Today on the Salesforce Admins Podcast, we sit down for an Admin Evangelist roundtable discussion with Josh Birk, Jennifer Lee, and yours truly. Join us as we chat about how AI can help you be a better Salesforce Admin and what you can do to improve your prompts.
You should subscribe for the full episode, but here are a few takeaways from our conversation with Josh Birk and Jen Lee.
With everything going on with Einstein Copilot and Prompt Builder, I wanted to bring the Admin Evangelists together to find out how they’re thinking about AI and what you should do to get ready.
The number one thing that everyone agreed on is to start practicing your AI prompts. Josh recommends seeing if you can get your LLM of choice to tell you a dad joke. Then try and get it to tell you a better one. Just like how we had to learn how to write a good Google query, you’ll quickly find out that some prompts are more effective than others.
Jennifer shares the story of how her husband used ChatGPT to help with their itinerary on their trip to Italy. They still had to double-check that the restaurants it recommended were still open and that the timing of everything made sense, but it was a great starting point for planning their vacation.
Both Josh and Jennifer also use AI to help with work. Jennifer’s found ChatGPT to be really helpful for writing formulas. She used to spend hours on Google trying to find an example that matched the exact scenario she needed. These days, she can just write a prompt with her specific parameters and get back something useful in seconds. If Salesforce gives her an error, she can tell ChatGPT about it and it’ll try to fix the code.
Josh, meanwhile, has been using AI to help generate Apex code from scratch when he’s spinning up a demo org. As he’s quick to point out, it’s not necessarily helpful for the maintenance and debugging tasks that most developers do on a daily basis, but it’s perfect for his particular use case.
One last thing we talked about that I want to highlight is the importance of the human in the loop.
We used the example of someone calling a power company to find out why their electricity bill is higher. If a human has realized that the weather has a major effect on usage rates and created a screen flow to call the right API, then an AI might be able to give the customer the right answer. But you need a human in the loop to do that second-order thinking.
We’ll have even more about how Salesforce Admins can use AI next week in Josh’s deep dive episode, so be sure to subscribe to hear more from the Salesforce Admins Podcast.
Mike Gerholdt:
So in the world of AI and GPTs, and I think one’s called Hugging Face, maybe it’s Hugging Chat, I don’t know. There is a lot to learn and people maybe you’re afraid of it or you haven’t tried something out. I don’t know. We’ve heard a lot as evangelists on the admin relations team. And so this week I wanted to dive in with all of the rockstar evangelists that I have. Josh Birk, our Senior Admin Evangelist, and Jen, our Lead Admin Evangelist. Also, everything flow about let’s dive into prompts and let’s start learning about prompts and what should we be afraid of or what shouldn’t we be afraid of, or what should we start doing? So that was a really long, highly caffeinated intro, but welcome to the show, Jennifer and Josh.
Joshua Birk:
Thanks for having us.
Jennifer:
Hey.
Mike Gerholdt:
So I won’t name names, but I have been around in the community, and I have heard people like, “Cool, oh, you’ve done something with ChatGPT.” And to be frank, if you follow me on Instagram, you’ll realize that my feed, sometime in March, quickly took over crazy images generated by Dall-E. Because I find it fascinating that I can give it words to a 1970s music classic rock and say, “Make a picture.” And it will produce something that would take me years to put together. And sometimes it’s crazy, and I have to share that with the world because I think that’s just so cool. But I guess I’m not afraid to try things out. So that’s where I was with the world of AI. Where do you guys fall?
Joshua Birk:
Well, first of all, I want to go back to something you just said about what’s dangerous about talking to something like let’s just call it ChatGPT. Just use that as the generic one since it’s the most famous and most popular. But I go back to-
Mike Gerholdt:
It’s like the Kleenex of AI.
Joshua Birk:
Yeah, exactly.
Mike Gerholdt:
Everybody knows about it.
Joshua Birk:
Exactly. And I’m going to say this and I’m going to add a huge caveat to it. But when it comes to going to especially a free one, and just tinkering around with it, it’s the same thing that one of the best pieces of advice I ever got before I started programming on a pretty basic on an Apple II. And my teacher at the time, I don’t even remember what the class was called, it wasn’t computer science or anything like that. But basically he was like, “Do whatever you want. You’re not going to break the computer.” There’s nothing you’re going to do that’s going to… Nothing’s going to blow up. Nothing’s going to go in smoke.
