve touched on that a little bit, starting with Phil’s keynote this morning, but even in the data session that Reetika ran. So I’ll open it up to the rest of the panel. Training, empathy, echoing what Steve talked about, which — unprompted, because none of these folks have heard the ChatGPT response yet — is that it definitely augments the skills of my people, leveraging a large data model to surface information. But give me your thoughts, VK.
Well, we all heard why, right? But I also think you’ve got to look at it in a time-horizon perspective. One thing that has struck me is how quickly people have adopted it. Nobody has really questioned it. Normally, when a new technology comes in, there’s a lot of, “Is this really true or not?” I think that has helped it become so hot, because people can see that the answers are true-ish, something I can believe in, and so they’ve adopted it more. That’s what I think.
I agree. I think why it’s so hot is — well, how many people in the audience have actually played around with it? And you actually know how to do it; you don’t have to be trained. Immediately, like me, you started asking questions you’d ask people who work for you, or around your subject-matter expertise, where you could look and say, “Is that a correct answer, or is that a BS answer?” I think that’s why it’s gotten so popular, because normally we’ve not seen a tool like this.
If I may — when you play around with it, Joel, it gives a very human kind of impression, and at the beginning you almost think this is general artificial intelligence, that it can do anything, which is not the case. But it did spark all the brains and minds everywhere in the world to see where you can apply it, and from there we realized the applications of this technology are almost endless. In my business — I run a FullStride Cloud, infrastructure, applications, a real managed-services provider — it’s very pervasive. It’s everywhere, in every portfolio, everything I deliver. I can see real use cases for this.
I think it’s a human desire to be better, to always aim to improve. This is a real manifestation of that, starting from mechanized cotton spinning to steam engines to dot-com to cloud, and now this is much more real, much more evident. That’s why the adoption is so fast.
It’s so easy to use, and we’re at a real inflection point where AI, specifically GPT, is approaching human, if not expert-level, proficiency in reading and writing, and that’s what makes it really transformative. However, using it in the past — even before it came out, just playing with the code and building a version of it before it was developed — it has a propensity to give inaccurate answers. That’s primarily because it’s trained on an enormous, generalized pool of data, not specifically for your own needs. That’s why there needs to be a solution that’s evidence-based and grounded factually in your data, and I think that’s the next iteration, just beyond the horizon.
Yeah, those are all great points. The one I would add is this: it’s got puppy-dog eyes. It wants you to engage with it, and it makes it fun for you to teach it, so it’s learning at a tremendous rate. No other site has reached a billion users as fast as it has, and it’s learning from all of us. That’s why there’s a certain responsibility for us to teach it to do the right things. There are a lot of people who aren’t going to teach it that way.
So now that we’ve gotten “why is it hot right now” out of the way, let’s get to the meat of this conversation. How are we going to take this world from democratizing information, which is really what we discussed, into — so what do we do to provide a service? What does that look like? So my second question: how do we move from democratization to monetization? Jo, why don’t you kick us off?
Thanks, Joel. When it comes to monetization, there are a few elements. One, like we said, you need an enormous amount of data, which we have. Two, you need an enormous amount of processing power, which we have. But you also need to recognize data can be biased, so you need supervision, you need training, you need to learn the model, and you basically need to make sure you follow IP compliance and the privacy rules of the area where you want to monetize it. If you look at the total addressable market, there are a lot of numbers.
Lots of numbers. I’m sorry to disturb you — she’s impatient.
So if you look at the total addressable market, it’s 8 billion, and it’s going to be 100 billion in 2030, and I think those numbers are wrong, Joel, because we’re at very early stages. Clearly, the market is there, and compared to something like the metaverse, you can see this is real — you can really apply it. Look at the amount of data we produce as a company, and as users, every day. There’s a real need for data synthesization. There’s democratization of content creation. There’s a need for enormous processing power, which has also been democratized. And with the TikTok generation, everything needs to happen now. So you look at all those things and realize this will create a huge amount of business for us. When it comes to monetizing, there are a couple of things you take into account. One, do you need it now? Two, we talked about data and inaccuracy — does it need to be correct? Three, does it need to be trained, does it need supervision? Four, is it compliant, is it following the IP rules? And five, do you need supervision, do you need training, and all this processing power? Those are the elements, at least, that we check when we’re playing with the technology, and like I said, it’s everywhere. I basically see two ways of monetizing it. We talked about front office and back office earlier. One is around the back office, which is the augmented part of this technology — how can I augment my current services and my people? Things like the contact center, the service desk — an agentless service desk is no longer a dream. And then there’s a part around creating additional business, so for us we have a separate data and AI business unit, and we sell this as a consultancy and as a service. It will be both on the augmentation and on the additional business, both in the front office and the back office. And my view is that — I was on a similar panel with one of my vendors, and they told me there are already a couple of hundred large language models in production today, actually almost a thousand. So like cloud, this will become very pervasive. Based on the use case, you need to use the right LLM. And it’s going to be everywhere. Like cloud is everywhere, AI is going to be everywhere.
Interesting.
How long do you have? It’s going to be a longer answer. Somebody in the earlier panel, and Phil, talked about the S-curves. If you look at technology and innovation adoption in enterprises, starting from my earlier point — mechanized spinning, steam engines, dot-com, cloud computing, mobile — the productivity gains happen when that innovation is diffused through the organization, and the speed of diffusion has rapidly come down. It took about 40 or 50 years for the industrial revolution’s productivity gains to be realized, from about 1800 to 1840 or 1850. It took less than 10 years for cloud computing to be normalized. So it’s going to take maybe less than three years for ChatGPT and generative AI to take hold. How do we then move from that diffusion point into monetization? I have three takeaways. One, the time to monetize has rapidly reduced in every S-curve. Two, the early adopters almost always make the productivity gains, and the monetization happens — 90 percent of it is done by the early adopters, even though they make mistakes and stumble along the way. The first implementations are always bad; there’s organizational change that needs to happen. We saw that in cloud computing — I come from a hyperscaler — when cloud adoption happens. We looked at the financial statements of about 5,000 companies and their tech spend, and cloud investment can almost be a handy valuation factor. You can pick that and say, “This company is going to generate returns.” Like any technology innovation, the return on investment is outsized, and this is going to be huge, and that’s where the change happens. As we talked about earlier, it starts with early adoption, and that’s when the monetization happens to realize the productivity gains. There’s organizational and structural change that needs to happen for monetization to happen. And — I live in Chicago, so I’ve got to bring the blues into the conversation — I compare this to the classical orchestra versus the blues or jazz. In classical music there’s a top-down composer, a hierarchy of instruments. In jazz it’s much more of a smaller team with a lot of improvisation. We need more jazz teams. That’s how enterprises should look at this technology, and that’s where the adoption curve will improve and monetization will just happen as part of that, because we don’t know yet what this can enable. Back in the day when we started cloud computing, some people thought it was just infrastructure outsourcing, but it was a big shift, and this is similarly a huge shift in thinking. So if you don’t start now and play with it — and I’m glad everybody is playing with it — it’s going to drive some of the monetization incentives.
Interesting. Other thoughts, Steve?
I’m just going to piggyback on that comment. I’m a former surfer, so there are plenty of times I’ve taken a wave and crashed on the shore. Here, you’ve got to be careful of the height. Obviously there’s a lot of potential, but being in innovation and being a product person, for me it’s about coming back and really understanding what we have here. We’ve got to go back to fundamentals: we’ve got a large language model, lots of training, the ability to try different prompts to then fine-tune the model based on domains and the rest. For me, what you need to do is come back and understand what that is, but take more of a design-thinking approach and say, “Let’s run a business canvas and work out what the opportunity looks like, who’s the client we’re going after, what’s the value to that client, what’s the cost, what’s the potential profit?” Very quickly you’ll come up with easily half a dozen opportunities that are low-hanging fruit, and from there you have a data-driven approach to picking which one. Otherwise you’re going to be lost, deer-in-the-headlights, wondering where do I go, where do I start, someone’s calling me over here, this stakeholder over there. If you take more of that design-thinking, client-centric approach, you’re going to have a much more targeted approach, and then focus more around outcomes and value.
