Table of Contents
At The Founder, we are thrilled to bring you an exclusive interview with Paul Meinshausen, the visionary CEO and co-founder of Aampe, a trailblazing company reshaping the personalization landscape for consumer apps. With an impressive background that includes co-founding PaySense (acquired for $185M) and a founding team of scientists specializing in cognitive and behavioral science, engineering, and experimentation, Paul and his team are leading the charge in deploying agentic AI infrastructure. In this conversation, Paul shares his journey, the innovative technology behind Aampe, and their mission to redefine user interaction across industries.
Q. Your journey from co-founding PaySense to leading Aampe is impressive. What inspired you to venture into the field of agentic AI infrastructure?
PaySense was a waypoint in a longer journey of trying to understand and improve how we use technology to communicate with each other in a complex world. My career actually started in the harder realm of human conflict: simulating battlefield environments for the U.S. and then later helping the U.S. military use complex unstructured data to make inferences about insurgent and local population decisions in Afghanistan during the way there in the late 2000s. Those experiences taught me that we simply didn’t know enough about how human cognition works, and that drove me back into academia to do research in psychology and implicit social cognition. Academia showed me that we didn’t have the technology or tools to really systematically model or learn much about human communication and behavior, so I left academia and dove into data science for consumer technology.
Building a mobile credit and fintech app in India helped me recognize some key limitations in how we built consumer software (e.g. mobile apps) in the first real decade of mobile app tech (2011-2021). So instead of building another consumer business, in 2020 I decided to solve these key problems for consumer businesses by developing what we came to call Agentic AI infrastructure. That’s what Aampe represents and we have a lot more work ahead of us.
Q. How has your diverse background in cognitive science and technology shaped your approach to building Aampe?
For anyone who has studied the human mind, what generative AI does feels familiar—it’s procedural memory. Procedural memory helps you ride a bike, shuffle cards, or form a sentence. It’s important, but it’s only one of at least four categories of long-term memory that drive human thought and behavior.
We knew the capability we wanted to develop required associative memory, which lets humans distill experiences into simple representations that carry a definite emotional load. Associative memory is what lets you recognize value (what helps you) so you can seek more of it, or what harms you so you can avoid it. Technologically, that’s reinforcement learning — not generative modeling. As we’ve expanded into other memory systems, like habituation and priming, we’ve repeatedly seen that true productive autonomy doesn’t come from procedural memory alone. This aligns with established cognitive science research.
Language impresses us because it feels uniquely human, but it’s not sufficient for human success. As a species, we’re tinkerers — we try, reason, and learn. That should be the essence of AI. Our agents perform well because we designed them to mimic human cognitive capabilities, not just surface behaviors performed by LLMs.
Q. Aampe’s AI infrastructure deploys intelligent agents to personalize user experiences. Can you elaborate on how these agents operate and what sets them apart from traditional AI solutions?
Aampe’s infrastructure allows our customers to assign a dedicated agent to every individual user of an app. Those agents autonomously learn by interacting with their assigned users. That interaction can happen on any surface - conventional marketing channels like push notifications and emails, but also on literally any screen on the app or page on the website. Any place you put content to try to satisfy or delight or engage a user, an agent can personalize that with information it's accrued over time by observing and curating that user's past experiences.
The main difference between our agents and naive implementations of generative AI is that our agents have a persistent memory and can act without being prompted. The agents are continually learning and then distilling those learnings into simplified representations in long-term memory. So a person can do stuff on an app, then go away for three months, and the agent won't have forgotten you. They'll assume that you like the stuff you liked before, though they'll temper that expectation - essentially, they'll assign low-confidence to what they think they know about you, because you've been away for a while. But as soon as you interact with the app again, the agent will see that and will start building up high confidence judgements again.
Each agent is plugged into the app's complete event stream, reading button clicks and page views the same way a human would read body language or tone of voice. And they have access to a catalog of approved content - copy, images, links - that they draw from based on what they think the user will most want to experience. And when an agent doesn't know what a user wants to experience, it'll talk to agents who are assigned to similar users and get ideas from them.
Once an agent gets activated, it doesn't need a human to feed it a prompt. It has everything it needs from day 1: information feeds, goals, and concrete actions it can take. You can feed it guidance, and then let it run. It will keep doing its job. Just like a productive human employee would.
That's the biggest difference between our agents and LLM-based AI solutions. An agent based on an LLM needs to constantly be told what to do. A human who has to be constantly told what to do is the human no one wants on their team. Our agents are capable of figuring out for themselves what they should do, and when they should do it, and when they should stop doing it, and when they should do something different.
Q. With over 100 million agents deployed, what has been the most significant challenge in scaling Aampe’s infrastructure across four continents?
