Pod Insights MarTech Podcast ™ // Marketing + Technology = Business Growth Mastering AI Personalization with Customer Identity Data About this episode AI personalization requires unified customer identity data. Joyce Gordon, Head of AI at Amperity, explains how brands are moving from broad segments to micro-segments using generative AI. She outlines the critical role of identity resolution in creating personalized experiences, demonstrates how Model Context Protocol (MCP) servers enable data integration across systems, and shares practical frameworks for implementing AI personalization with proper evaluation mechanisms. Show NotesConnect With:Joyce Gordon: Website // LinkedInThe MarTech Podcast: Email // LinkedIn // TwitterBenjamin Shapiro: Website // LinkedIn // TwitterSee Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.
Transcript 0:00 Speaker 00 The Martech Podcast is a proud member of the I Hear Everything Podcast Network. Looking to launch or scale your podcast? I Hear Everything delivers podcast production, growth and monetization solutions that transform your words into profit. Ready to give your brand a voice? Then visit iHereEverything.com.
0:25 Speaker 01 From advertising to software as a service to data, across all of our programs and clients, we've seen a 55% to 65% open rate.
0:36 Speaker 02 Getting brands authentically integrated into content performs better than TV advertising.
0:42 Speaker 03 Typical lifespan of an article is about 24 to 36 hours.
0:46 Speaker 01 If we're reaching out to the right person with the right message and a clear call to action, then it's just a matter of timing.
0:53 Speaker 00 Welcome to the Martech Podcast, a member of the I Hear Everything Podcast Network. In this podcast, you'll hear the stories of world-class marketers that use technology to drive business results and achieve career success. Here's the host of the Martech Podcast, Benjamin Shapiro.
1:15 Speaker 04 69% of marketers say that they're unable to consistently deliver personalized experience, according to Salesforce's 2024 State of Marketing Report. That's almost seven out of 10 of us that can't figure out how to turn an abundance of customer data into a tailored experience. Here's why. Your customer data is in silos. Your tools need clean and unified data to deliver results. So how can you turn your data chaos into the type of customer experiences your prospects and your customers expect? I'm Benjamin Shapiro and joining me today is Joyce Gordon, the Head of AI at Imparity, which helps unify fragmented customer data to powerful personalized marketing.
1:57 Speaker 04 And Joyce is going to explain how you can combine AI and a unified data foundation to transform your customer experience strategy. Joyce, welcome to the Martech Podcast.
2:07 Speaker 05 Thanks so much, Ben. It's really great to be here.
2:10 Speaker 04 Excited to have you on the show. Excited to dig into this combination of data and artificial intelligence and personalization. I feel like it's the topic of the year.
2:20 Speaker 05 It's been the topic of several years.
2:22 Speaker 04 Well, the ball bounced your way here. You're an AI expert and all of a sudden that makes you like the belle of the ball. So before we get into talking about AI and personalization, you're one of the people that actually know how this stuff works. Give me a quick bio on what it means to be an AI expert. Like, what are the skills you have and the experiences you had to make you actually an expert in artificial intelligence, not just using ChatGPT?
2:46 Speaker 05 Absolutely. Well, I've been working in AI and ML for the last decade. I started my product management career at a really small startup. I joined right after the seed round called Custora, which was focused on customer analytics, helped build the product team there and grew it from seed stage through Ampariti's acquisition. We grew from about seven to 70 people during that time. And at my time at Ampariti, I have focused on everything from building a platform for marketers to building our predictive analytics and machine learning project to now AI. I also studied statistics and math in college. So this is a space I've been in and a deep interest I've had for a really long time.
3:27 Speaker 04 All right. So when we're talking to you, we're not just getting the perspective of somebody that knows how to use ChatGPT and Claude. You actually understand how they work. Explain to me how AI has changed personalization in the past year.
3:39 Speaker 05 Brands have been talking about one-to-one personalization forever, basically since the dawn of time. But it's kind of been a dream that's super challenging to realize in the past. And the reason is one of creative constraints. Creative teams are continuously underwater and there never really was enough creative available to make one-to-one personalization a reality. Generative AI takes the cost of content creation and drives it to zero. So what that means over time is the creative constraints we've faced in the past, they're going to start to lessen. Now, it's not so easy. It's not as if you can take GenAI, throw it at your data, create one-to-one personalized versions for everyone. There's still a lot of challenges let down in the way. And so brands are starting to move in
4:23 Speaker 05 the direction of one-to-one personalization, but they're not totally there yet. Where we typically see brands start is with a segmented approach. So in the old world, Before Gen AI, maybe you had enough creative for five different segments when you were launching a campaign. Now, maybe with Gen AI and using AI to create the content, like the imagery, some of the email subject lines, some of the actual copy, maybe you can actually launch a campaign with 20 segments. And what most brands are doing now is they've got a human in the loop. So humans are reviewing the output, they're making sure the AI does things like match the brand voice, is sending the message they want, doesn't have any legal liabilities. So over the past year, we've seen broader segments to more narrow segments.
5:07 Speaker 05 And I think over the next year, we'll see even more movement towards one-to-one personalization. So those segments will get smaller and smaller as brands get better at using the models, as they lay their data foundations, as they think about how to introduce Gen AI into the workflow. And then also as they build systems to actually evaluate the Gen AI content at scale, enabling them to figure out where to focus.
5:28 Speaker 04 Let's talk a little bit about building the structure because it seems like that's the topic of the month right now. I've been reading and trying to research the concept of MCP servers, but it seems like everything we've been doing is human in the loop. I am writing a prompt, I am dropping some data in, I'm analyzing it, and then I'm trying to figure out how to apply that. Now we have this notion of being able to tap into your data sources and have LLMs be able to unify everything and make decisions and then actually take actions. We're sort of getting into the agentic part of the agentic AI era. Am I right? How big is this MCP servers? Help me boil this down.
6:08 Speaker 05 Let's start with what is an MCP server? So MCP stands for Model Context Protocol.
6:14 Speaker 04 I knew that one.
6:16 Speaker 05 Well, you're already ahead of the game. This was a standard rolled out by Anthropic. And really what it is, super simply, is it's kind of like an API for agents. It's how different agents talk to one another. So if you've got your customer data on one MCP server, you might want to get that data to your conversational AI agent on your website. Because if you think about conversational AI, for example, if I'm shopping for a dress to wear to my friend's wedding in Hawaii, you really want that conversational experience to understand what products I've purchased in the past, whether I'm in the loyalty program, what items I'm interested in. So the MCP server is a protocol that makes that data available to other agents, like that conversational agent.
7:00 Speaker 04 One of the things that in my research, I watched a bunch of YouTube videos basically on like, what is an MCP server? It seems like this is really a hot topic right now.
7:08 Speaker 05 Yeah.
7:09 Speaker 04 The best description that I had is all of these different tools, they can all speak, but they all speak different languages. Your Salesforce database speaks English, your Marketo database speaks French, and the MCP server is essentially the translator. You feed all of the data into it and it understands all of the language. So then you could have it all combined and read back into a single language. So it is the filtering layer that helps take all of the disparate data and unifies it into kind of a common language that you can then interpret through an LLM.
7:42 Speaker 05 Yeah, that's exactly right. I absolutely love that description. To your point then, it's a common language where all of your different agents can talk to one another. And I think one thing that's interesting to think about is where things are and how they might evolve. So I would say as we think about conversational AI experiences today, there's a lot of challenges there. Like there was the Air Canada example.
8:06 Speaker 04 Yeah, give me the Air Canada example. What happened?
8:08 Speaker 05 So Air Canada provided a fare to someone. I don't remember the exact details, but I think the AI incorrectly offered a customer a discount that should not have been offered.
8:19 Speaker 04 All right. Well, at least the door didn't blow off.
8:20 Speaker 05 Yes, exactly. But there are challenges around the AIs providing offers or experiences that might not exist. So hallucinations, there are legal challenges, obviously there are brown voice challenges. So where we typically see brands start with these conversational AIs is to really constrain the use case. So if you're a ski slope, for example, maybe your AI to start is only going to answer questions about ski school and not what meals you could buy. It's not going to let you book lift tickets. And that's super important because if you've got a constrained use case, it limits the risk and enables you to really deliver a resonant customer experience. And then you can expand beyond that over time.