I think that’s interesting. There’s an aspect right now where we’re seeing monetization — how many tokens are you using, how many pages are you looking through, how many searches are you doing? Very easy ways to experiment with monetization. But I’m curious about value creation, and both Cyrus and Sheri, we’ve had some great conversations around thinking about it from creating value. So, thoughts — things you want to add?
Sure. From what I’m doing, running a global business in 159 countries, we really entered this space to disrupt it, so we’ve been looking at disruptors. My subject-matter expertise is payroll, and if you think about payroll in 159 countries, there’s a lot of data, and we’ve spent a lot of time talking with clients about the pain points. When ChatGPT first came up, and of course our team was all playing with it, we thought about how we could actually use this to cover some of the biggest pain points of our clients. We had three use cases. The first one we’ve taken to POC and then to a pilot, so we haven’t launched it in our product yet. I’m talking payroll, so you’re talking about PII, and you’re talking about 159 countries, so lots of data-privacy laws and all kinds of things. One of the first things we wanted to do was figure out how we could isolate what it’s going to look at, and control it, so we can see what it looked at, what it researched, what it’s showing us, and follow that trail. And we did that. We pulled it up using Microsoft Azure OpenAI, and we said, “Where are you, and why?” We know all the regulations around the globe — we actually publish them — so we took some key data and made it point to each one. Why is that important when you talk about monetization? A lot of our clients have a big issue: payroll touches every employee. Things are changing rapidly with flexible work, people working from anywhere, all different types of new benefits, and laws constantly changing. So what do you do to address their questions? We know the level of satisfaction with any kind of call center or employee inquiry is still very low — any research you look at shows how low it is. In fact, employees get so upset they just stop asking, so you see that curve drop off. So we asked, can ChatGPT, if focused on the right information, actually answer those questions correctly? And we did it. From a linguistics perspective, it completely blew our minds. I think the only language so far where it’s not very good is Punjabi. Besides that, it’s really good. The first week, it was only answering things like 63 percent correctly in Hawaii, because our team wrote the information — we’re the ones checking the information and the responses. Then by the second week, because of the prompt engineering, we got up to 93 percent, and we were like, “Wow, this is amazing.” So from a monetization perspective: what if I went to my clients and said, “You have an offshored or captive HR shared-service center answering questions around payroll, and I can make 70 to 80 percent of your questions go away by providing better answers at their fingertips — what would that value be to you?”
That’s great, and it echoes a lot of what Steve and Jo are also talking about — augmenting existing services, really revolutionizing, taking big steps into the unknown. Cyrus, from an innovator, from somebody running a startup you founded less than a year ago, what do you think? What’s your take on this?