The challenge of scale for agents in some ways is similar to the challenge for humans. One of the biggest challenges to scaling was getting the agents to efficiently pick content in its final form. Most generativeAI chatbot applications today are B2C - they’re used directly by the end-user. The end-user interprets the agent’s output as responsive to their prompt and then directly modifies and gives it feedback, ensuring a kind of closed-loop value judgement. Enterprise AI involves another layer - the agent is interacting with an end-user on behalf of a business. This means it has to be a lot more efficient and accurate in discovering the relevance of the information it provides to an end-user. There are a lot of steps to that - pick a good value proposition, decide whether to do a personalized recommendation, schedule a message for Wednesday night, etc. But once an agent has made all of those discrete decisions, the agent needs to select pre-approved content that fits all of those criteria. You can brute-force your way through that kind of combinatorial problem, but it takes a lot of time and resources. So we designed our agents to rationalize their choices. Can they filter down the candidate set of content by first paying attention to the one aspect of the message that is most likely to matter to their user? For one user, that could be the call to action; for another, it’s the tone of voice; for another it’s the time of day. Deciding on that one most important lever and then doing the combinatorial stuff just on what’s left over makes the whole process more efficient. There are other examples, but they’re all similar: scaling the agents isn’t hard, but scaling the agent’s ability to deal with complex decisions took a lot of work.
Q. You’ve worked with some of the leading brands in food delivery, fitness, fintech, and entertainment. Could you share an example of how Aampe has transformed user engagement for a particular industry?
One of the most important implications and results of using AI agents to experimentally learn and curate an end-user’s experiences is that those agents are not reliant on statically generated and already existing data. Because the agents themselves are responsible for actively learning, they are able to generate the data they need. This overcomes a classic and massively important problem with the past 10 years of “big data”: garbage in, garbage out.
Take food delivery as an example: there are 2 categories of data that a conventional food delivery app operates on: 1) Restaurant and dish inventory data, and 2) User data.
Restaurant and dish data represents everything a restaurant provides about its business: address, description, all the dishes on the menu, and specific metadata about each dish (name, price, star rating, or various characteristics like ‘vegetarian’ or ‘chicken’).
User data represents everything an app knows about a user’s past search, browsing, and ordering history. The data shows that a user looked at these N restaurants, and ordered these D dishes, from these K restaurants.
What’s missing? All the additional ‘meaning’ level information about the end-users goals and preferences. Are they eating to achieve a health objective (protein-rich food to gain weight, or calorie-low food to lose weight), or an emotional objective (a dish or cuisine that contains some meaningful personal memory, or a dish that will serve a romantic interest). The user-interface for a conventional food delivery mobile application doesn’t have the capability to collect that kind of data. But an agent can explicitly and systematically insert information on any level into a piece of content and a surface in which they’re interacting with a user, and based on the user’s response, they can learn appropriately. This means consumer tech is no longer bound to and limited by the static data schema that was designed at its formation. Instead, the schema is virtually limitless and can evolve and support the agent’s service to the end-user.
Q. How does reinforcement learning enable Aampe’s agents to adapt and evolve in real-time, and why is this approach superior to conventional personalization methods?
Think about how you learn your favorite ice cream flavor. At first, you might try vanilla. It's good! What's not to like about vanilla? Then you try chocolate, find it amazing - better than vanilla. Then you try strawberry - it's good, but not chocolate-good. Over time, you add more and more flavors to your list of ice cream, and sort of mentally keep track of how much and in what situations you like each flavor. That preference profile started off pretty much flat - when you've never had ice cream, you both like and dislike every flavor just about equally. But as you get more experience, those "meh" opinions turn into "yes please" and "no thank you" opinions.
That’s reinforcement learning (RL): trial, feedback, adjustment. Every time you try something and get feedback, you change the mental "score" you have for that thing - you move it up if the feedback was good, and down if the feedback was bad.
Aampe's agents’ RL capability works the same way. The agents start clueless...well, that's not really true any more...even for brand-new users for whom we have no data at all, agents have a reference for what's been popular among other users so it can make some educated guesses. So let's say the agents start *relatively* clueless. Now they have a chance to message a user on a Thursday afternoon. Should they do it? They have no reason not to. So they message on that Thursday afternoon and the user visits the app! Great. That agent has a mental slot for "Day of week: Thursday" and another slot for "Time of day: afternoon", and as soon as that user starts doing stuff on the app, the scores for those two slots start to go up. And the message was about new shoes that they just got in stock. So the "Product category: shoes" and "Value proposition: new arrivals" slots also got a score boost. And the message was a push notification, so the "Channel: push" slot gets a boost too.
