Better Customer Outreach with AI

Author: Ivaylo Bahtchevanov

The Power of Notifications

If you’ve ever unsubscribed from a notification, you’re aware of what a poor notification experience is. Similarly, if you’ve ever clicked on a notification, you’ve likely found it to be useful and relevant. In this article, we’ll share what a good notifications strategy looks like, and how to use AI to improve your approach for every user.

Notifications via channels such as email, mobile / desktop push, and SMS, are among the most effective means for businesses to engage with their customers. From a user’s perspective, it ensures timely access to information when they are away from the app or the website. On the flip side, notifications can be a powerful strategy to drive engagement and even revenue for companies, as the growth of an SMS marketing platforms like Attentive proves. 

While it might seem intuitive that sending many notifications will lead to better engagement, the reality is that it can very easily backfire. Send too many notifications and you  risk them turning off notifications, or churning from your product or service altogether. Another big risk of sending too many notifications is getting blacklisted by Google, Outlook, Yahoo mail or Apple/Android. Once you’re dinged by the major providers, it’s difficult to claw your way back in cutting off a crucial communications channel with your users.

This means that ensuring the right timing, placement, channel of delivery, and curated content of your outreach is critical to your business.. When done well notifications should be a natural extension of your core product and enhance the overall customer experience.

For example, emails allow casual users to view recommendation content directly in their email clients without having to log onto the site or install the mobile app. Having a compelling yet accurate subject line for these emails will in turn maximize open rate, while minimizing unsubscribes. For active users, push notifications may work better for those who have the app already installed on their phones. 

AI can be leveraged to optimize notifications in a number of ways. There are three important aspects to consider for  an effective strategy using AI: 

• Volume Control Optimization 
• Channel Selection 
• Content Personalization 

Volume Control Optimization

The volume control optimization problem predicts, for each user, the correct frequency and timing for sending out notifications. The objective function needs to accurately assess the impact of long-term user engagement rather than just the short-term clicks and impressions. It will also predict, to some degree, whether a certain volume will lead to unsubscribing from notifications in the future. Achieving good performance requires complex, nonlinear models to capture this trade-off as well as efficient scalability to be able to learn from millions of users.

A good example of this is Meta’s notifications sytem. It demonstrated that when they decreased the notifications to just a few per user, the overall long-term engagement drastically increased. On the other hand, they found that increasing volume very quickly had a diminishing impact to the overall experience and resulted in a net negative experience. While in the short-term more notifications increased clicks and views, these notifications were very quickly perceived as spam and consequently users opted out. 

Similarly, the Linkedin system tunes a personalized volume capping per user. It balances this against conversations or opportunities a given member may care about, thus improving the overall user experience. These notifications play a critical role in how members engage with one another. Examples include connection invites, job recommendations, daily news digest, and messaging all of which are part of the core product experience.

Channel Affinity

Channel affinity determines, for each notification sent to a given user, the best channel to send the notification given the user’s likelihood to respond. Part of the problem is understanding each user’s unique preferences and ways of engaging with the platform. The other part is tying the preferences to the exact timing specified in the volume optimization problem. If a user has the mobile app, they might be more likely to respond to a mobile notification. Pinterest’s notification volume control system combines Volume Control Optimization with Channel Affinity to deliver highly effective and engaging notifications in a systematic way.

Content Personalization 

As we recently shared in another post on personalization, there are incredible benefits for ensuring all user content is as personalized as possible. This includes ensuring any notification messages are associated with content that has been determined is aligned with a user’s interests, whether it’s a piece of clothing in the case of a retailer, or a piece of content, in the case of a media company. Content relevance requires understanding a user’s previous interactions with the platform to develop the best recommendation. GNNs have been proven particularly effective in providing curated content for users. Another dimension of the content problem is to summarize the subject or heading: i.e. to identify the specific set of keywords that will strongly resonate with the user and convince them to open the notification.

Optimizing your Notifications like a Portfolio

Managing the overall volume, frequency, channels, and content of notifications can be thought of as a portfolio optimization problem. Given many different notification types and strategies, each company needs to determine what the best combination of notifications will be most effective for their particular set of customers. Arriving at the optimal portfolio of notifications means iteratively and agilely testing many different hypotheses and combinations and arriving at a balanced portfolio of notifications that will maximize both the long-term engagement and the overall user experience of your customers. 

The chart below shows one such experiment done by Pinterest. They found that the optimal volume of emails to increase the number of active days on the platform was as high as nine, higher than the default (control) of 7-8. 

The Kumo Approach

An effective organic outreach system outlined above requires solving easily over a dozen different machine learning problems, many of which will require dedicated pipelines. You need to build a nuanced understanding of your users, and then run many experiments on top of these models to optimize. All of this requires significant complexity and overhead to manage end-to-end.

Kumo’s approach is to make it very easy to quickly generate predictions that will help you understand future potential outcomes based on a set of actions. This is accomplished by building a graph from your enterprise data, learning the relations and interactions between all the entities in your tables, and providing a no-code interface for orchestrating many ML models concurrently with a single query.

By using the Kumo Predictive Query, you can perform a comprehensive what-if analysis that enables you to determine the optimal strategy from a set of potential actions and identifies the best path forward based on a desired outcome. You can flexibly test multiple hypotheses with a series of queries. 

For example, running the query below will score all notifications for a given user based on likelihood of positive engagement:

For each notification in the notification table:  
Predict the conversion in K days if the notification is sent out in D days / H hours

You can then run a query to predict the outcome given a specific notification. For example, you can run any of the following:

For each of the 3-month active users:
Predict the total sales from the user given that you send them a coupon next week.

For each user who has opened a session in the past 30 days:
Predict the number of times they will make a purchase given you send an email to them tomorrow.

Similarly, you could predict the number of deliveries ordered by users for an on-demand delivery service, the number of purchases made on an ecommerce platform, or the likelihood of a set of users churning or disengaging over a given future period. Each query translates to a fully optimized state-of-the-art Graph Neural Network, tuned to the specific data and problem.

The ability to launch dozens of concurrent queries means agilely experimenting with many hypotheses, and getting to the right approach in a matter of days rather than months, allowing you to capitalize on growth opportunities before they pass.

To learn more about how Kumo can build your optimal organic outreach strategy, reach out to us directly!