9:00 Speaker 04 I interviewed Jordan Crawford, who's a go-to-market specialist, and he's like next-level AI guy, too. And he always uses the example of, I think it was a car company like Toyota, where you can go on to their customer service chatbot and then ask it to write a Python script for you, and it'll like actually do a pretty good job. So it's like you need some parameters to be able to deliver these personalized experiences. We can feed real-time data, customer data, in theory, using an MCP server. Who's doing this well? Who's actually able to create customer experiences that are personalized and tailored? And where are they kind of falling flat?
9:36 Speaker 05 I would say the brands that are doing this well have kind of constrained the use cases to start. We're all still learning about this technology. It's hard to believe that ChatGPT was launched in November 2022, so we're still very early. I would say the companies that are doing things well today, and many of them are on parity customers, are focused primarily on how can we move from bigger segments to more micro segments? So you might not even notice as the customer, but you are getting a more customized email based off your loyalty status, your churn status, your price point than you were before. And then we've started to see conversational AI take off when it's focused on a specific use case as opposed to saying we can answer any question for the brand. So the example
10:21 Speaker 05 with the ski slope I mentioned before is like an actual customer. So there's a lot of examples like that out on the market.
10:28 Speaker 04 Okay, so we've got this notion that artificial intelligence is helping us move from sort of our static, we've got a couple of segments into very micro segments. Eventually, we're going to work our way down into one-to-one personalization. Give me the framework to utilize customer data for that one-to-one AI personalization.
10:48 Speaker 05 So before the framework, I think my overarching feedback probably as a product manager, this is just life philosophy, but always start with the use case. If you've got kind of a loyalty chatbot experience, the data you might need and the complexity of that is going to be different than if you are starting with a use case where you have a chatbot that needs to answer anything. So start with the use case. But once you've got the Use case, I'll talk through a couple items that are just really helpful from a data perspective. The first is if you have a conversational AI and you really want that AI to be able to answer a lot of questions about your brand. You want it to be able to answer questions about loyalty. You want it to be able to help find a dress toward your friend's wedding.
11:27 Speaker 05 You want it to understand how to recommend credit card offers to people so they can sign up for your credit card. It's going to need a lot of data across different sources and being able to deliver comprehensive experience there really starts with identity. So the first piece is developing an identity spine. So that's how do you know that the same customer shopped online versus in store versus transacted with your loyalty program?
11:52 Speaker 04 An identity spine. I've never heard that one before.
11:55 Speaker 05 That's a little bit of an ampanism. It's the idea that we can understand who Ben is across all of your different data sources and touchpoints. And that's so critical for a conversational AI experience, particularly if these experiences seem human-like. Customers' expectations are even higher than if the AI seemed a little bit more robotic. So that's really where to start. The second piece is making sure you've got the right attributes on the customer. So what are the things about Ben that are really going to play a role in curating that experience? These could be things like your predicted customer lifetime value, what products you'd purchased in the past,
12:34 Speaker 05 even embeddings containing maybe which types of creative you've responded to. So all of that data is going to be really key.
12:42 Speaker 04 Before you go on, I'm sorry to interrupt you. I want to ask a question about that. How much is too much? So if I'm feeding customer data to an LLM and we're giving, let's take a use case of I'm selling retail. I have a T-shirt company and I'm selling gender, name, location, size, creative they've responded to, maybe even questions they've asked. But I also might be able to map to like pages they've visited, all sorts of other data and information. When you start giving the more nuanced information to an LLM, is there a sense where you can give too much information and it can't parse what's important? Important as opposed to like, hey, I'm going to prioritize that this person once
13:24 Speaker 04 visited a women's T-shirt section because I was buying something for my wife. And now all of a sudden, it's like, he likes men's and women's T-shirts. But I'm a man and I want to buy T-shirts that are large or extra large for men primarily. How do you give the context of what information is priority as opposed to just giving everything and letting the LLM decide?
13:46 Speaker 05 If you were to rewind, let's say, a year ago, context windows, it's the number of tokens you can pass an LLM.
13:53 Speaker 04 How much information?
13:54 Speaker 05 Exactly. How much information you could pass the LLM? They were much smaller than they are today. Now they're often millions of tokens you're able to input into a context window.
14:04 Speaker 04 Hundreds of pages. Go on.
