Fleckenstein is co-founder, chief product officer and CMO of machine learning start-up Amplero, a Seattle-based marketing AI company that promises to “leverage machine learning and adaptive experimentation to help marketers achieve what's not humanly possible.” Despite that slightly OTT value prop, I was encouraged to dig deeper by a shared pedigree: Fleckenstein was VP marketing & product management for mSpoke, a machine learning start-up I worked with seven or eight years ago until it was acquired by LinkedIn.
The opportunity to become re-acquainted with Fleckenstein (who played senior roles for Microsoft Office and Salesforce.com’s Marketing Cloud between mSpoke and Amplero) was a bit of serendipitous good luck, because Amplero’s work is potentially very important for marketers. In a nutshell, Amplero uses AI algorithms to analyze and segment your customer data into thousands of “nano-cohorts,” and then again to learn and adapt to how each of those cohorts actually responds, over time — without the need for human decision-making (though that’s always an option).
Here’s an edited version of our conversation:
Amplero talks about use cases with millions of customers. How big a business do you have to be for Amplero to be useful?
Fleckenstein: I’m not sure there’s a magic number, but typically we’re looking to work with companies that have a million-plus customers. We’ve gone lower than that, down to half a million or so. But when you start getting too small, frankly, you can still probably get value out of our system — but the lift you would get versus using a rules-based approach or traditional segmentation, it’s just not the same delta.
We’re focused on B2C enterprises that are taking more of a long-term-value view of the customer, not necessarily short-term transactional revenue. Think telco companies looking to increase average revenue per user, or subscription software companies looking to increase retention rates.
So what is the big delta you’re promising to customers?
Our initial customers are seeing anywhere from a 1% to 3% revenue lift over their traditional marketing approach by using our machine learning approach. One to three percent doesn’t sound like a big number, but when you’re applying it to billion-dollar-plus businesses, the math gets big fast.
I’ve seen the phrase “no data scientist required” on your Web site. Come on! Are you serious?
A couple of factors make that claim credible. If you think about the kinds of customers we’re working with, whether telco or gaming, we’re basically taking data they already have available. Much of that is already fairly cleansed, structured data that they have in a data warehouse. They’re either running their existing marketing automation or segmentation programs off it, or they’re doing other analytics on it. In a lot of cases, we also get actual customer usage data, which can be fairly messy.
Second, we apply machine learning during data ingestion to get things into the format that we need, so we can use the data in very specific ways for running customer marketing programs. That simplifies the data wrangling side of things for customers.
The machine learning does two primary things: It takes the data and retroactively creates a historically longitudinal view of the customer. Second thing is to look across all these customers and say, what are the thousands or potentially tens of thousands of nano-segments we can cluster these customers into to build small little cohorts?
How does the personalization and offer optimization actually work?
It starts with the marketers, who end up having a lot of control without having to do a lot of math or technical analysis. They tell the machine what KPI to focus on: I want to optimize 45-day average revenue per user, or I want to focus on reducing monthly churn — whatever the KPI is you want to drive.
The marketer then selects which channels they want to put into play: Hey, I want to use SMS, in-app messaging, email messaging. And they select which creative, messages, and offers to throw into the mix. There are also “sanity boundaries” they can set — like, if we’ve turned someone down for a credit card, don’t offer them a mortgage.
Once they do that, the system goes into a test-and-learn mode. If you think about traditional A/B testing, it’s like that — but instead of doing it in a broad-stroke way, you’re literally testing thousands of different permutations across hundreds of nano-segments to learn what works and what doesn’t work. It says, "Let me go test these permutations of creative, message, image, offer, channel on these cohorts of users to see how they respond." Typically, after like 30 days, it says, "Let me start optimizing," and will play the known winners more.
It sounds like the system must get very hungry for an enormous number and diversity of creative executions. Is that a challenge for marketers?
We typically start with a small number. We work with one agency for a gaming company, and they will start with four or five sets of creative that have different imagery, different tone and different kind of messages. Those can be further split into their individual elements: hero image, subject line, body text. Our system has a great ability to test permutations of those. And then as it begins learning what’s working in a general sense, it feeds that learning back to the agency, and the agency will often do riffs off the types of creative we’ve seen working. The system will then throw those new executions into the mix.
Besides learning faster than possible for humans, it seems this platform automates some of marketers’ biggest challenges from big data. Do marketers see it that way — or do they fear for their jobs?
When we started down this path I was convinced that the thing that would resonate most with marketers was the quantitative lift in performance we could show. And certainly, that’s important. But the thing that has resonated the most is, there’s an incredible sense of relief. Many marketers feel like, over the last four or five years, they’ve had to become increasingly data-oriented. They’re spending more than half their time in some tool, writing a bunch of if-then targeting rules to try to get more granular in their targeting efforts.
So many of them feel like they’ve gotten away from the root of why they became marketers in the first place: to be creative, to think emotionally about the mindset of the customer, and how do I engage them in both an emotional and rational kind of way, and how do I think more strategically. So they all look at this and, to a "T," they say great, this frees me up to be creative, to be strategic, once again.
What’s the one thing you want readers to know most that I haven’t asked?
The concept and the notion and the terminology around machine learning and artificial intelligence marketing can be pretty daunting. But the reality is, it’s actually quite easy to get started. Every customer we have has come into it thinking their specific organization isn’t ready — that they don’t have enough data, or aren’t sophisticated enough.
Inevitably what ends up happening with almost every large enterprise we deal with is that they already have plenty of data, plenty of creative, plenty of offers. In other words, they’re already doing the stuff they need to do. They’re just not leveraging the right tools to fully optimize all their effort.