At Kard, we imagine higher information results in higher rewards — and that begins by understanding what individuals really purchase.
By categorizing transactions at scale, we’re in a position to assist manufacturers goal the correct clients, issuers improve card utilization, and customers get rewarded in ways in which really feel private.
Traditionally, categorizing transaction information was messy and guide. However with a brand new Databricks-powered strategy, Kard is now in a position to classify billions of transactions shortly, precisely, and flexibly, laying the muse for customized rewards that drive loyalty and long-term worth.
What Kard does
Kard drives loyalty for each cardholder and shopper via a rewards market.
Our platform provides manufacturers like Dell, CVS, Allbirds, and Spherical Desk Pizza entry to tens of thousands and thousands of customers by delivering money again provides via issuer and fintech banking apps, rewards applications, and EBT platforms. Seeing a ten% or 15% money again supply nudges clients towards a purchase order (typically one which’s increased so as worth).
And on Kard’s pay-for-performance mannequin, manufacturers solely pay when a purchase order happens, guaranteeing ample attain with out the excessive prices or dangers of conventional media shopping for.
Money again rewards profit the issuers and fintechs, too. By providing rewards that customers care about, they improve engagement and utilization amongst their cardholders.
However what makes Kard notably particular is the category-level insights it captures, offering perception with out exposing any PII.
Why category-level insights matter for rewards
Figuring out what customers spend their cash on helps manufacturers (and banks and fintechs) perceive their buyer bases in a richer manner. In mixture, the spend patterns Kard collects:
- Gasoline smarter advertising campaigns — you possibly can determine high-intent segments based mostly on habits. For instance, if a big proportion of customers usually use rideshare companies late at evening, banks and types can goal them with weekend-specific cashback provides.
- Inform product design by revealing unmet wants. If information exhibits that youthful customers are shifting spend from grocery shops to meals supply apps, a fintech would possibly prioritize rewards tied to convenience-driven classes.
- Encourage new partnerships by surfacing frequent service provider overlaps throughout consumer cohorts. For example, if frequent vacationers constantly e-book the identical chain of resorts and rental automobile businesses, there’s a robust case for negotiating co-branded rewards or unique perks with these companions.
Categorical patterns get much more highly effective while you zoom in on the person.
For example, maybe a selected consumer spends essentially the most on sports activities playing. A generic retail supply would possibly go unnoticed, however a promo for a betting app might drive on the spot engagement.
Say a special consumer has decreased spend on groceries however elevated their use of meals supply apps over the past 90 days. That alerts shifting habits — and a possibility to reward comfort over price.
Lastly, one other consumer flies typically, however all the time with the identical airline. That loyalty might be bolstered with focused rewards, and even upsold to that airline’s premium tier. Different airline manufacturers could not even need to goal that particular person. Or they could solely floor the very best money again provides to enhance their odds of stealing the shopper away from their most well-liked airline.
With out dependable transaction classes, although, none of those personalization situations are attainable.
How rewards platforms traditionally labeled transactions
Categorization is the important thing to unlocking high-ROI go-to-market methods for our manufacturers and issuers, however it’s more durable than it sounds.
First, you’ve obtained to label all of the transactions. Historically, there’ve been two methods to perform this:
- Have analysts evaluate every transaction, line by line, tagging each in accordance with a predefined taxonomy. As you would possibly guess, this technique is tedious, error-prone, and extremely laborious to scale.
- Let customers categorize their very own transactions. Whereas this strategy leaves much less work for analysts, it additionally riddles the info with inconsistencies. One consumer would possibly label Domino’s as “quick meals,” one other would possibly name it “pizza,” and a 3rd would possibly tag it “consolation meals,” making it extraordinarily troublesome to attract dependable insights.
As soon as a considerable quantity of transactions are labeled, engineering groups can begin coaching machine studying fashions like LightGBM, XGBoost, or BERT to predict classes for brand new, unseen transactions.
Over time, these fashions might remove the necessity for guide tagging. Nonetheless, they require upkeep and upgrades as companies evolve and transaction codecs change. Including new class varieties (say, for an rising trade or a brand new shopper vertical) might contain retraining and even re-architecting the mannequin.
To help our rising enterprise, we wanted a extra streamlined, correct, and versatile strategy to categorizing the billions of transactions we obtain every month.
How Databricks powers a contemporary categorization strategy
Working with Databricks, we’ve give you a novel, scalable system for transaction categorization:
- Leveraging Databricks AI Features to run batch, agentic workflow that categorizes transactions based mostly upon an internally derived taxonomy.
- The outcomes are constrained with structured output performance, utilizing the json_schema response format with the enum characteristic to restrict errors.
- AI brokers course of incoming transactions in opposition to the required taxonomy, one for every sort of categorization. In a single occasion, we are able to seize high-level classes like Journey, after which determine hierarchical classes like Journey → Airfare and even additional, Journey → Airfare → Regional Airline.
- Inconsistencies are handed all the way down to paths which are evaluated by agent judges, whichallows for re-categorization within the case of errors.
The light-weight prices of this new strategy have given our workforce extra flexibility. If a brand new line of enterprise opens up, we are able to alter our classes instantly — with out having to completely retrain the mannequin. In truth, we simply opened up some new CPG classes to help a partnership with a preferred rewards app.
A few of our shoppers have requested that we use their very own class mapping to align with their inner methods. Now, we are able to simply go that various taxonomy straight to our new system and it’ll translate outputs accordingly.
“With the ability to roll up retailers into their respective classes provides us loads of leverage with clients,” says Chris Wright, Kard workers machine studying engineer.
“For instance, we are able to inform retailers that customers inside their class sometimes discover supply varieties x, y, and z work greatest. We will additionally assist retailers goal a section of customers who’ve bought with them up to now and had a current acceleration in spend inside, say, meals supply or journey share. And we are able to inform our clients who they’re competing with of their class and area to allow them to refine their campaigns accordingly.”
What’s subsequent for Kard and Databricks: hyper-personalization
Transaction classes could seem to be a behind-the-scenes element. However the agility we get from the Databricks AI Features-powered categorizer makes it attainable for us to maneuver quick with out breaking our information basis, and believe within the scalability of the answer.
Plus, it additionally opens the door to new sorts of services for Kard clients, like:
- Customized card provides based mostly on shifting meals or journey habits
- Stickier rewards for loyal clients of a selected service provider
- Sensible nudges based mostly on time-of-day or seasonal habits
- Service provider-funded cashback applications focused by section, not simply demographics
- Earned factors applications (for manufacturers and issuers)
By investing in smarter categorization now, we’re laying the groundwork for a really customized rewards expertise that enhances buy frequency, will increase AOV, and sustains buyer loyalty for manufacturers and issuers alike.
Conclusion
On this weblog put up, we confirmed how Databricks AI Features are powering information enrichment for Kard’s categorization pipeline. This permits personalization at scale, and drives loyalty and worth at a fraction of the trouble it could usually take.
Taken with studying extra? Attain out to one in all our specialists in the present day!
About Kard
Kard is a New York-based fintech firm based in 2015 that gives a rewards-as-a-service platform for banks, neobanks, and card issuers. Its API allows monetary establishments to shortly launch and customise cardholder rewards applications, connecting customers to 1000’s of retailers and types throughout the US. Kard’s platform is designed to drive buyer loyalty and engagement by making it simple for cardholders to earn rewards on on a regular basis purchases. The corporate is backed by main traders and serves over 45 million cardholders via its issuer and associate community.