05 Sep Making AI accessible for everyone
Our mission here at ReSci is to make artificial intelligence accessible and usable for brands. We believe that everyone should have access to the predictive capabilities that have made companies like Amazon and Netflix into the powerhouses they are today.
So how did we seek to do this?
We had proven data science models that worked, and we ran constant validation tests to prove it. But it was difficult for marketers to do something with this information, never mind attempting to apply it to marketing campaigns in an automated way. The predictive power was there, but the accessibility was not.
So we set ourselves a few goals:
- Allow marketers to use these predictions without needing an engineer every step of the way
- Embed the predictions in campaigns such that a marketer’s day-to-day is not drastically changed
- Be true machine learning, with feedback loops so that the machines continue to improve over time
We didn’t do everything perfectly, but we did achieve all of these goals with our automation product, Cortex. It’s the only email platform on the market that is capable of making decisions on behalf of the marketer. And it’s built in such a way that a marketer can (mostly) set up campaigns in familiar ways, except a bunch of the little knobs and levers just—didn’t need to be there anymore.
Let’s break down how we applied predictions to automation to make this work.
Reimagining the welcome series
This was one of the first things we sought to innovate, and for good reason. Instead of marketers spending time trying to map out journeys, most of which were just conjecture and pushed users into restrictive flows, we wanted the machines to do the heavy lifting. We applied our predictive lead scoring models to emails that would trigger shortly after a user signed up with a business.
The emails would send on the days and times that were best suited for any individual user, as opposed to everyone getting the same message on day 3, or 4, or 5 in a traditional welcome series. So now, instead of intensively mapping users in a workflow tool, you can just put some emails in these stages, and all your new signups would be nurtured along, but more intelligently, and at scale.
Reimagining the winback series
Similarly, we wanted to replace the workflow heavy winback drip, and we wanted to use our already working models to do it. Sure, a user who hasn’t bought anything in the past 60 days is probably churned, but are you sure? And could you have caught them sooner? Our models could, and can also trigger automated At Risk and Churn messages all by itself. And the messages would try a few times, based on what our models predict a user’s optimal frequency is. Again, just add some emails and go get a coffee. Your job is done.
Reimagining item triggers
Everyone does triggers these days. It’s pretty easy, and also very commoditized. You just need a user to do X on your site and you show them Y. Except you have to build everything that determines the Y. With Cortex, things are different.
Each item trigger ONLY sends to users who have some interest in those items. And we used a proprietary model we called “item follower” to power this. We were pretty happy with this new innovation, as it sent the right triggers to the right users, while Cortex would find more relevant stages to send to everyone else. Our Buy Again (replenishment) stage, finds your items that are statistically frequently repurchased, and predicts the right time to notify the user. Just add some emails, and let the machines do the rest.
Product recommendations were our bread and butter from the beginning. Most importantly, providing the power of Amazon style product recommendations to anyone who plugged their data into our customer data platform, brought us closest to our mission. And it’s so easy to deploy. Just add some dynamic tags to emails, and the machines will just return recommendations for each individual user. And we give marketers control:
We’re so confident in our recommendations, you can take them with you via our API.
Reimagining A/B testing
Splitting your traffic 10% + 10% with variants and sending the winner to the rest is just—a delayed batch and blast, and is wasteful of precious touch points to your customers. If you can expect a .5% unsubscribe rate with each campaign, and you’re sending to 100k people, you’re losing 500 people EACH TIME you send. We came up with multi-armed bandits for marketing content, so you can just add subject lines and emails, and the machines will keep finding the right content for each customer, and get that content to them faster. More relevant content will be engaged with more, unsubscribed (and marked as spam) much less, and the revenue will come.
We know blasts will always be necessary. So we made sure we built a blast that plays nicely with all the other AI automated stages. This type of decision making will keep your revenue up without hurting your deliverability and sender reputation.
This is really only a portion of what Cortex is doing under the hood of your marketing campaigns. There is so much more capability, and so much more coming in the R&D pipeline. We think marketers can turn everything on with much less work, and without needing a technical resource to help them. AI is embedded in every nook and cranny of Cortex. And the results have been piling in.
We continue to do our best to make marketers feel comfortable using the AI, and our redesign, which you can see in the screenshots above, will be taking a big step in the right direction this week.
We will continue to keep our eye on our north star: our mission to make AI accessible for all.
About the author
Derek Kwan is COO and VP of Product for Retention Science. With 15+ years experience in marketing and ad tech, Derek previously led product innovation at Yahoo!. He also trains his game theory skills in his spare time as a poker player. Follow Derek on Linkedin.
Not a customer yet? Let us give you a demo.
Check out our free AI guides: