No Priors: Artificial Intelligence | Technology | Startups - AI is the new enterprise UI with Clara Shih, CEO Salesforce AI

Episode Date: December 7, 2023

AI is the new UI for enterprise customers, according to Clara Shih, the CEO of Salesforce AI. Salesforce released Einstein, now called Einstein GPT, in 2016, making it an early example of how beneficial AI can be when embedded in enterprise software. This week on No Priors, Sarah and Elad talked with Clara about what the evolution of AI in enterprise looks like, how Salesforce is adoption AI across the organization, and the onboarding process for companies looking to integrate AI into their workflow, plus the challenges of pricing for AI services. Clara Shih is the Chief Executive Officer of Salesforce AI where she leads the AI efforts across Salesforce including AI co-pilot and agent platform, model development, go-to-market growth, adoption, partnerships, ecosystems, and secure responsible AI. Before that was the CEO of Salesforce Service Cloud She is also the co-founder and previous CEO of Hearsay Systems. She is also on the Board of Directors at Starbucks.  Show Links:  Clara’s Linkedin Ask more of AI podcast Salesforce AI Sign up for new podcasts every week. Email feedback to [e-mail] Follow us on Twitter: @NoPriorsPod | @Saranormous | @EladGil | @clarashih Show Notes:  (0:00) Clara’s Background (0:50) From cloud services to AI (3:25) Internal Model Development vs Open Source (5:20) The Co-Pilot Approach (8:50) Enterprise AI Adoption (10:54) The future of Enterprise AI (13:23) Cross-team collaboration (14:40) AI is the new UI (19:11) Structuring the Dataset (21:25) What’s next for generative AI in Enterprise (23:18) Pricing challenges in AI (26:30) Startups and AI (28:22) Collaboration in AI Industry

Today, I know priors we have an entrepreneur and executive Clara She. Clara is currently CEO of Salesforce AI, and before that was the CEO of Salesforce Service Cloud and of hearsay social, a company she's co-founder of, as well as she was a board member at Starbucks. Clara currently leads artificial intelligence efforts across Salesforce, including AI co-pilot and agent platform, model development, go-to-market, growth, adoption, partnerships, ecosystems, secure, responsible AI. So much stuff, I got tired just going through all of it. So she must be exhausted. Today, I know prior as we talk with Claire about Salesforce is for Rays and degenerative AI and the future evolution of AI in the enterprise. So thank you so much for joining us today, Clara.

Sir, thanks for having me. I'm a big fan. So I was hoping to just start off with how you ended up taking on the CEO role for Salesforce AI. And before that, you're working on service cloud. And then we has where this big wave of innovation happen in terms of generative AI and Salesforce has been quite fast to adapt to it. So just hoping to learn a little bit more about how your role evolved and the kinds of areas that you focus on today. Yeah, I mean, if you go back to hearsay days, and Alad, you might know this. Hearsay had and continues to have NLP to mine the messages that come through. And hearsay mines it for both lead generation opportunities as well as to detect compliance and fractions. So that was like really when, you know, just from an empirical

standpoint, I got closer to AI and ML. And this is like all, you know, pre-large language models. And then when I joined Service Cloud, it's like almost three years ago, when you think about the customer service world, and there's a lot of AI. There's been chatbots for many years, and we were using very early, you know, pre-GPT types of transformer models to do that. And just as we started playing around with our own models and we saw OpenAI models and the ecosystems models get better and better, it just became obvious that this would be a core

part of service cloud going forward. So I'd say probably, you know, a year and a half ago is when, you know, in the service cloud world, my engineering leader, J.S. and I, we really started to double down on these experiments, more prototypes. We were working with a couple customers, including Gucci, to develop very early prototypes of what now has become service GPT. And we were just learning and iterating and figuring things out as we went. Well, then of course, fast forward to last year, chat GPT is launched, and now every customer is super interested in AI. And across Salesforce, I think there was a sudden, you know, wake up moment to say, how do we apply large language models to every cloud? And so I think we were in a position of saying,

hey, here's what we've learned, working with Gucci, working with these other prototype customers, and let's start to think about how this applies to sales and marketing and commerce and slack. And by the way, instead of each of us building this separately, how do we create a common platform and shared services for everything from model fine tuning to prompt builder to the trust layer and the gateway so that we can all go really fast and also empower our ecosystem to do so. So that was formalized into a separate role and this new new role that I took on about six or seven months ago. And I guess Salesforce for a long time now has been building a lot of its own models. You know, I had very early, in hindsight now, forays into AI, things like Einstein and other things. And I know that's evolved into, you know, there's an outstanding co-pilot and I send GPT and other things like that as well. How much of the model development that you folks do now is internal versus using external sort of model sources, be they open source or close source?