14:06 Speaker 05 That said, Ben, you're bringing up a really important point. Just because you can pass all that information, it might be too much for the LLM to really reason over and you might start to get some responses that are not as curated as you want. One very core part of agents is this idea of tool calling. These are different tools the agent has that it's able to access. Let's say that we take your question. We use a lot of amparity in our AI experience, but let's say that we take a question like returning to this one. I'm shopping for a dress to wear to my friend's wedding in Hawaii. You might feed that question to the LLM and actually ask, hey, what types of data do you think would be helpful here? And then you might actually have a tool call that retrieves that data specifically and
14:51 Speaker 05 passes it to the context. So having as much data as you can, as long as it's high quality is great. But then you're going to need to have some sort of retrieval step that looks up what are the important pieces of data for that question and probably passes that in. Now, all of this is trial and error. So for some rounds, maybe it's fine to actually pass the full customer profile to the context. But if you're starting to get answers that seem irrelevant, using more of a tool calling or like a RAG-like approach could be really helpful at honing in on the responses you're looking for.
15:23 Speaker 04 If I ever go through a career change and decide to go down the coaching route, I feel like my ethos and tagline would be make a plan to make a plan. And what you're talking about is akin to that, but for AI, which is you are giving information to have AI write a prompt. So AI can execute the prompt with the right information. And I know it sounds a little circular, but I cannot stress how much in everything that I've done in building our AI agents, like the production of our podcasts and our research and our guest sourcing, it's always here's a bunch of information. Help me write a prompt that I can use to distill this down and execute this task. And then I take the prompt that's returned and run it.
16:08 Speaker 04 And it does a much better job than if I just try to write the prompt myself. And essentially you're saying the same thing is you can't just feed the raw data in and say, give me the answer. It's feed the raw data in and say, refine it to figure out what I need. And then you execute based on that middle prompt.
16:24 Speaker 05 That's exactly right. To take things back to Amparity for one sec, we have a product called Amp AI, where you can ask and answer any questions about all of your customer data, which could be hundreds of tables with hundreds of fields each. It's really complex. And when you ask a question, actually, the first thing Amp AI does is it formulates a plan for what are the steps it needs to take to answer that question? And what's the research it needs to do along the way? Like, what are the research queries it needs to get? And what context does it need to load in order to answer your question? So Ben, that's very similar to the way you're thinking about things.
16:57 Speaker 04 I did this yesterday. I swear to God, I built one of our first AI agents, or what I'm calling an AI agent, for figuring out who should be on the podcast. And I have this big context document. Here's everything you need to know about the Martech podcast. Here's a second document that's a scoring rubric. So here's how we evaluate our guests. Like, here's the one out of 100 point scale. And then here's all the searches you have to do to be able to evaluate and effectively figure out what the scores should be. And then I take all of that information and I'm like, all right, here's the context of a guest application. Read through the context document, understand the scoring rubric. Conduct all these searches, now assign the scores. And it's exactly what you're talking about with Aparity. It's like it needs to think of how to execute the task,
17:42 Speaker 04 what research it has to do before it tries to just go and do it. Again, make a plan to make a plan, write a prompt to write a prompt. There's this repetitive task of refinement and strategy that you and the AI need to do to effectively execute.
17:56 Speaker 05 That's totally right. Yeah. When I think about success with agents, I really think about three things. Number one, do you have the right underlying data? Number two, it's all about the tools. So that's like making the plan and making sure the plan has access to the right tools to pull in the right data and also modify the agent's responses. And then the third piece, which we haven't touched on as much, is the evaluation. The evaluation is really, really important. How do you judge on an always-on basis if the agent's responses are good or not good? And more importantly, for the situations where things are not going as well as you'd like, how do you actually root cause what those challenges are? Because that feedback cycle is imperative for taking that information and actually using it to refine the outputs in your approach.
18:38 Speaker 04 The AI can learn and has memory. And when you're executing a prompt, you need to be able to feed back whether it was a successful execution or not. And the AI will start to refine itself. Let me ask you a question because I've been spending a lot of time, like I said, building our ranking agents and our topic sourcing, things like that. Trying to make each of our interviews a personalized experience. It's a ton of work right now. So what are the ways that you can implement an AI personalized experience without taking on truckloads of dev work?
