Most of the discussion around AI and marketing is about content generation via LLMs – but Hightouch is building AI products that fundamentally redefine what marketing’s role is. Their new AI Decisioning lets marketers automatically route the right content and campaigns based on signals from across tools and the data warehouse.
I sat down with Kashish Gupta, Hightouch’s cofounder and co-CEO, to talk about how they actually built this, what marketing teams want from AI, and why RL is making a comeback.
”Imagine there's one marketer for every consumer in the customer database. That's how the reinforcement learning agent operates. And it has access to every past action every customer has taken. So it learns on all customer history. From the email system, the SMS system, and the push notification system, it'll see all the copy and creative that is available, and determine based on all past actions, what is the best campaign to show this customer to optimize for an outcome."
“Reinforcement Learning is uniquely suited because it’s very good at understanding a sequence of events, both historical and future, as it relates to a goal state. Most marketers are trying to optimize for LTV – how much transaction volume am I getting from a particular customer – but LTV isn’t linear. There might be a bunch of intermediate actions you’d want a customer to do now, so that you set them up for a higher LTV later. RL does a really good job of optimizing not just getting to the goal right now, but getting to the goal eventually through intermediate states.”
“In our UI we let marketers define their own reward functions: they can decide that a click is worth some amount, a session is worth some amount, etc. And these end up being vastly different across customers. Companies within certain verticals to some degree resemble each other, but even then they’re still vastly different. For example, all of our grocery store customers right now are optimizing for order online or order ahead, because they know that if customers do that they’re likely to order from the same store every time. RL might take many cycles to figure this out, but marketers already know it.”
“We ended up building an entirely new engineering team for this product. We hired some ML engineers, ML scientists, and then restaffed a bunch of our distributed systems engineers onto the project as well. Hiring great talent is always hard, and we did a couple of things to stand out. First, we do something unusual, which is a single interview process: it’s usually 1.5 to 2 hours, and if you pass it you get the job. We also have an incredible VP of Engineering who used to run the ML team at Opendoor; people join Hightouch because they’re surrounded by the best engineers.”
.svg)