Today on the Salesforce Admins Podcast, we talk to Mehmet Orun, SVP, GM, and Data Strategist at PeerNova. Join us as we chat about how to use Data Cloud to create trustworthy AI experiences.

You should subscribe for the full episode, but here are a few takeaways from our conversation with Mehmet Orun.

The last time we had Mehmet on the pod, the big concern with AI was hallucinations. How can we be sure that the agents we create don’t start making stuff up? What’s more, how do we know that they won’t share data they shouldn’t? In short, how do we create trustworthy AI experiences?

As we’ve learned time and again, AI is only as good as the data you give it. But, as Mehmet explains, Data Cloud has changed the game for admins in terms of control over who sees what and in what context. Admins can create personalized experiences with Agentforce, constrained by the permission model and capabilities of Flow to ensure that everything is working as intended.

One thing Mehmet reflects on is the way that data management techniques have changed over time. Several best practices no longer make sense in today’s context of AI. For example, duplicate records that used to be a mortal sin make more sense when you’re trying to constrain what a customer can see vs. your employees vs. your vendors.

However, the personalized engagement that is possible with Agentforce requires a complete understanding of what’s happening with someone. At the same time, you want your agents to only act on information they’re “allowed” to see, or generate insights that are relevant to the outcomes you want to achieve. As Mehmet explains, “good” data and “bad” data is really about making sure your data is structured in a way that makes it easy to use.

The good news is that it’s never been easier to take care of your data with Data Cloud. Mehmet’s seen this with the nonprofits he works with. Data unification projects that used to take months or even years are now relatively simple affairs. You can identify bad data, filter out irrelevant data, and put the right data standardizations in place all in Setup.

Mehmet’s biggest piece of advice is to measure your data quality in terms of the business outcomes you are trying to achieve. As he points out, the amount of data you need to open an opportunity is different than what you need to close an opportunity. The same principle applies to AI agents. If you make sure they get everything they need and nothing extraneous, you’ll get good results.

There are a ton of great insights about data management best practices for AI in our conversation with Mehmet, so be sure to listen to the full episode. And make sure you’re subscribed to the Salesforce Admins Podcast so you can catch us every Thursday.

Mike Gerholdt:

Welcome back to the Salesforce Admins Podcast. Today we’re catching up with Mehmet Orun, longtime friend of the pod and true expert in data and AI. I’m going to tell you, a lot has changed in the world of artificial intelligence since our last chat. And Mehmet’s here to break it down from hallucination risks to the role of data cloud in creating trustworthy AI experiences. If you’ve ever been wondering how to make your data more meaningful and your AI outputs more reliable, well, you are in for a treat. So, make sure to follow the podcast so you don’t miss a single episode. And with that, let’s get Mehmet back on the podcast. Mehmet, welcome back to the podcast.

Mehmet Orun:

It is wonderful to be back, Mike.

Mike Gerholdt:

I know. You just come by with all these wisdoms and knowledge that you have in the world. Last time we were on, and I’ll link to that show, we were talking about hallucination risks. And it’s been a year. And boy, I tell you, a year in AI time, everything’s changed. So, what’s new in your world? What are you paying attention to in terms of AI and Agentforce?

Mehmet Orun:

To be honest, one of the interesting things about having been around a while is while the technologies are new, our overall objective haven’t really changed. And one of the things I’ve been really trying to look back to is what were past challenges we overcame, what were the parallels, and what were some of the best practices that people newer to the field around data integration, artificial intelligence, may not know about, so we can share this knowledge while absolutely picking up new ways of doing things? Also, because we definitely have new tools under our belt. And having a organized way to assess what may cause hallucination risk and mitigating it has been a truly hot topic. I have been visiting old friends, making new friends as I’ve been traveling across different Salesforce events as well. And the good news is people are excited about the potential. People are also excited about having tangible methods they can take back to their organization. I’m looking forward to sharing some of these with you today.

Mike Gerholdt:

Yeah. It’s interesting, you look at some of the stuff that’s out in the world and the spectrum for people looking at what’s going on with AI goes all the way from everything that it says is right to nothing that it says it’s right, and somebody falls somewhere in between there. But I feel like I did podcasts in 2024, early ’24, I think in ’23, even talking about hallucination. The one thing that it came back to, it seems to have gone away, because I think more of the conversation is around the quality of the data and what we’re feeding Agentforce and getting your data ready. Am I right?

Mehmet Orun:

So, given you mentioned ’23, ’24, a lot of the hallucination-

Mike Gerholdt:

This was a long time ago.

Mehmet Orun:

Yeah, in the AI world, right?

Mike Gerholdt:

Uh-huh.

Mehmet Orun:

A lot of the hallucination risk conversations that were happening, and this was mostly around ChatGPT, was because the information that was available was up to a particular date. It was predominantly unstructured data available on the internet. So, if something was published past a certain date, it was not going to show up in answers. One of the big changes I think in the ecosystem is in the past we talked about IT solutions, data warehouses, analytics, which was separate than marketing segmentation and engagement.

And then we had these really interesting LLMs and generative AI for the past several months, it feels like years. The focus is the idea of a truly enterprise scale data platform that can power automation, that can power analytics. That can look at structured and unstructured data in order to provide complete, compliant, and contextual information that can also power AI. I know we both like storytelling. I had a really interesting experience with my father a couple of weeks ago. Do you mind if I tell that story?

Mike Gerholdt:

Oh, please, tell me. I love a good story.

Mehmet Orun:

She’s a 90-year-old retired brigadier general. He’s a military engineer.

Mike Gerholdt:

90 years young, you mean?

Mehmet Orun:

Oh, man. I still barely keep up with him.

Mike Gerholdt:

See, that’s what I’m saying.

Mehmet Orun:

And as a military engineer, you are always given a mission and you have what you have. That is the typical mindset. And in every country, in every place, people are talking about artificial intelligence, what it may mean. And he said, “Okay, look, I think this is your field. Help me understand what is new versus what he was working with in older computing days, and why are people worried? Why are people excited?” So, I sat next to him. We brought up ChatGPT, and I asked a series of three questions. The first question was, I said, “Tell me what you know about my dad’s name, Sunday Orun, a retired engineer, a retired soldier, not even rank.” And it gave what rank he retired at, what branch of the military, where he went to school. Simple question, limited context.

I said, what else do you know about him? I said, him without the name. And I got information about article field-working for magazines in a couple of his books. Post-retirement, he did poetry, which is a wonderful way to retire. And he’s like, “Oh, it’s interesting. How does it know that?” I’m like, “Well, can people find your books in online storage?” He’s like, “Yes.” So, it’s available information. It can leverage all of these as it searches. He’s like, “Okay, it makes sense.” Then I asked the question, “What do you know about his son?” And ChatGPT says, “I do not know who his son is.” And he’s like, “So, why doesn’t it know we are related?” And I said, “Because the fact that you and I are related would be in a government database. It would not be in public records. It’s not something that’s on the internet.” And for him, this was an obvious separation.

So, you asked a question, this is a long-winded way of getting there, perhaps, so what has changed? A year ago, I could dump a bunch of knowledge articles or perhaps a meeting transcript and say summarize, or I could use knowledge articles to power chatbots. Now I can look at what do I know about a person in their transactional context based on their order history, based on their case history, based on their knowledge of the product. And I can give much more personalized recommendations, because the AI platform idea, as opposed to an LLM technology idea, is bringing together matching technology where we used to think about as duplicate management and CRM.

That mindset has evolved, and it is there to provide contextual interactions. Data cloud is not just powering the generative AI capabilities for Agentforce. It’s also providing the unified insights that can even be constrained to only what a person is supposed to know about, where admins and architects can control this, given the permission model and capabilities of flow. Which for me is incredibly exciting, because that means we can deliver more value. We can use the technology we are already deeply familiar with, and we can show the true potential of AI while minimizing risk to our organizations and minimizing confusion for our end users.

Mike Gerholdt:

That’s a fabulous story. I feel you’re spot on. Just the level of understanding why and what we have available to us is huge. In the email you sent me, I want to pull out a sentence, because we’re talking about data and we’re talking about a lot of things. But I feel this is a good foundation. You said, “Part of this helped them realize why historical CRM data management techniques do not scale versus benefits of data cloud.” To the uninitiated, and I’m one of them, so I’m asking this question for me, can you give me what you mean by historical CRM data management techniques and help me understand that versus the benefits of data cloud?

Mehmet Orun:

So, if I think about what is in the Salesforce admin data management toolkit, we talk about a distinct set of areas we expect admins to do. They configure objects, object fields with validation rules and some data management rules, such as, do you want a default value or not, if it’s required or not? We talk to them about duplicate management rules, which led the impression that all duplicates are bad. And we talk about storage optimization more around performance, because every org had a storage limit, you wanted to think about when you may want to offload storage either for cost savings or build skinny tables for large data volume handling. Those were the domains of data management we got to, which was fairly technical, focused on mostly data entry operations. Let’s fast-forward to even two years ago.

If you have a Salesforce CRM org with Experience Cloud, you need to have intentional duplicate records because the records and end user maintains their information should be separated, then how that customer’s information is maintained by employees. You may also have records maintained by partners using Experience Cloud that’s still about the same customer. So, already thinking that for a customer that should have one record is no longer sufficient and acceptable, because partners need to have their view of the information. Customers want to maintain their own perspective of what they’re called, what’s their best contact information. And companies want to be able to have their internal view as well, such as customer segment, customer risks, so on and so forth.

But personalized engagement requires a complete understanding of what’s happening with an organization, where you only act on information you’re allowed to see and you act on insights that is relevant to the outcomes you want to achieve. So, three things I really, really like that data cloud brought in to our solution kits is, first, I can provide a holistic understanding of the individual or a business contact, even though I have multiple contact or lead records in my CRM, even in the simplest of architectures. Let’s talk about a nonprofit example. Let’s say that we’re talking to Sam Smith, and Sam is a donor. Sam was a board member. Sam worked for an organization that gave us grants.

That is us interacting with Sam, the human, into a business context and in a donor relationship. We are going to want to track these through different departments, probably through different records. But when we want to know what do we know about the people we engage with, how do we send them a personalized thank you, this is where data cloud powers that unification perspective. Does that example make sense before I tie to the AI specific examples that extends this?

Mike Gerholdt:

Yeah, no, it does. I’m following along.

Mehmet Orun:

So, let’s say that we are now in a data model that we have accepted, we should maintain contextual transactions in our business applications, whether we have one or multiple CRM orgs, and of course, other systems. We first unify it around individuals’ business contacts and accounts. So, now, related transactions, related emails, donation history from external systems or cases, regardless of your industry, can come together in one umbrella. Now, if I want to create a personalized thank you message, we can look at overall interaction history and not just think that we have seen someone for the first time, because they’re using their new email address in their new corporate role, but they’ve been a lifetime member.

So, generative AI solutions work better when interactions across a person’s contact points can be made accessible within compliance rules, of course. And agentic solutions work better when it can understand what are all of the different type of transactions that may be associated to an individual or an account, even when they’re distributed across multiple account records, multiple contact records, even multiple CRM orgs. You can see me, I’m pointing to things on the whiteboard in front of me, but this is something that used to take organizations months, if not years to put in place. And having done this now for real with a few nonprofits as part of my pro bono work, I know we can do assessment and planning in a few days.

We can then onboard the data and configure data clouds, data unification capabilities in less than a month. And that includes identifying bad data that is in the system. [email protected], they’re still present, whether your org is three years old or 20 years old, by filtering out irrelevant data, by putting the right data standardizations in place. These are all part of a single umbrella of capability. Whereas an admin, you just worked with the admin tools in the past. And now many of these transformation capabilities, configurable rules are accessible still under the setup tree, still under the Salesforce tabs. That allows us to be more productive Salesforce professionals and allows us to decrease the total cost of ownership as we support our organizations.

Mike Gerholdt:

I mean, I’ve always thought when I’ve asked people a rhetorical question, what is the most important thing that your company owns? And 99% of the time when I ask people that question, they get it wrong, because they mention a patent or a brand, or a product that they produce. And I say, “No, it’s your data.” The data that you have is the most important thing for you to take care of. And ironically, it’s also the most least paid attention to, because we just throw things in, and we’ll sort it and figure it out later. Hurry up, move on to the next thing. And now, as you bring up the unification of all these systems, where we’ve put all this data and the management or mismanagement of it, now is the vital importance, because now we can truly link all of this information, and have AI sort through it and give us the relevant information that we need by just thinking through a few more processes.

Mehmet Orun:

I think what’s important in what you said is AI is additive to what we have had, because I agree data is the most important asset. And the fact that no organization I’ve ever been a part of or helped had perfect data. It’s something we just need to accept, but not live with.

Mike Gerholdt:

Right. I have a friend that has a small marketing agency. He probably has 200 people in his little CRM. I promise you, his data isn’t good. Even in that, I mean, nobody’s got perfect data.

Mehmet Orun:

So, what matters, and this is what we talked about last year, is we can’t assess data quality as a technical concept. We can’t just look at what is in my object. Is it good? Is it not good? We always need to look at data in context of a business outcome. I think an example I give often is how much data you need to start an opportunity is different than the amount of data you need to close an opportunity. What you want to gather if you lost a big opportunity is different than what you probably would ask people to capture if you lost a small opportunity. So, these are all proportionate to the business benefit, where I don’t think historically we did a great job explaining as professionals, whether we are admins, architects, business analysts.