19:09 Speaker 05 I would say a conversational AI is probably not the place to start, especially if you're new. I would start with something less risky and human in the loop. And you can actually get started really, really quickly here. You need some data, you need an LLM, and you need some people to review the output. Like you can even do it in ChatGPT, you could do it in Databricks. So I would start with a really constrained use case. So maybe, for example, you want to personalize your loyalty welcome email. And for that, you might need to know information on like someone's product. Preferences, maybe their gender, maybe their age, maybe what their signup source was. And then you can take that with a couple of examples for how you would actually want to personalize the email,
19:54 Speaker 05 pass it to the LLM and then have the LLM produce the output. And I would say one of my biggest tips are people will often try to use LLMs to personalize and they're like, this doesn't match her brown voice. It's not what we're looking for. If you can pass five to 10 examples of how you would actually personalize the output based off different pieces of Cospur data, you'll get far better results than if you just leave it all up to the LLM.
20:20 Speaker 04 So what matters more? Does it matter more that you give a tone of voice or some structure and rules for it to try to interpret? Or is it better to give less like, here is our style guide for the chatbot? And more, here's other examples of copy. Is it better to try to define it upfront or just give examples and let it interpret and figure out the style?
20:43 Speaker 05 This is anecdotal experience. There's no hard and fast rule here. I find that if you've got high quality examples, that's typically the best route. And then typically what I'll do is I'll pass the high quality examples or look at the outputs. And if there are things I don't like about the outputs, I'll say, don't do that. For example, I was prototyping an AI that would produce personalized emails. And one of the inputs was whether it was a family shopper or whether it was an individual shopper, so whether it was just a person shopping for themselves. And in the personalized email, it would say something like, oh, is your shopping for you and your kids? Here are things you might be interested in. And we all know just because you've bought kids items in the past,
21:27 Speaker 05 it doesn't necessarily mean that you have children or you're shopping for them. So in addition to the examples, I put in a rule and I said, if you see a family shopper, you can show some kids items, but don't specifically call out in the copy that they're shopping for their kids. So that iterative approach is really helpful and it ended up being a mix of examples and then also rules that you're going to hard code into the prompt.
21:49 Speaker 04 Yeah, whenever I am writing LinkedIn posts, if I give examples to an LLM, it always tends to too closely replicate the examples. I had this problem where in every LinkedIn post, I was using an LLM to help me edit. It always put in the term game changer, because in the two of the five examples that I had, it was like, this is a total game changer. So it always put that term, interpreting it as my tone. And I'm like, all right, but it's too formulaic. So I want something that is unique and change around how we do it. So I kind of lean towards here's rules and style guides with key phrases to use, but don't be too repetitive. Instead of giving a ton of examples, it's chicken or the egg, you could probably do it both ways.
22:33 Speaker 05 Yeah, and I think in this example, one thing I may have tried, and there's no hard and fast rule here to your point. I may have gone back and said, use the examples as inspiration, but don't replicate them too closely and do not use these following phrases or something like that.
22:49 Speaker 04 All right, so when you're doing personalization, you're basically saying, start with human in the loop. You're taking your data, manually uploading it, giving it to an LLM and having somebody publish it. You're not automating the process first to figure out if it works. Talk to me about when you're putting things into production, how are you taking that manual process and then building a personalized experience that doesn't need a person?
23:13 Speaker 05 I think this is really the billion dollar question. Most brands are still in the human in the loop phase. If you wanted to move more towards like an always on, one to one personalized experience, obviously you need to figure out how all of this fits in your workflow. You don't want to be uploading creative or email subject lines to your tool. It needs to all flow through. But the really big challenge here is one of evaluations. So you're going to need a system that continuously evaluates the quality of the LLM outputs before it actually goes out in market. And a really common approach here is what's called LLM as a judge. So you actually ask an LLM to review the work of another LLM. And it's up to your brand to really
23:58 Speaker 05 Define what the criteria are. So some things I've seen in the past are relevant. So if there are specific pieces of customer data that you want to make sure definitely is to personalize, does the output actually include those? Some other things are around safety. Is there anything offensive in the output that maybe should be flagged? And then there are also things around offers or other pieces of information that could be dynamically added, making sure they're in line with your brand. So the brand needs to sit down and kind of think about what are the key checks that they want always to be done on the creative? And then certain creative will probably get flagged for human review to take a look. But as I mentioned, most brands are really moving in this direction,
24:38 Speaker 05 but still using a segmented approach today just because it's so important to get those evals right and it's quite a challenge.