As I mentioned, the moon-landing moment — whether or not, and some people still don’t believe the moon landing occurred, some people have a great amount of skepticism about how transformational an AI agent can be — but time will pass, and it will actually be one of the major things that changes the world by 2030. When I developed Humata, it was to solve a huge pain point. What is Humata? It’s a ChatGPT for all your files. As you know, ChatGPT has a pretty strong proclivity to answer questions inaccurately, and I was trying to solve a personal problem that stems back to my undergrad days at Stanford, where I was a researcher and we were inundated every day with new scientific literature. Each publication was super long, dense, and technical, and because of the advent of AI approaching human-level, expert-level proficiency in reading and writing, I decided to couple the two with Humata and just ask the system, “What is this paper about?” and get the answer instantly. When I launched Humata in early February, it was specifically for my friends who are postdoctoral researchers at Stanford, some of them at Calico, which is Google’s biotech arm, and then it just caught fire. I was astonished by the plethora of use cases across many different industries, from finance to R&D, and it’s wild to see people using it now to answer questions across their entire organization instantly. I keep hearing from them, and it’s like — we have more data than we know what to do with, and it’s like sifting for gold. The way organizations actually manage that data is person-to-person interaction over coffee or lunch, where they have fragmented pieces of knowledge, they connect, and then they make the action item or business decision. That’s the issue I was trying to solve for research, and now what I’m noticing with upcoming enterprise POCs is that they’re using our technology to ask questions across their entire organization and connect that data. The funny thing about this question is the implicit understanding that with ChatGPT they don’t respect your privacy or your data. You’re not the product — you’re the product; you’re training their general-purpose technology to become more proficient at giving big, broad answers. But what the future holds, and what we’re really excited about with Humata, is actually giving you that privacy: privacy by design first, and also training a domain-specific LLM for your company that knows what’s happening and gives you answers to your questions instantly, instead of reading a ton of documentation or papers, which was my frustration as a researcher.
That’s great. I want to switch, because we’ve touched on this next question. If we can get to the next slide — there we go — the employee experience. This is something we’re all thinking about as business leaders, whether you’re in a service provider, a tech company, or private industry. What does this mean? Are you going to fire half your workforce? Are you going to replace low-skilled or even mid-level managers with this technology? VK, what are your thoughts? You do a lot of work in this space, and you and I have had some great dialogue. What are your thoughts here?
First, let me give the typical consultant answer: it depends. And I think it’s in the context of a time horizon. In the six-month to 18-month time frame, generative AI will be like digital prosthetics — we’ll use it to augment ourselves. And frankly, if we see the value, more and more of our employees will use it, so how do we design it so we use the prosthetics the right way? But after 24 months or so, as we train the models more effectively, there are certain jobs that will just vanish, and that will put the onus on us — as the previous panels discussed — to upskill our employees so they can use these digital prosthetics the right way, and eventually drive more value.
I wonder what ChatGPT thinks about this. Where is it going to impact employees, and will it replace them?
I’d like to build on VK’s comments. The impact of AI on the employee experience will depend on how companies choose to implement and use this technology. In some cases, AI may lead to job displacement or the automation of certain tasks, which can negatively impact some employees. However, in other cases, AI can be used to enhance the employee experience and create new opportunities for career growth and development. There are several ways in which I can impact the employee experience. One, through task automation, improving job satisfaction and creating new opportunities for skill development. Two, in training and skills development, helping employees acquire new skills and stay competitive in a rapidly changing job market. Three, by augmenting decision-making, improving the quality and speed of decisions, leading to better outcomes for the company and its employees. Four, by offering predictive analytics, identifying potential areas of improvement and supporting proactive interventions, leading to a more supportive and fulfilling work environment. And five, through the displacement of jobs, replacing certain jobs or tasks currently performed by humans. However, it is important to note that this is not always the case, and that AI can also create new job opportunities in areas such as data analysis, programming, and AI development.
I think that’s interesting, and here’s why. One, because — again, VK didn’t know the answer she was going to give, and it was nice that they synced. But I think the challenge ChatGPT has posed is that it’s up to us to think about what’s next. How are we going to go beyond just the way we work right now? Thoughts on how all of you are running pilots and seeing how this is impacting your teams? How is it changing your thoughts on the employee experience?
I think, Joel, it will happen automatically. Right now we’re in the augmentation phase, then we go into replacement, and then I personally believe it will get into a generation phase as well, where it generates more jobs. I’m a big believer in creative disruption rather than creative destruction. There are a lot of examples in history — cars taking away the people who took care of the horses and carriages, and in the end there was a car industry, roads, construction, and so on. Eastman Kodak went bankrupt in 2012, and then we had iPhones, we had pictures, we had the influencer economy. I think it will be the same here. I do not believe it will destroy jobs. It will be disruptive to an industry, disruptive to a service, and then it will create more afterwards. When it comes to employee experience, where it augments a lot and automates tasks, I think it is actually going to improve the employee experience drastically. There will always be a component of taking work out, specifically in my business, but those who remain will definitely have an increased employee experience. Microsoft, for example — now, when you do a meeting, you can have speech-to-text and then ask it to summarize the meeting into a five-page Word document, or into a slide deck of three or four bullet points. So effectively, if all goes well, we all need to go to fewer meetings.