The dangers of using AI right now is not the conversation you’re having with the AI, it’s what you do with the result. So Mike, you’re posting things to Facebook. What’s the nefarious outcome to that? There’s not, right? Now, if I use the generation of AI to submit my legal brief without checking it, you might run into some troubles. But what I always tell people is just jump in and just try talking to it, because until you do… I usually start with have it tell you a dad joke, have it tell you five dad jokes.
My favorite, actually, one of my favorites is go in and play 20 questions with it. Because having an AI guess the object that you’re thinking in your brain is actually, it’s an interesting way to prove how a conversational model works. But basically to go back to my teacher’s advice, don’t be scared. Go in an experiment.
Jennifer:
Yeah, I have to say that when it first came out, I wasn’t one of the people who went running to play with it. But when we were planning our trip to Italy last year, my now husband went and used it to come up with the itinerary. He said, “Okay, we’re going to start here. Here are the sites we’re going to, recommend some restaurants for me.” So then it came back and it said, “Okay, you should go here and here.” He, of course, human in the loop, had to go verify that these restaurants still existed.
Mike Gerholdt:
Yup.
Joshua Birk:
Sure.
Jennifer:
But it helped put together our agenda. And then nowadays I’ve been using it to help me with formula creation, because I am the worst when it comes to the parens, commas, nested if statements, they drive me nuts. I used to spend hours Googling to find the exact scenario I needed that someone else posted, and then tweaked it. But now I can go and say, “ChatGPT, I need a if statement for three things. Give me the structure,” and then I would put it in there. So it’s been really helpful to me in that regard.
Joshua Birk:
And not to throw down Jen, but being the worst, I might have to, I’m not sure. I think I might be worse than you are, because I’ve done the same thing. And the nice thing is, so ChatGPT is what’s called a chained conversational AI. Which means when you create that formula, if you get… Like you put it in and you get an error from Salesforce, you can give that error back to ChatGPT and be like, “You got this wrong,” and it already has the context of the formula it gave you, and it’d be like, “Oh, you’re right. Let me fix that for you.”
Jennifer:
Right.
Mike Gerholdt:
Yeah. And then just to clarify, in the context of what we’re talking about, a prompt, because we spelled this out in our workshops, but a prompt is a starting line for an AI conversation or task where you tell the large language model what you’re looking for using natural language, right?
Joshua Birk:
Yeah.
Mike Gerholdt:
So tell me this, because sometimes I struggle with this. How do you write your prompts differently, than say, putting in a search for Google?
Joshua Birk:
Yeah. So first of all, I think that’s one of the reasons we call it… People are like, why prompt? It’s a weird word. And I think the reason why we refer to it as prompt, because it’s a more generic thing than saying query or question or request, or anything like that. Because the thing that you need to provide in a prompt… So let’s compare those two directly. When you do a search in Google, it’s basically going to a database of a whole bunch of really fancy stuff Google’s done, and then you’re always going to get the same answer.
If you do the same prompt in ChatGPT you might get slightly different answers because that answer is being generated on the fly. And so in order to be successful with… The nice thing with ChatGPT, once again, because it’s a chain conversation AI, so you can start slow and then just keep adding context to it. But in a non-chained one like Prompt Builder, it’s adding in all the instruction and the explicit instruction, and the repeated instruction, and the context of, “No, that’s sort of the response I wanted, but that’s not exactly. Here’s how to correct it,” kind of thing.
So to compare it to a Google search, it would kind of be like, no, give me the Google searches that are actually relevant to the problem I have at hand. Which Google can’t do because it doesn’t understand context.
Mike Gerholdt:
Right. That makes sense. Jen, you said human in the loop, and I feel like we’ve naturally in just comparing Google searches to prompting a GPT to return a result, naturally put the human in the loop. But what did your husband do in terms of planning to put the human in the loop for other things?
Jennifer:
Yeah, I mean, he had to go, even though it would return an itinerary for us of sites to see in Italy, he still had to go and check and say, “Okay, is this the right path that it should take?” Confirm the hours that… Even though it suggested here’s the order, well, is it even open at that time, right? Can you imagine just showing up there and like, oh, nope, it’s not open until noon, but you have me going there the first thing in the morning.
