How OfferFit uses self-learning AI to stop customer churn

Running a business optimally is not only about finding and convincing customers to buy your product. It’s also about retaining customers who’ve already bought your product, so they will remain loyal to your business.

Stopping customer “churn” has long been a big business behind the scenes, but historically it’s been a slow, tedious process. Boston, Massachusetts-based marketing-tech startup OfferFit, which announced $14 million in series A funding today, has made halting customer churn its life’s ambition in part by replacing standard manual A/B testing with an AI implementation that’s faster to respond to customers, more accurate, and more efficient to use.

A/B testing (also known as split-testing) is the process of comparing two versions of a web page, email, or other marketing asset and measuring the difference in performance. Testers do this by giving one version to one group and the other version to a second group; they then can see how each variation performs.

Ensuring a good user experience

OfferFit’s AI platform gives enterprises the ability to deploy the same type of self-learning AI that marketers at Google and Facebook use today, the company said.

Large B2C companies want to ensure a good online user experience each time by making personalized decisions on how to interact with each customer. For example, enterprises ask: What products should be offered? What channels should be used? How frequently should the company reach out? The existing methods for doing this are mostly manual, involving designing, running, and analyzing numerous A/B tests. This makes it impossible for companies to effectively make personalized decisions at scale, OfferFit claims.

The AI-based platform gives enterprises the ability to move from whatever A/B testing they do to advanced self-learning models. OfferFit can be customized to use a client’s first-party data to make automated decisions on the best messages, channels, incentives, and communication timing for each of their customers. The software also remembers each customer’s preferences. For customers such as Brinks Home Security and Engie, OfferFit claims its implementation has increased campaign performance while saving marketers hundreds of hours per year.

A key use case is that of Brinks, a smart home security system provider which has about 1 million customers worldwide. The company has seen a 200%-plus increase in customer-retention performance in the last year, largely thanks to OfferFit, Brinks CEO Bill Niles told VentureBeat.

Accelerating A/B testing

“What used to take us months of A/B testing is now possible in a matter of days,” Niles said. “This means we are making quick, informed decisions on what our customers want. Our initial pilot of OfferFit far exceeded expectations, and based on that pilot’s success, we intend to expand from customer retention into cross-selling and upselling, prospecting, and lead nurturing.”

“The holy grail for us is to ‘quantumize’ our customer base so that we can design and offer a bespoke marketing and sales experience to each of our 1 million customers,” he added.

Brinks’ customer base had a high propensity to churn, Niles said. “We were running 18.5% churn two-and-a-half years ago, and that’s just not acceptable. Customers might disconnect for a variety of reasons. What we were doing previously was traditional A/B testing, running different offers, and having our phone agents call these customers individually; what we realized was that really is a very laborious and slow way to do this. To implement offers we were using in A/B testing and doing maybe two different test cases in tandem — it would take us four to six weeks to spin it up, run it through, and understand the results,” Niles explained.

“Our hypothesis was you could use advanced analytics and AI to actually reach out to this population more efficiently and effectively,” he added. “Our churn rate is now down to 13%, and we will go even lower.”

How the AI works on the back end

For IT managers, software architects, and engineers who are interested in understanding how OfferFit’s inside tech works, here are some data points from Dr. Victor Kostyuk, cofounder and CTO of OfferFit.

VentureBeat: What AI and ML tools are you using specifically?

Victor Kostyuk: OfferFit uses reinforcement learning (RL), also known as self-learning AI. Reinforcement learning has only reached maturity recently, so very few companies today have the capabilities to use it effectively. The specific RL models we use are ensembles of contextual bandits of varying complexity.

VentureBeat: Are you using models and algorithms out of a box — for example, from DataRobot or other sources?

Kostyuk: No, because reinforcement learning is so new, standard tooling for it does not yet exist (that’s why we’re building OfferFit). We build all of our models and our modeling framework custom, using only the foundational tools of data science (Python, NumPy, TensorFlow).

VentureBeat: What cloud service are you using mainly?

Kostyuk: Google Cloud Platform. Using BigQuery, GCS (buckets), and GKE (managed Kubernetes).

VentureBeat: Does that mean you’re using a lot of the AI workflow tools that come with Google Cloud?

Kostyuk: We’re not using AI-specific GCP services (for the reasons mentioned above). All pipelines and services run containerized on Kubernetes.

VentureBeat: How much do you do yourselves?

Kostyuk: We always use existing tooling whenever possible, since we’re all about efficiency. But for reinforcement learning, there’s not a lot of existing tooling available, so we end up doing nearly everything ourselves.

VentureBeat: How are you labeling data for the ML and AI workflows?

Kostyuk: In reinforcement learning, the analog of a label is the reward function. For example, a streaming video company might use OfferFit to maximize the conversions from a free trial. In this case, if a customer converts, they are labeled “one,” whereas if they cancel, they are labeled “zero.” The model makes decisions to navigate the exploration/exploitation tradeoff (this is the central concept of RL) — i.e., it tries to maximize the reward, while still making room to explore and try new things.

Thus, the “labels” come from customer responses (e.g., purchasing or not purchasing) to previous actions, such as sending a given offer. The actual reward function used to train agents is “shaped” to increase sample efficiency and improve learning.

VentureBeat: Can you give us a ballpark estimate on how much data you are processing?

Kostyuk: Daily flows depend on the company (their customer base and frequency and type of customer interaction — e.g., clickstream events can be massive) and the use case. Usually millions of rows per day per company.

What’s next for OfferFit

With the funding, OfferFit plans to create the OfferFit Portal, which will allow OfferFit customers to build custom reports and adjust their OfferFit configurations without the need for engineering support, the company said.

OfferFit was founded in October 2020 by George Khachatryan and Kostyuk, two Cornell mathematics PhDs and alumni of McKinsey and Boston Consulting Group.

The company’s funding round was led by Canvas Ventures with participation from Alumni Ventures Group, as well as angel investors Jeremy Stoppelman (founder and CEO of Yelp) and David Edelman (former CMO of Aetna).