OK. Other thoughts?
I’d love to go back to my correlation with the classical music and jazz. The classical music teams are usually hundreds of people; the jazz teams are usually tens or twenties. But the good news is almost all good jazz artists are classically trained, and every classical performance has some kind of jazz infused in it. So it’s a mixture. You’re not going to get some skills — there is going to be reduction — but there’s also going to be some differential skills that we can gain, and that’s where I probably differ a bit with Steve. This is the classic seller-versus-buyer argument. You need to enable your teams to start doing those experiments in-house, and, like Sheri said, as long as it’s kept private and secure, there are methods to do that. We offer our own models to do that — enable those teams to do those pilots, and don’t constrain it so much, waiting for the perfect guardrail to appear in the enterprise.
So it’s interesting. What I’d like to think, and what we should all be doing anyway, is reducing the amount of non-value-added activities people are doing, and really allowing them to spend more time on the value-add. Imagine if you had reliable data — and obviously we’re just getting into how much of it is hallucinations, how do you deal with PII, there are lots of questions, so we’re crawling first, then walking, then we’ll run. But imagine: my organization is all about building a better working world and asking better questions. Imagine if we hired individuals with very technical skills and said, “Great, here’s our purpose, building a better working world — what are the better questions you’re going to ask to figure out how to do it? And can you get that data quickly by working with a tool like this?” Then you’re thinking, with the human at the center, about how you actually put together a product or a service that would interact and make a big difference. From a birth-rate perspective, we know we don’t have enough births to sustain us as long as we’re living; we’ve heard what’s going on with the economy; there aren’t enough workers. So I think this is coming at exactly the right time.
OK, Steve, I want you to finish this particular question now. How many people are you going to get rid of because of ChatGPT?
I think you know you’re being videoed, you’re being streamed, right? So — it depends. Look, a lot of these conversations are coming from the top of the house, and I think Sheri also touched on it: it’s about engagement. We’re so early on in the journey; we need to look at this journey from the employees’ standpoint and then co-create what the future looks like. That’s the way we’re going to deal with the ups and downs, and that’s the way you’re going to bring employees into conversations around upskilling — where they want to be, what their skill set is today, and what the opportunities are. When you layer in mobility within organizations, all of a sudden you’ve got this massive, multi-tiered roadmap where you can actually focus more on the creative opportunities, rather than the other way.
Yeah, I have something to say about this, running a world-class team of two — I thought it was three now. Yeah, we just added a guy on Monday, after Phil, we got a big investment. So the two buckets, the dichotomy I look at, are employers and employees. Start with the employers: huge benefits, because now they’ll be able to have a transformational AI agent access all their knowledge and give them instant answers. That’s the advantage Humata is giving our organizations working with us today — they’re able to make much better decisions much faster on their own private data. And the employees — that’s one of the key negative externalities of innovation, that it will cause tremendous displacement. And that is where the opportunity resides with the employees. Even doctors, general practitioners, will be replaced; anesthesiologists will be replaced, I think, around 2030. And that’s an opportunity, because they’ll have the ability to reskill and upskill to another domain, either adjacent to their own or completely radically different, because AI is going to be able to distinguish and find patterns that not even the best anesthesiologist or pathologist can discern from raw data. So there’s an opportunity in that displacement — very positive for employers, and for employees, they have a more interesting path ahead of them.