Mike Gerholdt:
Yeah. Or it figures your itinerary is you just drive there, spend one second there, back out of the driveway.
Joshua Birk:
Right.
Mike Gerholdt:
Like, no, I’m going there for a reason.
Joshua Birk:
Yeah, it doesn’t always get the context of time correctly the first time. The last time I tried to do an itinerary, it gave me so many things to do in three hours that no human being could possibly do it.
Mike Gerholdt:
Oh my goodness. Yeah. I will say I have tried… So one of the first things I experimented with was, I think it’s called Gemini Now, because it was Google and it plugged into their Google Maps. And I think planning road trips and stuff, when you’re looking for specific stops, not touristy stops, but okay, a couple hours into the drive, I’m going to want to stop and get some diesel. I want to take a break from driving for 10 or 15 minutes.
I don’t want to just sit there and figure out on the map and scroll mile by mile on the map. Does that look like a gas station? Does that look like a gas station? And have it go through that. But oftentimes I also have to go back through the map and validate like, oh, is that still open?
Joshua Birk:
Right. Yeah. So I think they’ve gotten better because these large models at one point were impossible to pre-train on small amounts of data. They basically had to re-consume the entire internet over and over again. And that’s why when they first became popular, they warned you it doesn’t understand anything. Basically pre-pandemic, if I remember right. And that was a problem here in Chicago because a lot of very famous places had shut down.
So to your point, Jen, about itinerary, making sure they’re still there, it would’ve been spot on recommendations in 2019, but now not so much. And again, I think they’ve learned how to train them with smaller bits of data. So I think this is better. But yeah, it’s whatever moment in time it’s looking at when it was grounded in that data.
Jennifer:
Well, when I was looking at free GPT products, I used Perplexity AI, and that was using the internet and not data up until a certain timeframe.
Joshua Birk:
Nice.
Jennifer:
And it came back and said, here’s my answer, but then here’s my sources.
Joshua Birk:
Nice.
Jennifer:
So then if I wanted, I could go click on that to get more information
Mike Gerholdt:
Yeah, as an article from 2007. Like, “Oh, cool, five star Michelin restaurant.” What do you feel, in terms of using AI just for your everyday life, is helping you understand AI better as we work to understand the capabilities of prompt builder and copilot?
Joshua Birk:
So for work, I’m going to confess, because I’ve been using AI to generate a lot of APEX. And it’s not that I can’t write that APEX. And I’m going to say this because I think this is a very unique situation because I’m writing a lot of APEX that’s from scratch. I’m not an enterprise developer who has to go in and maintain large bits of code and large project related code and things like that. And there’s this huge question out there for the everyday developer, is AI something that’s really useful? Because a lot of everyday development is maintenance and it’s bug fixes and it’s very detailed work.
And there’s probably a role for AI when it comes to code reviews and things like that. But for what I’m using it for, it’s like I know how to write this and I know I’ve written something like this, but AI will do it 90% correctly in about three seconds. I can’t type that fast, especially these days. I cannot type that fast. And so if I just kind of go back in and I’ll reiterate it and stuff like that, and because of that in very short period of time, for instance, the demo works we’re spinning up. I have APEX that can recreate all the data for me in these styles.
And that’s, I think, the important thing is it’s in the style that I want the data to look like. It’s in the level of realism that I want the data to look like. Which actually would’ve been… That’s the part that would’ve been very difficult me to do because to sit back and be like, I want 30 companies that sound like they would work with Northern Trail Outfitters.
Mike Gerholdt:
Oh yeah.
Joshua Birk:
Right? I’d have to sit back and actually ideate about that and spit out ideas and all of this kind of stuff. And ChatGPT is just like-
Mike Gerholdt:
I’d come up with a whole bunch of Southern Trail Outfitters and Western Trail Outfitters and Eastern Trail Outfitters.
Joshua Birk:
Exactly, exactly. It’s not shy. It will just try to give you that kind of slightly creative information shoved inside of an APEX class. And it’s definitely, I mean, I can’t even calculate how much time it saved me.
Jennifer:
I think for me, outside of the formula piece, which it has definitely helped me build formulas faster, but I use it to shorten things that I’ve written. Here’s this thing, I need to narrow it down into this many characters. So that’s been helpful in shortening that up. I have a problem of condensing things.
