Billions-scale personalization with AI agents: Our investment in Aampe

The most successful products in the internet age have all been built around personalization. Tiktok’s feed, Netflix’s recommendations, and Spotify’s discovery playlists are so powerful because they provide truly unique experiences for billions of different people.

For most businesses, the most common interaction they have with their customers is outreach – emails, texts, push notifications. But the vast majority of these lack any real personalization. Marketing and customer lifecycle teams blast out effectively the same message to thousands or millions of users, knowing that only a tiny fraction will respond.

This is why we are thrilled to lead the Series A for Aampe, which has built the first agentic infrastructure for personalization. For each customer, millions of agents explore different personalization strategies to determine the most effective one for each user. As new preferences and behaviors emerge, Aampe agents evolve to reflect them. This has a dramatic impact on the business metrics teams care about, allowing them to focus on personalization strategy and content creation instead of execution.?

Why you can’t personalize with rules-based automation

Last week, we wrote about jobs that AIs do better than humans. Personalization is a prime example of this, as AI can make millions of data-driven decisions in ways people simply can't.?

Marketing and customer lifecycle teams are experts at understanding customer needs and figuring out how to best connect with them. But even the largest and most skilled team in the world couldn’t possibly hand-write the millions of messages they send each day. This forces them to build rules-based journeys that decide what to send to whom. This is the welcome sequence we’ll share with every new user. Here are a set of reminders we’ll send to every person who leaves an item in their cart.

The problem with journeys is that they force you to design for the typical user. But of course, every user is different. Even if you correctly identify and speak to the most “regular” groups, you won’t connect with all your other users with different preferences.

Existing platforms integrate AI into the rules-based journey. An email on Black Friday sales might highlight a product category a customer frequently peruses. That is a step in the right direction, but only a small one. Real personalization is multivariate: Do they want to buy the same product they bought earlier, or something complementary? Do they care more about value, convenience, or quality? Do they prefer email or text? Messages that are short and to the point, funny, or full of pictures? A couple reminders, or to be left alone after one??

The action space is enormous and further complicated by several factors:

• Preferences are always changing, both on an individual and group level. A customer might be shopping for a relative around Christmas but for themself in the new year. They might just want simple discounts during the week, but a fun story on the weekend. Everyone wants to be bothered less on Thanksgiving than a regular Thursday.?
• Customer engagement happens over time. Today’s customer engagement platforms show what percent of users clicked on a given message. But teams typically cannot tell if a message they sent led to a downstream action like a post or purchase. They certainly can’t track if it led to a purchase days later. And they are blind to insights across campaigns: for example, that a given customer always engages with messages that highlight convenience, versus ones that highlight value.
• It’s hard to observe counterfactuals. Every message teams send to customers changes their subsequent behavior, and when they send the same thing for everyone, it’s hard to know counterfactuals. If a team schedules all their engagement between 8am and 8pm, how do they know there aren’t users who would have preferred earlier or later? Would sending a reminder followed by a discount have been better than a discount followed by a reminder? The only way to know is to run many experiments over time, which is impractical to manage.

It’s clear why even with the best tooling and team, it is impossible to do broad-scale personalization with rules-based automation.

How AI agents solve the problem

Aampe builds AI agents that make 1:1 personalization decisions for each individual user. They connect into existing data sources (e.g. customer, content, and inventory management systems) to understand a company’s products and customers. Next, marketing teams craft content. Instead of perfecting a single message to blast out to everyone, they can experiment with dozens of different offers, messages, and tones, creating tens of thousands of combinations that could resonate with different audiences.?

Finally, Aampe agents explore the actions they can take – choosing what to say, how to say it, when, and via what channel. They experiment over time via reinforcement learning, deciphering what strategies are most effective for a given user at a given point in time. While traditional customer engagement platforms typically only track click-through rates, Aampe tracks hundreds of product and business outcomes. For example, the agents might notice that a message generates a lot of clicks but no purchases. Another might not prompt an immediate response but substantially boost retention over time. Across large-scale consumer apps, Aaampe has demonstrated double-digit improvement in business metrics like retention and transaction volume compared to previous engagement strategies

Building these types of systems is extremely hard, and traditionally has been done only in the Amazons and Netflixes of the world. Aampe founders Paul Meinshausen, Schaun Wheeler, and Sami Abboud are a rare group, combining expertise in both large-scale data and quantitative social sciences, with experience building? similar infrastructure across industries – from defense to consumer applications.

A new class of infrastructure

Deciding how to engage with customers via email, text, and SMS is an easy place for AI agents to add value over rules-based journeys. But Aampe’s agentic infrastructure is much broader than that.

Already, Aampe is starting to personalize in-app experiences, another area where one-size-all approaches leave customers dissatisfied.

They are also the first personalization company to generate valuable new data on customers. Unlike existing customer engagement platforms, Aampe (1) knows what kind of content is in each message, and (2) runs continuous experiments matching messages to users. This experimentation not only improves personalization, but also creates a rich new dataset of user preferences and behaviors that is valuable for other parts of the organization — product, data science/ML, merchandising, etc.?

Aampe identifies which customers prefer certain products, respond to discounts, or resonate with particular value propositions. For example, they have discovered brand new user segments that were previously unknown to customer teams (e.g. late night snackers).?

Aampe is leading the charge on agentic infrastructure as a core technology underpinning the future of consumer applications. We are so excited to partner with them and lead the Series A, with participation from Z47 (Matrix Partners India).