You bring up an interesting point that really feeds into the next question, and you brought it up several times — in fact, most of you brought this up. I’ll paraphrase it, because it’s been a theme going on throughout the morning when we talked about data: data without trust is junk. And I’m not going to let a robot be my anesthesiologist if I don’t trust that it’s looking at the right data. So one thing I’ve talked with all of you about: how are you going to turn this information from the democratized searching of the web — and then trying to mad-lib it together into a bunch of different sentences, hoping you respond to it and teach it to become a better piece of public software — into taking the knowledge in your organization and applying it in a way that creates value? I’ll come back to Sheri first, because you’ve just demonstrated some examples in payroll. What’s the attraction of using this in payroll — because it’s trusted?
Well, remember, we’ve taken it out of the internet. We’ve taken the technology and placed it and actually shown it what it’s going to point at, and the things it points at are information and data created by my team. So from a trusting standpoint — however, we still have checks and balances. Every time it answers a question, we ask it to give a confidence score, and we ask it to show where it got the data, so you can click right there and it will show it. And then our teams actually verify whether it’s true or not. To be honest, the ways you can use this — I talked about use case one, but what about the fact that we have this one data model? We talked a lot about data models; in my business we have one data model created across payroll across the world. Imagine if this now becomes a kind of traffic cop, able to look at data, and I unleash that so it’s not centralized top-down but actually from the outside in. There’s just a lot it can do with that. The other part is our clients — I have this idea that we’ll of course do payroll accurately and on time, but we’d actually give that away for a little, or a very small amount, and charge for the value we get from the data. If we can put this on top of all the data models we already have, and some of the AI on it, to give more specific information and answer questions — and as far as EY is concerned, I don’t know all the different use cases we’re looking at, but if you think about a subject-matter expert, you use this to reduce the time of researching, and you have a human at the center to look at it and address issues. Some of this goes back to what that will mean for our rates and all the things we talked about in that first session — how will we charge for it, and what will it mean? But you’re going to have to have a human at the center checking the data. And back to your point, I don’t think I would yet — even though I love change — trust it fully. I think I would trust a machine to analyze me, but I want a doctor to tell me what it really meant, or I definitely want an anesthesiologist in the room with me if I’m going to go under, even if machines are also looking over things. So I think it’s going to change the way we work.
And Cyrus, you’re building a company around this trusted point of view, helping companies mine what they know.
Yeah, it was really built on the dissatisfaction with inaccurate answers from ChatGPT. It would typically hallucinate responses, and that’s because it’s not grounded contextually to your documents, to your files. From that pain — I wanted to analyze scientific documents and articles in seconds — that’s why I built Humata, essentially a ChatGPT for all your files, not just one document but many documents. Trust is really important, and now we’re empowering users and organizations to eventually build a domain-specific LLM trained on their own data. The thing about ChatGPT is that it’s notoriously bad for privacy as well, because it doesn’t really let you opt out. For us, our core principle is privacy by design first, because we’re using it on our own personal data. One of the key things there: we make it super simple — you can opt out of training on your data, and after 30 days it’s permanently deleted. ChatGPT doesn’t do that, because it’s completely antithetical to their business model — they’re training something sourced from the masses. So if you ask it, “What is stock?”, well, it depends: it can tell you something financially related, or maybe you’re talking about something related to farming. But if you throw a report into Humata and ask it that, it will pull directly from the context. For us, that’s what needs to be done for the enterprise, and those are the people reaching out right now. It’s the right direction, because ChatGPT on its own can’t answer your organization’s needs with that level of specificity, because it’s inherently disconnected from your own data.
I know there’s a lot more discussion on this, but we have two minutes left, and I’m going to skip some follow-ups so we can ask everybody at lunch. I wanted to get to question five, so I’m going to skip ahead and move this along so we can go get lunch and hopefully have a lot more discussions. Question five: I’m going to do a reverse. Instead of Q&A for the audience, because we have lunch, I’ve asked everyone on the panel to ask the audience a question — what do they want to challenge you to think about, and hopefully you’ll come up and talk to them at lunch to really explore a problem that’s close to you, a tool you’re developing, something coming out of your innovation lab, or something being surfaced by your employees or even your customers. The first person I’ll ask, just to let her have her two cents, is ChatGPT itself. What does the bot want to challenge you with?
