Leader Spotlight: Building an AI-native product inside a 75-year-old brand, with Mike Bal
Mike Bal is Head of Product & AI at David’s Bridal, where he leads Pearl Planner, an AI-native wedding planning platform built on knowledge graph architecture, agentic AI, and multiple specialized LLMs. Before David’s Bridal, he worked at Automattic, where he led Woo Express and the Blaze Ad Platform. Earlier in his career, he held roles at Engineered Innovation Group, 10up, and DocuSign, after starting out in digital marketing.
In this conversation, Mike talks about building an AI-native product inside a 75-year-old brick-and-mortar retailer and why he doesn’t think AI is married to any one business problem. He walks through the graph database that lets Pearl Planner understand a bride’s personal, interrelated preferences, his guiding principle for when to surface AI and when to let it work invisibly in the background, and the 20-plus-agent architecture running behind the scenes. He also shares the story of a critical API dependency that got deprecated days before launch and what the team did about it.
Betting on AI at a legacy company
Why make such a significant bet on AI, and what business problem were you actually trying to solve?
I think as a technology, AI is not married to one business problem or one customer problem. AI is an opportunity. Having been close to engineering and close to design, but never been the person who’s as good or able to actually do that work myself, AI is the thing that really takes down the barriers so you don’t have blockers.
So if you’re a resource constraint, say you have a tight budget, you have something you’ve committed to for the year, you can’t pivot as quickly as maybe a new startup would with funding sitting in the bank and three people, you can use AI to give the org, the team, or the company, more agility in terms of what can be turned around and who could do what.
The other piece of it comes from a UX perspective. Having built software for years, I couldn’t tell you how many times we came up with an ideal experience and then heard “That’s too complicated or too expensive to build,” or, “We’re not going to be able to do that right now. It’s not possible.” AI shortcuts most of that in a lot of ways. The standard in the industry for a long time has been that if you want information, you put a form in front of the user and try to get them to fill it out.
When we talk to brides, they tell us that one of the hardest things about planning the wedding is getting other people to understand their vision because it’s hard to put into words. And so a really good opportunity we saw early on was like, “What if we didn’t make them tell us? What if they just showed us?” So instead of filling out a form or going through multiple steps, the bride literally just shows us pictures of things they like, and then we leverage our expertise as a business that’s been in the wedding industry for 75 years to identify the things that actually matter to the decision-making process. We make a bunch of decisions on the backend that brides don’t have to worry about, regarding what we surface or how we personalize the experience. Little things like that just make it much more streamlined to what they need, especially in a space where they’ve been underserved.
Are there changes in the market that influenced that decision, like the growth of digital wedding sites like The Knot?
When you have a chance to start fresh, you don’t have legacy tech debt or a platform that you’re already invested in. What we saw early on was that nobody had actually done planning from day one through the day of the event. It’s usually just a checklist with the big milestones, but most people can think of those milestones off the top of their head, so that’s not very helpful.
It’s also skewed toward where those companies monetize, and brides see that instantly. When our platform launched, brides were like, “Hey, why are there so many dresses in here? That feels too pushy. It feels like you’re just trying to sell dresses.” It was only like 5 tasks out of 300, but they noticed it right away. So we took a different approach. When you start fresh, you don’t have existing infrastructure, legacy revenue, and restrictions, so you can just focus on the human, which is nice.
Building a data foundation for something deeply personal
In adding new technology to an existing business with legacy systems, were there any gaps that you found in data that you had to try to address?
I’ve never worked with a company that hasn’t had data gaps as an ongoing thing. There’s always improvements to make. But what I brought was the experience in the new tech that says, “Based on where things are going, the standard database is not how we want to solve these very personal interrelated things for storing and recalling and doing all this.” So we went with graph storage so that we could emphasize the relationship between — “My wedding color is red. I don’t know which shade of red it is, but this one is too dark, this one is too light, and my shade is somewhere in the middle.” Or, “I like this color because it suits my theme and I have a personal connection to it.” All those things make that experience much better when they talk to the AI or when we surface a recommendation or even when we default to certain settings to just shortcut a process so that a bride doesn’t have to make as many decisions.
We had to work out things like whether it’s the same user across touchpoints, or whether a wedding date means the same thing in every context — that’s the semantic layer we built out. Luckily, David’s has invested over the years in a strong data foundation. We have Snowflake, so all of our data ends up in one big place that we can pull forward. We have our own middle layer that actually transacts data from one point to another so that if we change our platforms over time, we still have the relationships and ways to pass things back and forth. That was one thing I was really impressed by from the tech side of David’s. Maybe it could have had better documentation, maybe it could have been updated a bit more, but overall the architecture was pretty solid on the data side, which I was happy with.
Deciding where AI shows up — and where it hides
As you rebuilt parts of that digital experience, were there certain principles that guided you in terms of where to bring in AI automation and where that human touch is still needed?
As we build the platform, the principle has been: if the user can do it, the AI can do it. So we have a budget feature: the user can enter an expense, or they can just upload an image, and the platform will scan it and put the expense in for them. Similarly, they can just type a chat to Pearl like, “Hey, I spent $150 on supplies for my centerpieces,” and she’ll add that expense to their budget. They don’t have to click to a tab, open the form, and manually enter the expense.
But over time, as they start pushing Pearl — “Actually I don’t like that you grouped these tasks in this milestone. Can you move four of them to a new one that’s called something else and push it back a month?” — Pearl can just do that. They might not be using Pearl that way yet because we have to do some education, but we’re building a future state where more people are assuming it can do that.
