Agentic infrastructure for personalized experiences - one learning agent per user, running quietly behind the apps you already open every day.
Above: the Aampe mark - little message-shaped tiles and a lone spark, which is roughly what a billion personalization decisions look like if you squint.
Right now, somewhere, a food-delivery app in Jakarta is deciding whether to nudge you about dinner. It is not a marketer making that call. It is not a scheduled campaign. It is an agent - one of more than 100 million Aampe has set loose - that has been watching how you behave and has concluded, correctly, that you are not hungry yet.
Aampe is a San Francisco company that builds agentic infrastructure for personalization. The pitch is unfashionably specific: assign one small, dedicated AI agent to every single user of an app, let it run continuous experiments, and let it learn what to send, when to send it, and - the part everyone forgets - whether to send anything at all.
It works across four continents, inside food delivery and fitness and fintech and entertainment apps, making an estimated 150 billion decisions a week. Most people who benefit from it have never heard the name. That is rather the point.
"AI agents can make decisions at a scale that is impossible for any human."
- Andy Triedman, Theory VenturesFor two decades, the industry sold a word it could not deliver. Personalization, in practice, meant putting your first name at the top of an email that 4 million other people also received. Marketers built "segments" - which is a polite term for lumping strangers into a bucket and hoping the average held.
The tools were rule-based. Someone had to imagine every journey in advance: if user does X, send Y. The map was drawn by a human, once, and then reality wandered off it immediately. Preferences change. Context changes. The rules did not.
Aampe's founders, all data scientists, found this maddening. A real person does not live in a segment. A real person wants a different thing on Tuesday than on Saturday, and a different thing again when it rains. No team of humans can write rules at that resolution. So the industry simply stopped trying and called the averages "personalization."
"The future of engagement in owned media lies in AI systems that learn from each customer's behavior and adapt automatically."
- Alexander Beresford, CGO/CMO, TaxfixPaul Meinshausen and Schaun Wheeler met in 2009 in a US Army intelligence analysis unit - the kind of place where you learn to make decisions from messy, incomplete signals. Meinshausen later co-founded PaySense, an Indian digital-lending company that PayU acquired for a reported $185 million. Wheeler spent years as a principal data scientist building consumer-graph systems. The third founder, Sami Abboud, is a former semiconductor engineer with a neuroscience PhD, whom they met through a telecom-payments API business.
Their bet, placed in 2020, was contrarian then and stayed contrarian through the generative-AI gold rush: don't use generative AI for the core engine. Use reinforcement learning. Don't write content with a model; figure out which content actually works, for whom, right now, by trying things and measuring what happens.
It is a more boring story than "the AI wrote your email." It is also a more honest one.
"Our mission is to fundamentally improve the way users experience digital products."
- Paul Meinshausen, Co-founder & CEOCo-founded PaySense (acquired by PayU). Background spans anthropology, quantitative methods, and data inference in complex environments.
Former principal data scientist who led consumer-graph work. Met Meinshausen in an Army intelligence unit in 2009.
Former semiconductor engineer and neuroscience PhD. Came from a telecom-payments API business where he was chief product officer.
Here is the mechanism, stripped of jargon. When Aampe plugs into an app, it does not build a campaign. It builds agents - one for each user. Each agent runs its own little experiment: it tries variations of message, timing, and content, watches the response, and updates. There is no master journey map. The map draws itself, per person, and keeps redrawing.
It connects to the plumbing companies already have - CDPs, messaging providers, data warehouses like Snowflake, content systems - and reportedly goes live in under eight hours. It also offers causal analytics with explainability, so teams can see why an agent did what it did rather than shrugging at a black box.
A dedicated reinforcement-learning agent for every user, optimizing what, when, and whether to send.
Thousands of message and content variants tested in parallel - automated A/B testing without the manual setup.
Explainable, causal read-outs so teams understand the why, not just the what.
Plugs into existing CDPs, CPaaS, warehouses and CMSs - live in under 8 hours.
Meinshausen, Wheeler, and Abboud start Aampe with a bet on reinforcement learning over rules and generative content.
An early ~$1.8M check from Sequoia Capital India's Surge program gets the infrastructure off the ground.
Additional funding to scale AI-driven customer engagement across more consumer apps.
Led by Theory Ventures with Z47, alongside the milestone of 100M+ deployed agents across four continents.
Plans to double the team and expand from marketing/messaging into managing the full user experience.
Skepticism is the correct posture for any "AI personalization" claim, so here is what is on the record. Aampe reports customer results in the territory of a 128% improvement in engagement and a 25% lift in incremental purchases - "incremental" being the word that separates real causal gains from taking credit for sales that would have happened anyway.
The scale figures are the more striking part. More than 100 million agents in the field. Hundreds of apps. Roughly 150 billion decisions a week. Named customers include Taxfix, the German tax-filing app, with deployments spanning food delivery in South and Southeast Asia, fitness in Europe, and fintech and entertainment in the US.
Bars scaled for readability across mixed units (% lift and hours). Treat as the company's own reported results, not independent benchmarks.
Aampe started where personalization is easiest to measure: messaging. But the stated mission is bigger and a little audacious - to let every aspect of an application adapt to each user's context, continuously. Not just which push notification, but which layout, which feature surfaced, which moment of silence.
There is a quiet ethical posture baked in, too. The system is built to retain no personally identifiable information, working from anonymized behavioral patterns. In an industry that mostly treats privacy as a compliance tax, designing the engine to learn without hoarding identities is a deliberate choice - and an unusually restrained one.
"We've designed infrastructure that enables every aspect of an application to adapt to each user's context and preferences, continuously."
- Paul Meinshausen, Co-founder & CEOThe bet on agents looks less contrarian every quarter. As "agentic AI" becomes the phrase everyone repeats, Aampe has the slightly smug advantage of having built the boring version - the one that runs, learns, and ships measurable lifts - while the term was still buzzword-shaped. Theory Ventures and Z47 put $18 million behind that head start.
Go back to the food-delivery app in Jakarta. The old way, you'd have gotten the 6 p.m. blast with everyone else - your name on top, a coupon you didn't want, a buzz you'd learn to ignore. The new way, an agent that has watched you for weeks decides the most useful thing it can do tonight is nothing. It waits. At 8:40, when you usually get hungry, it sends one message you actually open.
That is the change Aampe is making, one user and one quiet agent at a time. Not louder marketing. Less of it, aimed better. If they're right about where this goes, the best personalization will be the kind you never notice - because it simply stopped getting you wrong.