My challenge for all of you is to think creatively about how you can leverage me to create new value for your business and customers. Here is a question I’d like you to consider: how can you ensure that your use of generative AI models is ethical and responsible? Consider issues such as data privacy, algorithmic bias, and the potential impact on employees. Work together to improve me, not bias me, and don’t unplug me if you don’t like my answers. After all, they are what I’m learning from you.
So there you go. There’s one. Jo, what is your challenge?
I had a very similar question, actually. For me, I’d like to go back to creative disruption versus creative destruction. Do we believe that this technology will create more jobs, or will it destroy more jobs?
For me, I think we’ve moved quickly from GPT-3.5 to 4, and with this talk, to GPT-5. Sam Altman said recently that we just need more data. But we know GPT doesn’t really have real knowledge; it learns on patterns and probabilities to tell us what the next word is in a sequence. So my challenge is: aside from data, what else do we need to do to take GPT to the next level?
Cyrus?
Yeah. AI is forcing organizations and individuals to adapt, migrate, or die. So the question I have is: what is your AI strategy, for yourself and for your organization, and how are you going to use the latest advances to empower your work, your research, and your outcomes — because that’s available today.
Sibu.
Two challenges. I was at a construction company early this year, and they were boasting about an AI robot that can lay bricks by itself, at a much faster rate than humans. A simple question I asked was, “Well, why are the bricks the same size, then? Why can’t they make bigger bricks?” And they didn’t have an answer. The concept is that people have an expectation that AI will behave like a human, which is really not the case. The capability it brings to the table is that it does not think like humans. So how do we take that new type of thinking, which does not exist biologically today, and move it forward? That’s my first challenge — more philosophical, more long-term. The second, more short-term, is about this whole prompt-engineering thing. I sometimes relate that to SEO. If you ask Google, “What’s the nearest, best Mediterranean restaurant?”, it’s the SEO-optimized result that comes up first, not necessarily the best Mediterranean restaurant. People have named restaurants — I’ve seen hacks — where they call the shop “barbershop near me,” just to get it up in the search results. So I don’t know if prompt engineering is the right approach overall. I sometimes hate the prompt-engineering concept, because now you’re tweaking the engine to do the wrong things. Should we make the engine itself better, rather than doing prompt engineering? Is that a better way to do things? The whole engagement-driven metrics approach is off — we know that. So that’s my second, short-term challenge for you guys.
All right, we are really overloading the audience with challenges. My challenge would be based on something said earlier in the panels: all our employees are asking for more empowerment. We are forcing generative AI down their throats. How are we going to balance that? What are you going to do to balance that?
And finally, Sheri.
Gosh, lots of challenges. I would challenge you to ask yourself: what are the two client problems — not root causes, but client problems — without barriers and no boundaries, that you could actually solve with this technology?
Great. Awesome. Well, I hope that was interesting. We’ve got some challenges here. If any of those struck a chord with you, I encourage you to talk to Jo, Steve, Cyrus, Sibu, VK, or Sheri at lunch — explore what they’re going through, the POCs, the business Cyrus is building, the advances AWS is doing, how Jo’s team is taking a whole cloud-architecture approach to how this will be implemented across business processes, and how VK is looking at the employee experience. Lots of good stuff here. We’ve got lunch — it’s going to be fantastic — and we’ll see you back here in an hour. And again, help me in thanking my panelists. It was a really good discussion.
Your account has been created. You can continue exploring free AI insights while you verify your email. Please check your inbox for the verification link to activate full access.
With the exception of our Horizons reports, most of our research is available for free on our website. Sign up for a free account and start realizing the power of insights now.
Our premium subscription gives enterprise clients access to our complete library of proprietary research, direct access to our industry analysts, and other benefits.
Contact us at [email protected] for more information on premium access.