Someone glancing at their watch mid-conversation is a small tell, easy to miss, easy to talk past. Someone recoiling when you say the wrong thing is not. A good conversationalist reads the difference and adjusts accordingly. Most customer engagement software never gets that chance, because it's built to watch for exactly one thing, a purchase, a signup, a completed checkout, and treat everything before it as scaffolding, or worse, exhaust.
That's a strange way to run a system that's supposed to be learning about a person, and it gets stranger once you notice that the moment being watched for is usually the worst one to learn from. It's rare. For a lot of the outcomes that matter most, retention, a subscription renewal, it doesn't even exist as an event. Nobody fires an event called retained a customer.
Aampe starts from a different premise, and it's a sharper premise here than it would be for a segment-level tool. Each user gets their own learning agent, one that has to build an individual read on that one person rather than lean on a stereotype of people like them. A segment tool can get by on a handful of funnel events, enough to sort someone into a bucket and move on. A per-user agent needs the whole stream, every glance, not just the moments someone finally says yes. Once that much is coming in, the work isn't deciding what to keep. It's figuring out which of those events are lessons and which are just facts, two different categories that get treated as one far too often.
Some things the agent shouldn't have to guess
Start with the easy case. Whether someone added something to their cart and didn't buy isn't a mystery the agent needs to solve through trial and error. The business already has that fact. Handing it over as a guardrail, this person is eligible for this message, frees the agent's judgment for questions nobody's already answered. Frequency rules work the same way: don't ask someone to refer a friend twice in a row, don't push a feature at someone who used it five minutes ago. None of that is learning, just information the agent should have had on day one.
Reading the room
Take those facts out of the stream, and what's left is a run of small reactions, the app's version of a glance or a flinch. A purchase counts as the full win. Adding to cart is most of one. Searching and filtering is a small fraction. Landing on the homepage is a sliver of a percent, and it still isn't nothing, because the person didn't ignore you. The one real exclusion is the event that isn't a reaction at all, the ping that fires because a screen loaded, with no person behind it. Everything a person actually did, however small, earns a grade instead of getting thrown out.
When the goal isn't an event at all
The thing a business wants more of, purchase, renewal, whatever counts as success, tends to be rare, and for outcomes like retention there's frequently no event to point at in the first place. Skip waiting for it. Look at who reached that outcome over some stretch of time, compare them to who didn't, and pull out the repeatable behaviours, the everyday reactions from the section above, that separated the two groups. Point the agent at those instead. The business still gets to report on the real outcome directly. The agent just learns from a proxy that actually exists as something to act on.
An agentic system doesn't need the KPI and the learning target to be the same thing, a real break from how most legacy systems are built. An insight that used to die in a slide deck after some VP asked for it now has somewhere real to go, because the agent isn't in that meeting. Telling it what came out of the meeting, the way you'd train a new hire instead of expecting them to absorb the business by osmosis, is the only way it gets included.
Events that are real and chosen but still bad goals live in this same bucket. A direct deposit lands on a schedule no message will move. A fantasy team gets created once and the impulse runs out. Keep both in the stream. Neither should ever be the thing the agent is aiming at.
When someone goes quiet
Location, time zone, age, mostly sit outside all of this, because an attribute can predict but it can't explain. Knowing a time zone correlates with response to a discount doesn't tell you whether to hand that discount to everyone in it. Attributes are most useful before individual signal exists, a way to borrow a pattern from similar users, and they get phased out as real behavioural history accumulates.
Silence gets handled the same way, as a gap to fill rather than a verdict. An agent that sends five things and hears nothing back keeps going, working from the assumption that it hasn't found the right thing yet rather than that the person is gone. Unsubscribing or uninstalling is different, scored as a real loss rather than a weak positive, because it closes a door the others don't.
The point was never the ride
A ride share app that only ever pushes take a ride is optimising for the fraction of users who need one today, burning its shot at learning anything from everyone else, what tone they respond to, what would actually get them to open the app. Two weeks later, when that quiet majority is finally in the market, a system that spent the downtime learning already knows where to start. A system that spent its time pushing the same line is starting from zero, again. That glance at the watch turned out to be the whole education.
This builds on a conversation between Aampe co-founder Schaun Wheeler and Patricia Lazatin on how decisioning agents learn: what job each event should do, and why. Watch the full conversation.
