Why removing items from a wishlist is a strong buying signal: AI agents and app retention

Should you have AI agents assigned to every single app user? Imagine the results if you could do exactly that …

Imagine smart AI agents able to suggest new options, notice when a user or player seems confused, help when there’s a problem, suggest the next best thing, offer the right coupons or discounts or packages for each and every person, all in real-time. It’d be almost like having a real person there with people in your app, offering suggestions and, in some cases, trying to close a sale.

Is this just a dream? Or is there some substance to the promise?

I chatted with Shaun Wheeler, a data scientist at Aampe. Hit play, keep scrolling …

Maximizing engagement via AI agents

Everyone wants to maximize engagement, retention, and monetization, right? That’s why tools like Braze, CleverTap, and Aampe, among others, are so popular. Old-school segmentation and audiences are blunt instruments, crude, slow, and unable to react in realtime to user activity.

“Every user on an app is an individual,” says Wheeler. 

“They have a different background, different preferences, different patterns, different things are going to motivate them or activate them, and there isn’t any data science team big enough that they could actually tease out all those individual patterns. Even if there was, there’s no CRM team big enough that they could actually act on all those patterns.”

So a rules-based approach triggering specific messages when players or customers do specific things, he says, is substandard.

Even worse, there’s a lack of learning.

If User A does Action B, what does that mean for the future? Does it mean they’re actually really primed for Offer C? And if so, what’s the next logical projection? A rules-and-trigger system doesn’t really know anything or store anything: it just reacts. All the “knowledge” resides in human heads that are setting up the rules and triggers and messages.

And that just can’t comprehend all the myriad of paths users/customers/players might take, or how previous behavior will impact current action. Plus, it can’t react in realtime and build new models of what people do as things change.

AI agents can do better.

AI agents, not quite like Agent Smith

AI agents sounded a lot more sci-fi before ChatGPT, Wheeler says. Ultimately, they’re pretty simple.

“An agent is just a type of AI that can receive guardrails and then it can act autonomously within those guardrails.”

The ideal dataset isn’t just 5 things you’ve instrumented in your app, or even 50. It can be everything: the entire datastream of what people do in an app.

“We work with apps that easily have over 400 different types of instrumented events,” he says. “That’s all information, like every button click, every page visited, whether it’s on a site or an app, and you can take all of that into the agent and process it in a way that the agent can then make decisions about.”

The goal is to be able to understand how likely each step — and each stimulus the app provides in response — is likely to move a player or user or customer towards a certain goal. How the app responds and what channel it uses, whether an in-app notification, a push notification, an email, or even potentially some modification of how the app acts or what it looks like, is on the table. That’s exactly what AI agents are intended to do.

“All of those kinds of behavioral decisions are normally made by a CRM team,” Wheeler says. “Many of them can actually be made very reasonably by an AI agent that’s properly structured.”

Better engagement through smart messaging

One example?

I play a game multiple times daily that opens with 3 pop-up messages that are essentially the same, with minor variations, every single day. Some of them seem to be power-ups for parts of the game that I don’t play and don’t understand. Others are announcements and events in such tiny text I’d have to work hard to figure out what they’re about.

All of them tend to train me to close pop-ups instantly. 

They also put roadblocks between me and what I opened the game to do: have fun. 

I’ve often thought a metric that app developers and marketers should track is “taps to X.”

• For games that’s taps to fun: how many taps does it take before a player is enjoying a game?
• For retail that’s taps to buy: how many taps does it take to purchase something?
• For fintech that’s maybe taps to pay or taps to deposit …

Knowing that number — and any changes up or down — is more important than I think most app publishers admit to themselves. And whenever that number goes up, that’s a massive disincentive to use an app.

The point is: messages seem free because you can just pop them up on your app, or click to send them out via push notifications.

In reality, they’re extremely expensive. They can cost engagement. They can cost retention. They can cost monetization.

So ensuring that each message is a smart message that is both relevant and timely, and is sent to someone who will be happy to see it — or at least OK with seeing it — is critical to generating better app engagement. AI agents could notice that I’m not paying attention to these pop-up messages, and either stop sending them, or just sending 1 that is actually relevant to what I do in the game.

Saying less but achieving more

One example of doing exactly this is a nursing app that Wheeler mentioned that was using SMS to message nurses about shift availability. 

SMS is good because most of us see most of our text messages, as opposed to in-app messages that we only see when we’re in the app, or push notifications that may or may not be enabled, and even if enabled are not tremendously effective in all cases. But SMS can be expensive, so messaging this way is literally financially costly rather than just metaphorically costly as mentioned above. 

The result of smarter messaging getting better engagement?

“We increased their activation by, I think if I remember correctly, it was 9%, but we reduced their SMS volume by 75%,” says Wheeler. “That was savings in the hundreds of thousands per year.”

How? 

The AI agents recognized nurses who never took weekend shifts, or never took night shifts, or followed other patterns of behavior. (I personally know a nurse who only takes night shifts.) That personal knowledge — almost like a human who deeply knows the individual people using the app — made all the difference.

Sometimes AI is smarter than us

Ask a human about a potential customer removing an item from their wishlist, and you’d probably get an answer that it’s likely a bad sign: a sale won’t happen here.

That’s actually opposite to reality, and the AI agents discovered it by following the data.

Wheeler talks about what he learned from a retail app:

“There were several events that you wouldn’t be surprised led to a higher probability of a purchase. Adding to cart is a very strong signal, but one that really surprised us was the wishlist. It wasn’t adding to wishlist … it was removing from wishlist that made them more likely to purchase.”

What was happening is that people who are adding to their wishlists are just curating products, sort of like a Pinterest board. But removing from a wishlist often means that you’re actively making decisions about what you might be just about to buy, right now or in the near future.

Much more in the full podcast

Check out the full Growth Masterminds episode for much more.
You can subscribe to Growth Masterminds on any podcast platform you prefer, or watch the videos by subscribing to our YouTube channel.