Why you cannot "DIY Your AI"

Lisa DeMille explains why 95% of AI pilots fail: LLMs alone can't run a business without context, processes, orchestration, and engagement. Salesforce's AgentForce platform combines these four systems to turn raw language models into working agents, with Slack handling the engagement layer and saving Salesforce over 140,000 work hours.
Chapters
AI models alone can't run a business, but Salesforce can. And with a rebuilt platform tailored for the agentic revolution, when it comes to the public sector, we are mission ready. Now let's learn what makes Salesforce uniquely positioned to deliver the agentic enterprise. Carey Chow joined by Lisa DeMille, SVP of Solution Engineering at Salesforce. Lisa, always love talking to you. Thrilled to have you here. A lot of big announcements here in DC. I'd love to know what are the What are the big takeaways, though, that people should have following that main keynote?
Well, Kerry, so great to be here at AgentForce TC. And wow, what a keynote delivered by Kendall Collins, our president of Mission Force. You know, today is all about seizing the Agentek opportunity. We have over 6,000 government representatives and partners here, and we're doing nothing but talking about Agenteks and AgentForce and so on. You know, the thing that stuck out to me in that keynote, and I've seen it a lot, is that 95% of AI pilots fail. That is a staggering number. I mean, I talk to government customers from around the world every single day, and I will tell you that they don't want to be a part of that 95%. They don't want to fail. But the reason why these, these pilots are failing, and Kendall was doing a great job talking about this today, was that, you know, Mark Benioff likes to say, You cannot DIY your AI, meaning that an LLM is not enough. You could have the best LLM in the world, but that's not enough. An LLM is not going to know the difference between your passport, your driver's license, and your Social Security card. It really needs context. That data feeding into the LLMs needs context. So huge piece about that keynote.
All regarding that context, tell me more about this because that leads directly into my next question. How does AgentForce work?
Right. So, again, you can't just drop in an LLM. You have to have a plan. And this is what I love about Salesforce. We always have a plan. And in order to get to the agentic enterprise, you need to have a 4-process plan. Number 1, you need to have context about your data, right? System number 1. So, system of context. Laying on that, you need to have a system of process. So this is like business processes and business rules, right? You got to tell the data what to do and when to do it. Over top of that is you have to have a system of orchestration. We call agency. I like to say system of orchestration. It's like a conductor in an orchestra going, hey, you guys do this. The drums take over and the flutes are going to like bring it on home. That's what— so it's orchestrating all of those processes and data. Now, here's where the magic happens. Then you put on a system of engagement. This is Slack. This is AgentForce. AgentForce is going to be sitting on top of all of these processes and this data, and it's actually the way that we're interacting with AgentForce, not just asking it these pointed questions, but talking to agent force like I'm talking to you right now, that's huge.
I love that analogy of the symphony orchestration there. But you're talking about Slack, and one of the biggest advantages, as you mentioned, of all the Salesforce innovations, we put all that innovation directly into the flow of work, which is Slack. What truly makes Slack such a game changer?
Oh my gosh, you don't have enough time for me to talk about how much I love Slack. So Slack has revolutionized the way we work. It has definitely revolutionized the way I work. A year ago, I start my day off, I've got my monitor, I'm in all these different applications, and I deal with a lot of products around the world, a lot of people around the world. Most of the time I can't remember if it was a Word document, an Excel, I can't remember who I talked to, I got a lot going on. Today, my day is completely different. I actually have My screen is half Slack and then half whatever I'm working in, and I engage with Slack just like I'm talking to you. I actually, I have 10 customer meetings here today. To prepare for those customer meetings would have normally taken me 2 weeks. I want to know everything about that customer. You know, when did they start with us? What products do they own? Any support tickets? That's a lot of work for anyone, 2 weeks. So, you know, I prepared last week, I asked Slackbot to do it. And I just wrote a 2-minute prompt, and within 60 seconds, I had all of my briefings. I have— we have seen this at Salesforce. Slack has saved over 140,000 hours of work. I mean, think about that at an enterprise level. It is— I don't know how I've worked without Slack, ever.
It's contributed to my— I've contributed to that number as well, because Slackbot saves me plenty of time prepping for meetings and so forth. I'd love to know, take me through a use case example when it comes to like productivity of what we're able to offer. I'd love to, because we talk about the innovation, take me through a use case.
So let's say that you had to provide permits or benefits within the government and in that, those processes, the government has a lot of data and they have a lot of forms. And imagine if you could go to Slack and just say, you know, Slack could actually surface up the cases that you have to work on every day. You can go to Slack to say, where are the ones that I need to engage? What data do I need to engage with this? How do I further help that person? You know, General Motors, so you know General Motors, right? They've deployed Slack across their 97,000-person workforce. I mean, it is literally changing and revolutionizing the way that we work today. So it's shortening the timeframe that we need to find data. It's shortening the timeframe for us to get back to citizens, back to consumers. It is just, it is incredible because remember, it sits on top of all that data. Yeah. So Slack is like looking into all of this data, this context, all these systems, and agencies have a lot of systems too.
A lot of systems that need modernizing, and that's where Agentforce comes through. Lisa DeMille, thank you so much for joining us and for sharing your insights on how AgentForce can make such a big difference. To learn more about how AgentForce Public Sector can help your organization scale, scan the QR code on your screen. Now check out this highlight from the main keynote.
Citizens are expecting 24/7 availability, instant expertise, personalized answers, and they need all of you. They need all of us in this room to come together and bring our energy and our ideas and innovation to serve them, because we cannot open up a skills gap. We cannot open up a trust gap. We cannot have a chasm between what we experience as citizens from versus what we experience as individuals. We have to unlock this superpower for the government and for nonprofits and for education. It's absolutely critical to the country and to all of us. But we know from doing this the past 2 or 3 years that AI models alone can't run your organization. Who's been told here, oh, all you need is a model, just roll out a model, it'll be great, it'll solve all your problems? It's just not the case. You know, models can't act on their own. Sometimes they get things wrong. They can't guarantee outcomes. They lack transparency. Sometimes models are confidently wrong. Does anybody here— do you know anybody, you have a friend, maybe a family member, don't say it out loud, and they're confidently wrong, sometimes about politics, sometimes about a recipe at Thanksgiving? There's nothing worse than having somebody who's confidently wrong and doesn't know your business, but they're really in your business. You know, that makes it even worse. And this is why 95% of AI pilots are failing before they make it to production, because the models alone can't do it.










