Let’s address the elephant in the room…the world doesn’t need another AI blog post right now. The skeptic in me looks at the explosion around generative AI (GenAI) and thinks the biggest impact GenAI has had to date is that it’s helped write and proliferate blog posts about generative AI. But, like any product leader I balance that skepticism with optimism and potential for what could be, not necessarily what is. While the skeptic usually beats out the optimist, with AI it’s the other way around. I’m excited, and I think we’ll look back on the period of the next few years as a major turning point in the applications and impact of AI. I’ll also balance that optimism with a follow up post on the challenges I see in realizing this potential.
It’s worth calling out that I’m not new to this discussion. Throughout my career, I’ve been AI-adjacent. During my PhD, I worked at Brown and MIT to develop systems and infrastructure for both supporting and leveraging a new class of machine learning (ML). I spent years managing data infrastructure at a quantitative hedge fund, Two Sigma, widely viewed as one of the pioneers of leveraging and operationalizing ML at scale. During my nine years at ActionIQ, we’ve always explored and integrated ML into our product, whether it was through probabilistic identity resolution, model-based audiencing (e.g., lookalikes) or models to inform orchestration (e.g., channel affinity). So after months of conversations at conferences, I thought I’d distill some of the most exciting (and practical) opportunities for AI with respect to customer data and customer data platforms (CDPs).
Before we dig into the specifics, let’s first evaluate what exactly happened in the last few months (as of July 2023). In the context of AI, you can’t separate data and models. The recent advances weren’t solely a breakthrough in the context of the model itself. Probabilistic natural language deep learning models have been around since the early 2000s, and specific implementations like ChatGPT and Bard have been around since at least 2017. Recent advancements have been made in how the models are trained at scale, with growing corpuses of data, and more specialized hardware (see: NVIDIA). The importance of data in ML is neatly summarized by Peter Norvig, long-time Director of Research at Google, who in a 2009 paper wrote, “Invariably, simple models and a lot of data trump more elaborate models based on less data.” That statement is now nearly 15 years old — a lifetime in the context of tech — and implies that there’s a choice. However, we increasingly find ourselves in an era where hugely complex deep learning models can be combined with massive data sets and supported by scalable cloud infrastructure and specialized computing. In that view, data, models and the cloud are really three parts of the same trend culminating in recent breakthroughs in AI.
Ok, now let’s get into it in the first of a four-part series we’re launching to explore GenAI challenges, opportunities and customer experience (CX) use cases.
Opportunities in GenAI and CDP: Riding the Hype
Opportunity 1: AI as a UI
When we first started ActionIQ, we were trying to make it easier for marketers to build audiences/segments for their marketing campaigns. The challenge was that the data was inaccessible to non-technical users. So marketers relied on more technical ops/analytics/data teams to “pull lists” for them, often through custom SQL or ad hoc data pipelines. The process took weeks if not months, and there was no way to iterate on that timescale. The way the marketers would communicate the requirements was usually a ticket describing the characteristics of the audience. For example, “active high value customers who have an interest in shoes.” The problem with that is two-fold. First, there’s a tremendous amount of ambiguity in that statement. What qualifies a customer as active, high value, or interested in shoes? Is it their past purchase behavior, most recent browsing behavior, self-selected preferences, or even a lookalike model based on other “similar” customers? Second, the marketer might not know exactly what they want upfront, or how many customers will qualify with this initial criteria. These descriptions were directional at best, and they would be able to iterate on that initial criteria from there. Unfortunately, if it takes weeks to pull an initial list, iteration becomes impossible.
So what we did was we built a drag-and-drop UI (along with the appropriate modeling and abstraction layers over the raw data) to provide non-technical marketers an interface to be able to build audiences themselves, completely self-service. This takes the iteration time down from weeks to minutes, and allows them to continually tweak an audience to find exactly their right customers for a campaign. The results were tremendous, and we regularly see customers with higher campaign ROI with a self-service model. However, this model requires the marketer to take a much more hands-on role with the data, and while it certainly isn’t like they’re writing SQL, even a drag-and-drop interface isn’t as “easy,” from their perspective, as the natural language descriptions they used when writing up tickets.