24:44 Speaker 04 We're still using AI to execute manual processes for segmentation and building personalized experience. It's doing the work and we're the editor and publisher. Give me your forecast for what's the next step? What does the future look like when it comes to these personalized experiences?
25:00 Speaker 05 Well, maybe take this in a slightly different direction relative to our conversations so far. I've seen a couple announcements over the past several months that I think are really interesting. Stripe announced an order intense API where it can purchase on behalf of users. Perplexity and Amazon announced something similar where they're also able to kind of purchase on behalf of users. So I think what this points to is we're going to see more aggregator platforms, like the Amazons of the world, the perplexities, the Open AIs have their own shopping agents. So instead of going to a brand directly to shop, you might go to one of these agents and say, hey, I'm going skiing. I failed this weekend. Can you help me identify a jacket?
25:41 Speaker 05 And what those shopping agents might do is using MCP or potentially a different approach, we'll have to see. They're going to reach out to different retailers or different brands in general. And they're going to pass some information on the user's context and question and ask for recommendations for that given question. So with that fail skiing example, they're going to send along information about the user's question, demographics and Try to get recommendations back and then they'll kind of be the aggregator for what those recommendations are. And the user will be able to purchase actually directly through those experiences. What that means for brands is maintaining the first party connection is going to become a little more challenging and really important.
26:26 Speaker 05 So there's a couple of things that I think are really important there. Obviously, product quality, differentiation and brand are important, but also things like making sure you're delivering a fantastic personalized experience on your own website or property. And loyalty are going to be really important for maintaining that first party relationship as well. So I think we'll see more personalization within the context of brands and their experiences and also more agentic personalization with aggregators as well.
26:52 Speaker 04 Yeah, it's funny. My interpretation of what you're saying is your website's not going to matter anymore. And I know that you're saying you have to build a better personalized experience. There will still be the value of brand and probably loyalty programs. And that's how companies will drive sort of first purchase conversions. But it sounds like I'm looking for a ski jacket, rank all the ski jackets for me. Here's the ones that I liked in the past. Ooh, I like this color. No, I don't like that one. You know what? Send me all three and then you can execute in the LLM to actually conduct the purchase and have the product shipped. So now we don't actually have to go from I did my Google search. I clicked to the link. I'm in the first party website. I placed the transaction.
27:34 Speaker 04 You also don't necessarily have to go on to Amazon to execute. So there's this sort of interesting change. First party data is going to be much harder to come by if everybody is executing their transactions in the LLMs. Even if it means there is more volume, more velocity, less friction, less data. It's going to be a brand new world.
27:55 Speaker 05 An analogy that I think about often is Amazon. I think there's a lot of parallels. So some brands, they basically make their entire business on selling on Amazon and that works for them. There are some brands that don't sell on Amazon at all. And then there's some brands that do some of their transactions on Amazon, but not all of them. And for the brands in the middle in particular, I think we've all gotten the experience where you get a package in the mail and there's a little piece of paper in the package inviting you to sign up for loyalty or offering you a special offer. So they can tie that transaction back to you. I think brands are really going to be focused on ways to maintain that customer relationship in light of this new world.
28:34 Speaker 04 All right, and that wraps up this episode of the MarTech Podcast. Thanks to Joyce Gordon, the Head of Generative AI at Imperity for joining us. If you'd like to contact Joyce, you can find a link to her LinkedIn profile in our show notes or visit martechpod.com. You can also visit her company's website, which is imperity.com. And if you haven't subscribed yet and you want a daily stream of marketing and technology knowledge in your podcast feed, hit the subscribe button in your podcast app or over on YouTube, and we'll be back in your feed every week. All right, that's it for today. But until next time, my advice is to just focus on keeping your customers happy.
29:32 Speaker 00 Thanks for listening to the MarTech Podcast, an I Hear Everything production. Looking to launch or scale a podcast like this one for your brand? Then visit ihereverything.com.
Previous Master existing tools vs invest in new AI technology Next Which team should be responsible for integrating AI into your product? Page updated November 7, 2026 · transcript and summary are machine-generated