We're taking really an open architecture approach because we have, we serve such a diverse set of customers. Some of our customers are large enterprises. They have their own models or they want to find. tune their own. Others are all the way down to SMBs who don't want to have anything to do with model selection and just want us to figure everything out for them. And so we're kind of taking the best of what's out there and we're offering customers choice. And then there's a set of customers who have kind of asked us to take it on, right? They want us to figure out based on the data and the feedback that we're getting and given cost performance and latency objectives.

They want us to choose the right model for the right task. So it's really a combination of using whether it's co-gen from our research team, which powers Apex co-gen GPT that we have in our developer GPT, where you also fine-tuning versions of that for domain-specific models in customer service and for sales and for specific industries like healthcare and financial services, whether it's those in-house models or it's working with our customers to allow them to very easily spin up and fine-tune their own models using the data that they have within sales Salesforce Data Cloud, or it's offering the choice of external third-party models, be it Anthropic and Cohere, which are both Salesforce Ventures investments, or OpenAI, which is

the close partner or Google Vertex, and offering people either the choice to buy those through us or to bring their own API keys. And then I know they also provide other things that are integrated in the platform or taking a platform-based approach to things like co-pilots, agents, and an agent-based platform. Can you tell us more about what Salesforce is doing there in some of the directions that you're hoping to go in? And then also, I guess, related to that, how early do you think agents are and how do you think they evolve over time? Because it seems like we're kind of in the nascent phase of these things, but they're still very exciting. Yes.

So the way the Salesforce is rolled out, I mean, as you mentioned, the earliest foray into LLMs were models. We've had models for, you know, four or five years, large language models that we've developed. And we've open sourced many of these on Hugging Face, which is another Salesforce Ventures investment. And then in March of this year, we announced our plan to introduce out-of-the-box AI features into every existing Salesforce cloud. So this is what I mean when you hear the words service GBT, sales GBT, marketing GBT. It's these prompt templates that Salesforce product managers have created based on where they see opportunities and operational bottlenecks for the jobs to be done for their buyer and user base. So a great example of this, the most popular one, is service reply recommendations for contact center agents.

So customer sends an email in or they chat something in, and then we provide a suggested response grounded in the knowledge article and passed similar cases for that particular customer. So that's what we have out in the market today. It's GA. We have customers using it, giving us feedback. So then in parallel, our platform team is building up the platform, as you mentioned, right? And the platform itself is co-pilot, which is the natural language interface that will span

across all of our clouds as well as Slack. And then it's also co-pilot Studio. And within co-pilot Studio, there's three platform, like really big platform areas that we're building out. The first one is prompt builder. And, you know, as you can imagine, a lot of our customers, they want to take the prompt templates that the sales cloud product managers have created. They want to customize it. They want it to be in their brand tone.

They want to point it to a different model. They want to make all kinds of tweaks. They want to ground it in different data. That might be a custom field in their instance of Salesforce that doesn't exist in the out-of-the-box sales force, et cetera. So Prompt Builder, we actually just launched our pilot of that last week. And we're already getting customer feedback, which is incredible, just the speed at which is this is all going. The second part of Copilot Studio is Action Builder. And that's where we start to give, you know, empower.

the co-pilot with agent powers, right, with whether it's workflows or it's integrations. You think about our customers have spent decades building all of their customer workflows within Salesforce. They have all of their sharing rules and permissions. They have all of their integrations and the SLAs and the security guardrails for their integrations using MuleSoft. So any of those now with one click can be designated as an action for the co-pilot agent, which is pretty incredible.

And then the third part of co-pilot studio is Einstein Studio, which is bring your own models. And this is a capability that customers have if they want to train or fine-tune their own predictive or generative models using data that they have within Salesforce and or indexed by Salesforce in our data graph. So I think, again, Salesforce has done a flurry of really amazing work in a short period of time. One thing I always wonder about enterprise AI or the adoption of AI by enterprises, just the rate at which they're actually really using it. because as far as I can tell, a lot of people woke up to the importance of this industry just a year ago, right? When ChatGPT launched, it was almost a starting gun for generative AI. And obviously, you folks have been doing a lot and broader areas of AI before this. How much adoption do you see on the generative side so far?