But what if an interface into data could be self-service and based on natural language? If you’ve used ChatGPT, you know part of its value is in the conversational interface. A natural language interface to customer data would have a lower barrier to entry for new users and, if done right, could significantly reduce new user training and onboarding time.
Opportunity 2: AI-Augmented Insights
The next opportunity is in what we’ll call AI-augmented insights. Brands don’t always know what to do with their customer data, and don’t always know what to ask. Traditional insights and analytics tend to be questions-based, meaning you need to know which questions to ask in order to get the most meaningful insights. Even if you do identify what is happening, it isn’t always clear why that is happening. For example, consider a retailer that notices that a new cohort of users is spending significantly more on their first purchase than other first-time shoppers. Assuming they even recognize that this is happening (the what), understanding the why is a lot more nuanced. Is it the category they’re shopping in? The region they’re buying from? The channels they’re engaging with (e.g., app vs. in-store)? These, of course, are just very oversimplified possible answers to the why. Identifying the what and understanding the why of customer behaviors is critical for any fast-moving business or market. But as the scale, complexity, and usage of customer data only grows, those questions become harder and harder to answer.
The role of the user in this context is to provide the right business metrics and goals to help the GenAI determine which “whats” are interesting. And when digging into the why, help remove any obvious features that “overfit” to the what. In this way, the user can refine what the generative AI is looking for and iterate on the why to get to truly valuable insights. Moreover, discovering these insights should not be the role of limited and expensive data science resources. Productized GenAI has the potential to bring this discovery process to a much broader group of users. This is very much in line with our vision of democratizing customer data to redefine CX.
Opportunity 3: AI-Driven Decisioning
The third opportunity takes AI and layers it into activating personalized customer experiences. This third category represents AI deployed in a more operational way, making real-time decisions on-the-fly. Now, I know that this has long been the holy grail of CX, a way to move beyond audiences and campaigns to true 1:1 personalization. Many companies have promised this nirvana, but they’ve all fallen short. That’s not to say that they didn’t do this in some capacity across some channels, but the scope in which this has been deployed has always been limited. Why is that and what’s different now? Well, we’re seeing a convergence of trends not just in GenAI, but in CDP and cloud infrastructure. CDPs are bringing together all the customer data and making it accessible and actionable. Advances in GenAI bring new models with many new applications — and advances in the cloud are bringing increased infrastructure and AI-as-a-service offerings to operationalize these advances at scale. All these pieces are critical, but we’re reaching a critical mass of maturity.
The role of users in this model is to set the goals, constraints, and provide the proper governance. We’ve seen what can happen when AI goes off the rails. And while it might just be a funny series of tweets when it’s done in a personal context, there’s real monetary consequences to companies that get it wrong. In addition, this decisioning needs to align with human strategies. For example, maybe a certain offer should be prioritized for a specific region because of a strategic decision for the company to invest there. Or for regulated industries, the set of data that can be used for decisioning must be tightly governed due to legal restrictions on what can be used for targeting. In practice, the ability to input and manage the constraints and governance is harder than decisioning in a vacuum, and next-generation AI-as-a-service offerings will need to factor in that visibility and governance into how they build their products.
Accelerating Our Vision with GenAI
Our vision has always been to democratize customer data (with governance) across the entire organization, and generative AI is providing new mechanisms to accelerate that — and we don’t take the dystopian view that AI is going to replace everything CX professionals do today. In fact, one consistent theme in all of the above opportunities is that humans are always in the loop. The concept is simple: humans and AI combine to perform tasks that neither would have been able to solve as effectively or efficiently independently. If you’ve used ChatGPT personally, you know how important the initial and follow up prompts can be to get to a desired result. This is human-in-the-loop.
So what’s the gap? Why are we not in this AI-powered promised land today? The short answer is that we won’t get there with off-the-shelf generic models. Models like ChatGPT are trained on open-web datasets and are designed for broad, but not deep, use cases. Companies need to build and train their own models, on their own data, and the platforms and expertise to do this will take time to mature. And while this might take longer than integrating ChatGPT into yet-another dumb chatbot demo, in the long term it will help brands truly differentiate. Just as there’s wide acknowledgment that customer data is a differentiator and brands need to “own” it, GenAI models will be the extension of that.
This is the first of a four part series we’re launching on AI with challenges, opportunities and how our team sees use cases evolving with audiencing and more. Stay tuned.