Mike Gerholdt:
No, I think that’s great. I mean, that’s actually to look back at admin track at Dreamforce a few years ago, the biggest session that was always attended was documenting your org. And I think it’s because you always would have to start from scratch as opposed to, Jen, I use AI for a lot of that as well. I will write a description in X number of characters following this style of writing and then give me three versions of it. Because Josh, to your point, I don’t have three versions in my head, or I need a fourth version, but I need you to give me three to kind of push the cart down the hill a little bit.
Joshua Birk:
Yeah. And it’s like, why do we keep going back to summarization as such a classic AI use case? Part of it is because it can read it and then write that summarization faster than you could read it by a 100th percent. It can do it. So it’s not just the speed of generation, it’s the speed of consumption as well.
Mike Gerholdt:
So working with, looking ahead, because Prompt Builders GA, we’re doing workshops. What are some of the things that we should think about as admins for what organizations may be looking for in terms of prompts, so that… Because I literally think the amount of creativity that we have in our heads is the limitations we all have for delivering on super useful prompts.
Joshua Birk:
Well, Mike, I don’t know if you know this yet, but we have an upcoming episode with my friend Ravish [inaudible 00:14:47], where we talk about this a little-
Mike Gerholdt:
Next week.
Joshua Birk:
Right. So one of the things I asked him, because Ravish actually is working with customers and working with billing out some of these solutions. And I think the answer to that question is start looking at… Let’s focus very specifically on this field generation prompt builder template. Look at your page layout and ask yourself, is there something here, or can I add a field to this that would be useful? Because going back to our friend summarization, that we could summarize not just the object that we’re looking at on the record, but summarize it in its related data. And it’s a data next to it and data that Flow can find to it and data that’s in Data Cloud and all of these points of data.
And what Ravish said, not to spoil my own content, but he’s like, “One of the most useful things is either meeting prep or call prep.” Like if you have a call with a customer coming up, you want to pull up the account and you want to get a one paragraph overview of all the activity of that account so that you are enlightened almost right away as to what to do with this. So I think it’s going to be hard until you see the fields actually being generated, because it’s a very different kind of data that we’re used to.
But the question is going to be, I’m trying to think of if, and maybe one of you two have a good example, because Salesforce has evolved so much over the years, right? I always joke with people, you don’t know how good you have it because you can do GO selection and SOQL queries. I had to write an entire APEX class in order to make that work, and I couldn’t even do it with Radius. I had to select zip codes by a square because we didn’t have the capacity to do sign and co-sign back then. The things that we’ve added that make the platform so much more intelligent, and this is another step in that evolution. And so it’s time to start asking yourself, what layer of generative data can I add to my object model that’s going to make my users more successful?
Mike Gerholdt:
Yeah, I would agree. I think along those lines, looking at what’s available, looking at what GPT, even if you can’t get your hands on anything Einstein and all you do is Trailhead modules and you listen to some podcasts like this. I think the one thing, not that worries me, but the one thing that I would have a serious conversation sit down with stakeholders is I would look at what’s all the anecdotal data that we store in Salesforce and where is it at?
And so I bring that up because Josh, your example of zip codes, phone numbers, hard data, super hard data, a zip code, a phone number, even a street address anymore, you can write a call out and verify that. The number of websites I go to now, and I literally type the three numbers of my house in, and it’s already narrowed it down to the five possible addresses this could be, I think is amazing, it didn’t exist before. But what I think the power of what people are wanting is, so give me a summarization of the last five calls and what their pain points were.
And if you’re not running some sort of reports or like Jen, if you don’t have, I’m thinking screen flows with really good fields to prompt people on, you need to fill this stuff out and we need to prepare for it. I don’t care how good the AI is, it’s not going to weed through bad call notes if the salespeople didn’t put that stuff in.
Joshua Birk:
Yeah. And I want to call out one of the demos I saw here internally, because I think it touches on that a lot. And the demo was case support, and I think this was Copilot. And the ticket was, “Why is my power bill so high?” And the AI doesn’t know how to answer that question properly. And this is like when we keep the human in the loop, the human in the middle, we have to remember that’s in the whole… It’s from beginning to end, because a human had to realize, oh, we charge people for electricity.