Is it large numbers of customers? Is it a handful? Is it mainly experiments? Is it pilots or people doing this in production? I'm sort of curious about sort of the real traction that's being seen today. Well, we have a lot of customers. So the answer is all of the above. Right. We have some customers who are, they've rolled out service GPT or sales GPT. It's operationalized across their contact center. It's already changing the day and the life of their contact center reps, which is pretty amazing, right? Just to talk to some of these individuals and hear them feel like they're doing the best work in their careers because a lot of the manual look up and wrote tasks that bogged them down before and made customers angry can now be largely automated or much accelerated with generative AI.

So we have examples of customers that have done that. Of course, most customers are in the middle, right? They're still experimenting. They're realizing how important it is to get their data ducks in a row. And they're starting to do things like connect their Salesforce data cloud with their various data lakes in their organization. And of course, the Fortune 500, every one of them has multiple different ones. And so one thing that's really exciting for us is we just announced and rolled out zero ETL data sharing. partnership integrations with BigQuery, with data bricks, with Snowflake, et cetera, so that really

customers can bring all of their structured and unstructured data into one place to really power these generative use cases. I guess if I were to think about it from a macro perspective and not a salesperson specific question, but when do you think we'll really see large-scale adoption of AI and big enterprises? Do you think that's, and I know it's always hard to predict these things. You think that's a year away, two years away, three years away, because part of what I always wonder relative to the ecosystem is, for example, you see all these tool companies, you know, around Eval or around observability or other things that Salesforce may not really touch as much, but that other companies are focused on. A lot of their future is sort of dependent on how

rapidly enterprises adopt these things or how rapidly they ramp. And so I'm a little bit curious about your viewpoint in terms of, you know, are we in the first inning? Are we in the third inning, like where are we relative to sort of enterprise adoption? It's hard to generalize because there's a distribution, but if I were to try to aggregate across everything, I mean, it's early, right? Probably the second or third inning. I'm not a baseball expert, but that's like probably roughly where it is. Like there are enough companies now, though few and far between, but there are enough of them where it proves out the value. It proves out that you actually can transform business processes in a big way.

But most companies, especially in the enterprise, as you know, their data is just like all over the place. And so that's kind of like step one. And we're seeing our data cloud grow as the fastest organically developed product in Salesforce's history. And a lot of that is driven by this need for data to power AI, whether it's for training and fine-tuning or for reg. That makes sense.

So you're basically saying step one is get your data in order. And then as far as I can tell, at least in my experience, step two has been either prototype something. something for external use, but it's still a prototype, or start using it for internal tooling or internal efficiency gains. And then step three always seems to be okay. Now we're actually going to push it out into our own end products or to our own end users or customers.

I would largely say that's true, but there's kind of like smaller pieces that you can bite off. I think the most common thing that we see in probably because we're a CRM company is, is in the customer service world. You don't have to have all of your enterprise data cleaned up, right? That might take a little bit longer. but can you have all of your knowledge articles across multiple knowledge silos? Can you bring that together using data cloud with our connectors and with vector search and embeddings to drive really good rag for any customer service question, whether it comes

in to the self-service agent, formerly known as Einstein bots, or whether it comes into a person? How did you think about just given the breadth of the Salesforce product suite and your role to advance, like, AI across the organization. How did you think about educating the rest of the product management and engineering organization or teaching them, like, how these experiences can change with AI capabilities? I'd say it's really, it's not one way, right? There's so much interest in all of this. And we have such an amazing team that everyone is just curious and wants to learn.

And they're coming up with ideas. I mean, so much of what we're building in their roadmap is coming from people. from all across the organization. So I say it's been a very collaborative effort, but it is, that is something kind of an ongoing effort, and especially as we think about, you know, as you're alluding to a lot, agents maturing and being able to do more,

I think it's really going to dramatically transform how we approach software development, right? A lot of what was explicitly hard-coded as different branching and execution paths and painstakingly, speccing out every screen in a user experience. Like a lot of that, maybe you could just like hand off to the agent to be able to do dynamically. Is there anything that you're excited about from like a change in end user experiences?

If you project out like a year or two or three, right? Like if we follow a lot's framework of, you know, or your phrasing of like you get your data ducks in a row and you have some internal and external use cases, if we just think about the externally facing experiences, I think it's much more intuitive for people to think about efficiency in sort of, like, for example, customer service versus like what can you as an end user expect that will be better? Yeah, it is such an exciting area. Like, AI is a new UI or maybe Slack is a new UI for AI. And that's also been really amazing, right? It's just looking at first, like these acquisitions that Salesforce made that at the time,

like admittedly, they were not made in the name of AI. But whether it's Slack as an interface, conversational interface, or its tablo as visualizing data and pulling in more data sources. Its Mulesoft is having all of the plugins and extensions that you could possibly want in an enterprise. Like, it's really played out nicely. But back to your question on the user experience,

I'll define, I'll answer that in both the literal, like, product, UX, but also the day-to-day experience that we're hearing users of our Einstein GPT product share. So from a user experience standpoint, we have this pretty awesome prototype. It's called Generative Canvas, where, you know, as you're conversing with the Einstein co-pilot, it's kind of just popping up different components that you would need from within what you're doing.

So if you're asking about a particular sales opportunity, as you ask questions, it'll surface up and kind of drill down the visualization of that. And so that's an example of previously what we would call a lightning web component page that you would have to hard code and hardwire. But through generative canvas, we already have all the components there. It knows to call the right items and visualize the right things from Tableau,

Salesforce reporting, be able to update records. So that's, it's pretty exciting. And it's not ready for prime time. It's very raw and messy, but that's kind of how we're operating. It's just we're learning from showing that to different customers, testing it at ourselves, and then we're going to eventually roll out

something that's pretty radically different. The other part of UX is working with Slack and thinking about how agents can be used, not just by one person, but by teams of people. And there's a lot of exciting UX work being done there. Now, in terms of experience, like the day-to-day experience of these users, even now, as I alluded to earlier, we're seeing customer service representatives, their day-to-day experience get completely transformed by generative. AI. So Gucci is a great example. They hired a number of new service representatives during COVID. There was high turnover in the early days of the pandemic. The thing about being a Gucci service advisor is you really have to know your product. People are spending a lot of money. They have high expectations. And so it's been really great to basically use retrieval

augmentation to basically help arm every Gucci service advisor with the right brand storytelling, the right troubleshooting that an expert would have. And what we've seen is that the average handle time on support issues has gone down. But instead of hanging up, the service advisor is able to have a deeper conversation. And because Salesforce has a 360 degree of view of the customer, a service advisor can see that, you know, Sarah, you have, we solved your issue with your broken buckle and your purse, but we see that you have a belt and a pair of shoes in your Gucci cart

or you've been browsing on the website. And so now I'm also able to empower the service advisor to have a sales conversation and a marketing conversation with you to tell you about the heritage of these shoes and how, you know, Jackie Kennedy used to wear these shoes too. So it just is changing the job. And it's really breaking out of these traditional department roles and functions into what is the customer really want and how do we empower that individual

who's working there, even if they're a new hire, with all the knowledge that they need to know to be able to address the customers wants and needs. That's very cool. One of the things that I feel like you would have a special view into is how enterprises are thinking about how their data interacts with all of these AI products. Right. I think that has been one of the biggest concerns to resolve or or just you know issues as enterprises think about adoption here not just do we have data of the quality and structure that's useful to retrieve or train in these AI models but actually like who who is going to be managing it and what happens to these models and ownership like what have you learned from working with customers around some of

their most sensitive data there's so much I mean we don't have all the answers but we've learned a lot so far, I mean, both how unstructured data gets treated and not all unstructured data is alike. There's unstructured data like a PRD or a service knowledge article where it's been written specifically with the intention of communicating a certain set of things. And you can probably assume everything in that unstructured document is important. Conversely, there's unstructured data that's in the form of transcripts, whether it's called transcripts, chat transcripts, or Slack channels, and it's the opposite.

And there's, like, you can assume that most of what's in there was not intended for other people. There's, like, a lot of back and forth and clarification and just, you know... Talking about my dog. Yeah, so then you have to, like, pre-process. You have to, like, mine that.

You have to do a step further before you use that for something like retrieval. And so that's something that we've learned and we're building in the capability to do that. Within that unstructured data, of course, There's data that should remain unstructured and can be vectorized and embeddings. But there's some data. There's actually a lot of structured data in there sometimes, right?

In a phone transcript that a retailer might have with the customer, the customer might reveal that her favorite color is blue and that she has a teenage daughter that she also wants to shop for. Those should then populate the structured data fields in Salesforce, which is also something that we're doing. What do you think is the most unexpected things that you've seen emerge out of generative AI relative to the enterprise? Or is there anything that really stands out? Everything. The fact that it works so well. And yeah, I mean, I feel like I'm surprised every week.