If the user responds poorly - the agent sees their activity drop off after sending them something - the scores go down. If the user doesn't respond at all, the scores start to shrink toward the middle, not "like" or "dislike". As a technical implementation detail: our agents aren't actually storing one score for each of these slots. They're storing two. One codifies how much the agent thinks the user likes or dislikes that thing. The second codifies how confident the agent is in that judgment. So a preference based on only one interaction gets a lot less confidence that a preference based on 20 interactions.
Those two scores are used to parameterize a bunch of beta distributions and the agents then use Thompson sampling to make decisions. So every value proposition has its own statistical distribution. When it comes time to serve up some content, the agent draws a number from each of those distributions, and the slot with the highest draw is what's sent. This is how the agents balance trying new things (exploration) with sticking to what they know works (exploitation). Just like you might randomly try pistachio even though you're normally very loyal to chocolate.
This kind of online reinforcement learning works really well in messy, unpredictable environments. The agent doesn’t need to know the “right” answer upfront. In fact, it never needs to know the right answer. It just needs to know where to place the next bet, so it can then adjust its understanding of the odds. Over time, it gets more opinionated and more confident, but if the user suddenly changes their behavior - they learn they like a new brand, or they change jobs and don't have the same purchasing power they use to, or their significant other finally told them their style needs some work - the agent will pick up on those shifts and change its behavior accordingly.
That's the real strength of online RL: it's not about prediction; it's about responsiveness. Humans are inherently unpredictable. In most situations, we simply don't have the data to predict really well. But that's ok, because humans themselves - the thing we're trying to get AI to emulate - aren't very good at prediction either. We're good at tinkering - at trial and error. We're constantly looking for clues that maybe we could be doing something a little better, or making a choice that works out better for us. That's all the RL agents are doing. They're changing as fast as the users are changing.
Q. How does Aampe balance the need for personalized user experiences with the imperative to protect sensitive user data, especially in light of increasing data privacy regulations?
Aampe exclusively uses first-party user data that is directly shared by each enterprise customer. Aampe agents are trained on the entirety of each customer's data, without sharing data across customers, ensuring isolation and robust security. If a customer chooses to share personal data, Aampe uses strict data privacy principles, such as data minimization, purpose limitation, strong encryption, and data subject rights, to ensure the security and privacy of this information.
Q. What role do partnerships with investors like Theory Ventures and Z47 play in advancing Aampe’s mission?
Aampe is very much a mission-driven company and this means we have to be selective about everyone involved. Those who join us as teammates understand the future they’re building toward and actively help discover the best way to engineer or market or sell toward that future. We select customers not just based on their willingness and ability to pay for our service, but also based on their ambition and determination to evolve toward a future that’s systemically better for their business and their end-users. And we selected investors who care deeply about our mission and the future it represents and embodies.
In addition to providing capital, what they do is provide top-cover on that capital. When you’re involved in truly deep innovation, your metrics don’t just come out of a preset box from some business school ready for use. You have to figure out how to measure progress because progress is unfolding on so many levels: statistical, technological, sociological and cultural, and operational and financial. The best venture investors are able to patiently observe and thoughtfully contribute to your evolving success metrics.
Our success is also, like anything else in today’s world, contingent on earning attention. Before enterprises can adopt an innovative technology they have to pay attention to the specific problems the technology reveals and solutions it offers. Those enterprises’ leaders are operating in a world of fast-moving and noisy information. So our investors also have to be capable of working right alongside us to earn and allocate the market’s attention towards the problems and opportunities that are central to our mission (as opposed to some other set of problems and opportunities some other competing technology might target and address).
Q. As Aampe expands its capabilities, what are your key priorities for scaling the team and infrastructure by 2025?
We are expanding our real-time capabilities and working to enable our agentic infrastructure to seamlessly inter-operate with a large landscape of existing marketing and product and data tools and software. To be successful, agents can’t wait until all SaaS is redesigned AI-first; instead agents need to be able to operate a significant chunk of the SaaS tools that today’s human marketing and product teams rely on. We'll continue evolving our cutting-edge AI infrastructure while ensuring robust and secure functioning to empower our customers to deliver personalized experiences to their users.
Q. Looking ahead, how do you envision Aampe transforming the way businesses approach user interaction and personalization in the next decade?
There’s no reason human Product teams need to design static User Interfaces for their mobile or web applications. Instead they will design flexible sets of interaction patterns, and then let agents adjust and curate specific experiences for end-users within those patterns and human-determined guidelines. Similarly, Marketing teams shouldn’t be manually designing and implementing static Campaigns or workflows or canvases. Instead they should be giving their agents brand guidelines, business objectives, and creative inspiration and letting the agents work from there. The result will be a world better prepared to deal with the 2 parallel forces transforming human society: material abundance (leading to attention scarcity), and human diversity (destroying the kind of uniformity represented by traditional ‘customer segments’).
