Originally published on the Databricks blog
At Kard, we believe better data leads to better rewards — and that starts by understanding what people actually buy.
By categorizing transactions at scale, we're able to help brands target the right customers, issuers increase card usage, and consumers get rewarded in ways that feel personal.
Historically, categorizing transaction data was messy and manual. But with a new Databricks-powered approach, Kard is now able to classify billions of transactions quickly, accurately, and flexibly, laying the foundation for personalized rewards that drive loyalty and long-term value.
What Kard does
Kard drives loyalty for every cardholder and shopper through a rewards marketplace.
Our platform gives brands like Dell, CVS, Allbirds, and Round Table Pizza access to tens of millions of consumers by delivering cash back offers through issuer and fintech banking apps, rewards programs, and EBT platforms. Seeing a 10% or 15% cash back offer nudges customers toward a purchase (often one that’s higher in order value).
And on Kard’s pay-for-performance model, brands only pay when a purchase occurs, ensuring ample reach without the high costs or risks of traditional media buying.
Cash back rewards benefit the issuers and fintechs, too. By offering rewards that users care about, they increase engagement and usage among their cardholders.
But what makes Kard particularly special is the category-level insights it captures, providing insight without exposing any PII.
Why category-level insights matter for rewards
Knowing what users spend their money on helps brands (and banks and fintechs) understand their customer bases in a richer way. In aggregate, the spend patterns Kard collects:
- Fuel smarter marketing campaigns — you can identify high-intent segments based on behavior. For example, if a large percentage of users regularly use rideshare services late at night, banks and brands can target them with weekend-specific cashback offers.
- Inform product design by revealing unmet needs. If data shows that younger users are shifting spend from grocery stores to food delivery apps, a fintech might prioritize rewards tied to convenience-driven categories.
- Inspire new partnerships by surfacing common merchant overlaps across user cohorts. For instance, if frequent travelers consistently book the same chain of hotels and rental car agencies, there’s a strong case for negotiating co-branded rewards or exclusive perks with those partners.
Categorical patterns get even more powerful when you zoom in on the individual.
For instance, perhaps a specific user spends the most on sports gambling. A generic retail offer might go unnoticed, but a promo for a betting app could drive instant engagement.
Say a different user has decreased spend on groceries but increased their use of food delivery apps over the last 90 days. That signals shifting habits — and an opportunity to reward convenience over cost.
Finally, another user flies often, but always with the same airline. That loyalty can be reinforced with targeted rewards, or even upsold to that airline’s premium tier. Other airline brands may not even want to target that individual. Or they might only surface the highest cash back offers to improve their odds of stealing the customer away from their preferred airline.
Without reliable transaction categories, though, none of these personalization scenarios are possible.
How rewards platforms historically labeled transactions
Categorization is the key to unlocking high-ROI go-to-market strategies for our brands and issuers, but it’s harder than it sounds.
First, you’ve got to label all the transactions. Traditionally, there’ve been two ways to accomplish this:
- Have analysts review each transaction, line by line, tagging each one according to a predefined taxonomy. As you might guess, this method is tedious, error-prone, and incredibly hard to scale.
- Let users categorize their own transactions. While this approach leaves less work for analysts, it also riddles the data with inconsistencies. One user might label Domino’s as “fast food,” another might call it “pizza,” and a third might tag it “comfort food,” making it extremely difficult to draw reliable insights.
Once a substantial amount of transactions are labeled, engineering teams can start training machine learning models like LightGBM, XGBoost, or BERT to predict categories for new, unseen transactions.
Over time, these models could eliminate the need for manual tagging. However, they require maintenance and upgrades as businesses evolve and transaction formats change. Adding new category types (say, for an emerging industry or a new client vertical) could involve retraining or even re-architecting the model.
To support our growing business, we needed a more streamlined, accurate, and flexible approach to categorizing the billions of transactions we receive each month.
How Databricks powers a modern categorization approach
Working with Databricks, we’ve come up with a unique, scalable system for transaction categorization:
- Leveraging Databricks AI Functions to run batch, agentic workflow that categorizes transactions based upon an internally derived taxonomy.
- The results are constrained with structured output functionality, using the json_schema response format with the enum feature to limit errors.
- AI agents process incoming transactions against the required taxonomy, one for each type of categorization. In one instance, we can capture high-level categories like Travel, and then identify hierarchical categories like Travel → Airfare and even further, Travel → Airfare → Regional Airline.
- Inconsistencies are passed down to paths that are evaluated by agent judges, whichallows for re-categorization in the case of mistakes.
The lightweight costs of this new approach have given our team more flexibility. If a new line of business opens up, we can alter our categories right away — without having to totally retrain the model. In fact, we just opened up some new CPG categories to support a partnership with a popular rewards app.
Some of our clients have requested that we use their own category mapping to align with their internal systems. Now, we can just pass that alternative taxonomy straight to our new system and it’ll translate outputs accordingly.
“Being able to roll up merchants into their respective categories offers us a lot of leverage with customers,” says Chris Wright, Kard staff machine learning engineer.
“For example, we can tell merchants that users within their category typically find offer types x, y, and z work best. We can also help merchants target a segment of users who have purchased with them in the past and had a recent acceleration in spend within, say, food delivery or ride share. And we can tell our customers who they’re competing with in their category and region so they can refine their campaigns accordingly.”
What’s next for Kard and Databricks: hyper-personalization
Transaction categories may seem like a behind-the-scenes detail. But the agility we get from the Databricks AI Functions-powered categorizer makes it possible for us to move fast without breaking our data foundation, and have confidence in the scalability of the solution.
Plus, it also opens the door to new kinds of products and services for Kard customers, like:
- Personalized card offers based on shifting food or travel habits
- Stickier rewards for loyal customers of a specific merchant
- Smart nudges based on time-of-day or seasonal behavior
- Merchant-funded cashback programs targeted by segment, not just demographics
- Earned points programs (for brands and issuers)
By investing in smarter categorization now, we’re laying the groundwork for a truly personalized rewards experience that boosts purchase frequency, increases AOV, and sustains customer loyalty for brands and issuers alike.