The other piece is that we don’t always have to tell the user we’re leveraging AI. If we have a color feature and we know the bride has picked out certain colors, we build on that without labeling it. When we generate palettes like, “Hey, how could this look with different palettes and different themes?” we don’t communicate that it’s an AI feature. We’re just like, “Hey, let’s see what it looks like with different combinations.” And then when the bride picks one, we preview what their vision board can look like with those colors applied. We don’t say that’s an AI feature, it’s just part of the experience.
We’re working on a new version of the vision board right now, and we want it to feel like magic. The only case where I tend to say, “Let’s explain a little bit,” is when the AI is doing a lot more behind the scenes. So if it’s like, “I checked your wedding day, I looked at information from your venue, I checked other comparisons, I read reviews, and here’s my recommendation.” I think it builds trust. There are certain scenarios where you should surface that you’re using AI, because otherwise it looks like you pulled something out of thin air and that could erode user trust.
For many things, it’s not necessary to communicate that we’re using AI, but if we’re sharing critical information or advice that’s needed to make a decision, then let’s at least verify we did the right things to get the information.
Taking anxiety out of wedding planning
You mentioned that one of those core challenges is the bride’s dress. Are there other examples of things you’ve done to reduce that anxiety level?
The budget tool is a good example. The perception is that a lot of brides already know everything about their wedding and have been planning it since they were little, but that’s really not the market. A lot of people we talk to are like, “Yeah, I figured I would get married, and every now and then think about it, but I don’t have a vision board. I don’t know anything about this.”
We take a more opinionated approach. The app says, “I already know my budget,” or, “I need help.” And if you need help, OK, where are you getting married? How many people do you want to have there? Elina, our president, had this plan a long time ago, and I thought it was one of her best ideas: What if we just asked them what things were most important to them and we did the math? So instead of like, “How much do you want to allocate toward apparel?” it’s like, “Do you want to splurge, or do you want to save here?”
And if it doesn’t check out, if they are going to have 300 people and they want to splurge on everything and their budget is small, we’re going to say, “Hey, here are the gaps. Let’s brainstorm where we can save.” They can lean on Pearl to help with some of those decisions. We don’t make them math at all. We can do the math. We just give them the results, let them think in human terms of what’s most important, what’s less important, and what they need. That gives them a starting point, and they can go on from there.
When your API dependency disappears days before launch
You mentioned that a few days before launch, there was a critical API dependency that was becoming deprecated and you had to pivot. Could you say more about that?
The initial idea, in line with everything we’d talked about, was: maybe brides have already collected a set of images in a popular social media tool, and it would be better if, instead of making them look through more images, they just showed us what they had. So we built it out that way. The fundamental principle, the shortest path: just share a link, we’ll go find the pictures, and you don’t have to do anything else. That was the ideal. Then we found out right before launch that the API was going away. So our backup plan became the primary plan, and we had to curate more and build a different loop for that content.
The fundamental principle is still there. The entire pipeline — look at these images, identify traits from different categories, create a file, send it downstream to be embedded — was all still there and usable. But the user experience side of it was a bit of a bummer. Since we’re attached to the appointment booking piece too, it didn’t hit as hard as I thought it would. We still saw good throughput for everybody who got that far — I think we had a 60-70% completion rate. And some people really loved it — instead of five pictures, they picked 20 or 30. That’s great data for us to have.
Building lean with AI-native tools
Operationally, you’ve described a relatively lean team that’s moving much faster than the traditional product organizations. Since you started incorporating more AI tools into the workflow, are there certain buckets of work that have disappeared or that you’ve gained very large efficiencies with?
We started out playing with Replit to prototype and get some initial code. Then we thought, “You know what? No. Cursor’s out. We can actually keep our engineering workflow.” My team did try Copilot for a little bit; but I was not a fan, and they ended up not being fans either. Cursor’s 10 times better.
It’s crazy how much product management has changed in the last year. Agent skills roll out, and I share the insights. And this is across David’s, to all company leaders. That’s kind of how we’re handling rollout — give them access, tell them what’s possible, let them apply it to their specific area.
One of our engineers, Valerie, did a workshop with the rest of the engineering team and IT, showing that, for each repo that she works in, she has her own set of skills for the agent, different workflows and changes in her system prompts so it always checks the design system, or it always goes to Figma and pulls in the code. She’s a 10x engineer on her own — not because she’s working ridiculous hours, but because of the way she thinks about getting the work done and the way she thinks about building the systems to do the work.
More broadly, how do you think about architecting AI adoption across the organization?
At David’s, we’re leaning into Claude pretty heavily. I would say Claude is a harness, and it already has the agentic under the hood, so it’ll spin out sub-agents and things like that. So what I’m doing from an architecture standpoint is trying to find the leaders and the use cases and then build up a range of impact as they get more comfortable with it. Start with, what can you do with chat? What can you do with Cowork and MCPs integrations? And then how can you reshape that to be something you do consistently or offload with skills?
Generally speaking, most companies don’t need something complicated to maintain, like a custom agentic workflow. If you’re building that on the product in the backend, great. There’s a lot of benefits to that. Pearl Planner uses agentic on the backend, and there are 20-plus agents in there that handle everything, from updating budget, to updating colors, to gathering context from what’s on the screen at the time, to updating tasks, or researching different things from the web, or pulling content from our library for specific moments. We have our own orchestration layer with all those specialized agents within it.
Where do you see this going next?
We’re hoping someday it’ll branch out beyond weddings, and people can just use it to plan anything. I did not have wedding retail on my bingo card, but it’s a very unique strategy, unique opportunity, a lot of the dimensions that I’ve never worked in before, so I’ve had a lot of fun.
What does LogRocket do?
LogRocket’s Galileo AI watches user sessions for you and surfaces the technical and usability issues holding back your web and mobile apps. Understand where your users are struggling by trying it for free at LogRocket.com.