But when it comes to AI, because agentic AI puts so much focus and emphasis on use cases and the persona we are empowering. If it’s a sales agent, we want to find out what is the job a sales agent is supposed to do. What is the information they need? What are the rules they should follow? And whether you have 100 fields or 800 fields in your constant opportunities objects, we still need to look at what data is reliable today, if that’s sufficient. If it is not sufficient, we need to go through some type of a data improvement process or when to look at a different use case. If we have sufficiently reliable data, we need to look at how do we ensure our prompts both use data from those fields that have reliable data and sufficient metadata, and also know when a subset of records don’t have sufficient data quality in those very same fields.

And then third, just because it works today, we shouldn’t assume things are going to be the same tomorrow because processes are changing, configurations are changing, people habits are changing. So, by monitoring what’s happening in our business applications and catching deviations, we can avoid unexpected batch surprises, also in the flows. Honestly, these are things with a time machine we should have thought of and incorporated into our automation flow, into our reports, but the attention wasn’t there as much as it is today. So, people being excited about AI, but worried about hallucination risk is one of the best things that happened to ensure we can provide reliable data for all types of decision making through Salesforce.

Mike Gerholdt:

Right. Well, I mean, what do they say? 2020’s hindsight? If you could go back and know the future, then you’d obviously plan for it, but it also creates opportunity for us to be creative and corrective in how we move forward. Which means that every solution you’re thinking of today moving forward is going to look very different than before Agentforce.

Mehmet Orun:

One of the things I’m still noodling on, and I’ll probably spend a few more years noodling is how do we make sure we can take better advantage of unstructured data that is the vast maturity of all interactions. I remember being excited about Einstein Activity Capture, which was a few years ago. And it’s still an untapped potential, but now we are analyzing that data, we’re incorporating that data. The more we can streamline the end user experience to capture information, know when information may be missing, incomplete, potentially out of date, so they can improve it in a tactical, surgical way, and then be able to explain to them why we’re making certain recommendations in AI assisted suggestions.

I think that’s also going to increase the confidence for agentic experiences, where humans are engaged in a secondary level. I know that I would like AI tools to give me results I can believe in when I’m directly engaging first, before I’m willing to expose it to perhaps less savvy or less aware of my underlying processes end user. I think a lot of people are going to go through the journey, so thinking about process mapping, thinking about testing strategies are also going to be important considerations for all of us.

Mike Gerholdt:

I mean, that’s the whole point, is to test. You want something reliable and to ask why. I think that the important thing is people get things wrong, too. We sometimes look at some of these technology solutions as infallible, as they’re always perfect and they’re not. They’re imperfect because they’re built by imperfect people. But that doesn’t mean that you can’t constantly iterate on your solution. I remember long time ago when I was an admin, I feel like it was Josh Burke or it was another developer who was always like, “Every year, I look at the code I wrote for the previous year and wonder why did I write it that way.” And it’s because you’re a year smarter.

Mehmet Orun:

100% agree. I have 13 years of Salesforce presentations in my Dropbox folder. And when I look at it’s fascinating to see what is still true. And it’s interesting to see when some recommendations have completely changed over the course of the last 15 years, because we are learning. And I think we need to be honest about, look, yes, this was the recommendation based on what we knew, here’s what we learned since, and here’s why we are recommending X today that is different. I think on the other side, as professionals, we need to remember none of us have all the answers. And what we knew yesterday might have changed today. So, look, I love your podcast. I love some of the things that come out of the various blogs, because people share what they’ve learned at a level of frankness, including what we stopped doing. And that is a sign us being learning humans. And it’s the best way to be.

Mike Gerholdt:

Absolutely.

Mehmet Orun:

How do we learn to be better every single day?

Mike Gerholdt:

Well, I feel like that’s a really good place to end this episode on, because I always want to learn more. And I appreciate you coming on the podcast and helping everybody else learn more.

Mehmet Orun:

It’s my pleasure. I know that there are so many thoughts we can always get into. I hope these sessions enable more personal connections. And if you’re listening to it and we run into each other at an event, let’s grab coffee. Let’s talk about data or life, because we’re going to learn from each other. We will make each other better. And thank you, Mike, for this opportunity.

Mike Gerholdt:

Absolutely. So, Mehmet took us for a ride from a 90-year-old general, his father, all the way to data that doesn’t quite behave. And the takeaway, well, AI is like a great intern. It’s only as good as the notes you give it. So, let’s feed it well and ask better questions. But anyway, a huge thanks to Mehmet for the wisdom and the stories. If you learn something today or you just enjoyed the ride, can you do me a favor and just share the podcast and spread the data love? Now, until next time, we’ll see you in the cloud.

The post How Data Cloud Enhances Contextual AI for Salesforce Admins appeared first on Salesforce Admins.