Yeah. Definitely predictions in terms of, you know, we look out two, three, four years from now, any major changes or overhauls in terms of how we think about enterprise software. Because you mentioned, for example, AI is the new UI, for example. I'm sort of curious, as we think ahead a couple years, how does it substantiate? in terms of software or business models or sort of changes in terms of how people interact with all those stuff? Yeah, I mean, I kind of liken it to, I imagine that it was like this when cloud first became a thing.

There were some applications where you're like, clearly this needs to move into the cloud. There's so much value to having it accessible on demand and on a mobile device and wherever you go. And then there were some applications where you really wanted, I mean, even to this day, you want to keep it on premise. And it's probably going to be the same is true with AI, right? There's some workflows or some decisioning and branching that you want to be fully deterministic.

You want to reproduce the exact same way every single time. It might be authentication. It could be a financial transaction. It could be a health care procedure. And it'll stay the way that currently is. There's a lot of other ones where the job of the software engineer and product manager and designer, is going to shift from prescribing the how to prescribing or describing the why and the what and the

goal. And we leave it to the AI to figure out stochastically the how. I have two more sort of business-oriented questions for you. Unlike much of the software developed over the last five, 10 years that it wasn't, I mean, much software is still very data processing heavy. So it's not free, but AI products in particular, there's real cogs, right, in terms of the compute. How do you guys think about this at Salesforce in your product launches or in thinking about new skews and pricing? That is such a difficult question. And it's something that we talk about all the time. We've put pricing out there so far for our AI products.

It's really difficult. You know, I think that you have to cover your costs, but you also have to provide it in a way that customers can easily understand and you don't need complicated calculators to try to predict token usage. So I think that's the balance that we're trying to strike right now. Overall, though, the key thing that we have to achieve is to show value, show ROI, right? Like in the case of Gucci and other retailers, are we reducing average handle time? Are we driving sales conversion uplift over the baseline? And so long as we're doing that, I think customers are willing.

to pay it can't just be an added cost without a clear benefit yeah maybe two reasons i'm still pretty optimistic about this question is you you certainly have a more complicated unit economics equation than like it's a web app and like we're using database services that are like very efficient today but in so many applications you have net new capabilities or like massive productivity gains right so wherever you see orders of magnitude improvement in terms of value you can give to the customer and then on the COGS side, you know, being able to do the same task with these AI models decreases over time monotonically, right, as we improve at every level of the stack for AI. And so I definitely think it's a complicated question, as he said, in terms of presenting that answer, getting the

answer, like, right to a really fair trade for customers. But I think there's just a lot of opportunity. Yeah. I mean, just to give an extreme example of that. I met with Christob Alenzuela yesterday, who is the founder and CEO. of Runway ML and he was talking about their involvement in the movie everything everywhere all at once which is such a good movie and I didn't know this but they actually powered a lot of the special effects that you see in the movie to the to the order that of you know only requiring seven people on the video editing team versus traditionally a movie like that would have 700 and so I mean just thinking about I mean there's a lot of questions that that

talks about from jobs and the Hollywood strikes and whatnot, but just from an ROI standpoint, right, I'm sure between the studio and runway, they're able to figure out a business model that works. Yeah. Yeah, I think it's a great example. One last question for you, just because you have such a unique viewpoint as both an executive at scale and a founder, where do you think startups should focus their efforts, right? You have Salesforce and companies of its scale that have all of its product and distribution and data advantages. What do you think is most interesting in generative AI putting on your entrepreneur hat? I mean, I see so many exciting startups out there. I think the startups that are, there's a foundational model, startups, if you can even call them

that anymore, given how big they've become or domain-specific startups that focus in an industry like legal or medicine. I think that's super interesting. I think at the tooling layer, we talked a little bit about that earlier. There's a lot to be done there to address different types of needs that different types of organizations have. And there's just so much, like, we don't know what we don't know. And so it's a time to invent and test and see what's out there. I mean, Salesforce, like, we're doing a lot, but we can't do everything. And then on the applications layer, there's a lot of applications to be built. Just like with our own internal teams at Salesforce, the way that those applications get built in the future will be very

different than they're built today. I think that for startups that to understand that and to maybe align themselves with with datagraphs that are out there because that's so essential for those applications to be relevant. Thanks for doing this with us, Clara. It was a great conversation. Yeah. Thanks for us for us tonight today. Thank you. Find us on Twitter at No Pryor's Pod. Subscribe to our YouTube channel if you want to see our faces. Follow the show on Apple Podcasts, Spotify, or wherever you listen. That way you get a new episode every week. And sign up for emails or find transcripts for every episode at no dash priors.com.