And what is one of the things that determines how much electricity you use? Oh, it’s the weather. So what I’m going to do is I’m going to build a Flow that calls out to the weather API that brings me back the weather for the last whatever period of time this ticket was for. And then the AI is like, oh, well, your power bill was high because it was super cold, or it was super hot, or whatever. But the case support person can basically just ask, analyze the support ticket and with the power of Flow that a human created, then the AI can add into it and be like, oh, this is why we think this is wrong.
Jennifer:
Yeah. I think back to the demo that we created for Dreamforce, for the admin keynote, we were using NTO, right? They did expeditions and we were able to say, okay, take the customer’s expeditions that they’ve already been on and take the customer state, and then now go and based on what they’ve done in the past, recommend something similar. And looking at all the expeditions that were available. So that would’ve taken some time for the rep to then go research that and see what expeditions they’ve been to, what they might like, based on those things, to recommend other things. But with Prompt Builder and field generation, click a button, AI does it all for you.
Mike Gerholdt:
I would agree. Well, thanks for sitting around and jawing about prompts and GPTs, because I feel like it’s June, we’ve been talking about this… Has it been for over a year or has it just been a few months? Or has it just been a year-
Joshua Birk:
It feels like over a year.
Mike Gerholdt:
Right. Right. But it’s a little more advanced. It’s interesting because I wonder in a year from now if we listen to this episode, if it’ll sound super dated or not.
Joshua Birk:
Yeah, I think that’s an interesting question. And it’s hard to answer because you have OpenAI saying the AI you’re talking to right now are, oh gosh, I think I’m trying to remember their exact quote and realizing it might not be something we want to record on a podcast. So let’s just say not nearly as evolved as the models that they are trying to bring into the near future. And the near future ones are supposed to be much more intelligent, much more capable of reading your contacts, much more behaving like a human, kind of thing. And so I think that poses the question of when I say how much context and how much repetition you have to put in your prompt to make it do what you want.
That statement might sound dated in about six months to a year. Because the person listening to it at that time might be like, “What are you talking about? The AI just knows who I am and they just talk to me like a travel agent. It’s just normal.” But I think in general, what we’re going to get is better results, and prompt building in general is going to remain relatively the same. We might get to the end faster. We might get more better concise data. It’s going to realize that that great pizza shop is actually closed in Chicago, and it won’t recommend it, and things like that. But I think the actual concepts of prompt building, I think they’re here to stay for a while.
Mike Gerholdt:
Wow. Well, that’s good to know. It doesn’t feel like fly-by-night anymore. Well, I’ll stick a bow on this episode for now, and thankfully none of our AI’s hallucinated.
If you enjoyed this episode, do me a favor and share it with one person, just one. It’s not that hard. If you’re listening in the Apple Podcast app, you can just tap the three dots up in the right-hand corner, click the share episode, and then you can post it to social. You can text it to a friend and be like, “Let’s listen to the prompt building episode together.”
If you’re looking for more great resources, of course, admin.salesforce.com. The number of people I know that don’t know about admin.salesforce.com is hopefully dwindling in the world because I keep bringing it up. But the good news is the reason you can go there, all the resources, if we mentioned any in the show, will be there, a transcript of the show, and links to our Admin Trailblazer group on the Trailblazer Community will be there as well. So with that, until next week when Josh talks to Ravish, and you already had a preview of that episode, we’ll see you in the cloud.
Today on the Salesforce Admins Podcast, it’s another deep dive with Josh Birk as he talks to Bobby Brill, Senior Director of Product for Einstein Discovery. Join us as we chat about how you can use Model Builder to harness the power of AI with clicks, not code. You should subscribe for the full episode, […]
Today on the Salesforce Admins Podcast, it’s time for a deep dive with Josh Birk, who talks to Raveesh Raina, Principal Solutions Engineer at Salesforce. Join us as we chat about what Prompt Builder can do and how to write effective prompts. You should subscribe for the full episode, but here are a few takeaways […]
Today on the Salesforce Admins Podcast, we’re replaying our episode with Sarah Flamion, Research Architect on Salesforce’s Research & Insights Team. Join us as we chat about what recent advances in generative AI mean for admins and the Salesforce Platform. You should subscribe for the full episode, but here are a few takeaways from our […]
