If you’re familiar with Teradata, you probably know us as a data warehousing company with the biggest, baddest massively parallel processing engine of them all. There’s history and legacy there that we’re very proud of. The largest Teradata systems in the world are measured in petabytes, and the busiest Teradata system in the world runs more than 200 million queries per day, every day.
And data warehouses built on Teradata aren’t simply passive management information systems—they’re mission-critical platforms that support “run the business” operational analytic workloads.
At some of the largest banks in the world, Teradata’s platform is vital for managing customer sales, marketing, credit risk, fraud, compliance, and liquidity management. One of the chief data officers we work with at a major U.S. bank likes to joke that the one call he never wants to take is the one from the corporate treasurer asking why the liquidity management reports and models are down. And that’s precisely why he runs those reports and models on Teradata’s platform, and nowhere else.
Outside of financial services, some of the biggest airlines in the world run their operations on Teradata’s platform. Nothing animal, mineral, or vegetable moves in or out of one of the largest economies in Asia without a query hitting a Teradata system.
But Teradata’s platform also includes ClearScape Analytics™, a rich suite of AI/ML capabilities, including:
- The industry’s most scalable and performant data processing and feature engineering functions
- A rich library of native model training and scoring algorithms
- The ability to run R and Python in parallel and at massive scale
- The ability to deploy and score models regardless of where they’ve been trained, through our Bring Your Own Model (BYOM) capability
In the U.K., for example, one of the top 3 mobile operators has built a man-in-the-middle fraud detection system based on a complex ensemble model using ClearScape Analytics: A model that was developed on Teradata’s platform against features engineered on the platform; a model that was evaluated and deployed on Teradata’s platform; and a model that runs in production on Teradata’s platform multiple times a day, every day, driving seven-figure benefits to the bottom line.
In Brazil, we’re proud to work with Sicredi, one of the country’s largest credit unions. Sicredi has developed a complex income estimation model to allow it to extend credit to workers and entrepreneurs outside the formal economy. That model was developed with Databricks but runs in production on Teradata's platform via the magic of our BYOM capability. It runs 25 times faster than it did on Databricks, enabling Sicredi to rescore its entire client base multiple times a day, avoiding six-figure annual credit agency bills.
At one of the largest European airlines, we’ve just run a proof of concept (POC) that successfully demonstrated that the same BYOM capability enables us to use our platform as a “parallel harness” to run pretrained, task-specific language models at massive scale. This means we can analyse inbound and outbound customer and agency communications to support topic detection, message classification, and sentiment analysis. Unstructured data—the free text conversations—are vectorized in database using pretrained models from the Hugging Face repository, with performant and deterministic similarity search using our vector distance functions for accurate classification.
In plain English: We’ve demonstrated that we can support demanding “AI for CX” use cases without the requirement for data to cross the corporate firewall—even as prompts—and avoiding expensive pay-per-token models. For these use cases, at least, our approach leads to improved accuracy and consistency compared with prompt engineering of large language models (LLMs).
We refer to this process of leveraging unstructured data to reveal signals of customer intent in interaction data as “AI for CX”—and right now, “AI for CX” is a major driver of demand in our customer base. In fact, we’re running POCs like the airline’s with several of the largest banks in the world.
When we look to the future of AI, we see two vital roles for Teradata in the industry.
Role 1: Help organisations with the increasingly urgent challenge of inference economics
We’re currently working with a midsized Australian bank that wants to apply agentic AI to improve the customer experience for customers calling its contact centres while also driving increased efficiency. It’s estimated that merely automating the call-wrap process at this institution would save 36 million minutes of (human) contact centre agent time annually.
This bank takes 50,000 inbound calls a day. Let’s assume that to improve efficiency and enhance the customer experience, we’ll need 10 agents. Each of these agents will need to make 10 inferencing calls to task-specific models—a speech-to-text model, a summarization model, a topic detection model, a named entity recognition model, a sentiment analyser, a credit risk model if we want to cross or upsell the customer, etc.
That’s 50,000 calls x 10 agents x 10 models x 363 days a year—no one in Australia works on Christmas Day or Australia Day—or 1.9 billion inferencing calls per year. And that’s just to improve call centre operations at a midsized bank.
These are rough estimates—and I could easily be out by a factor of 10. If you’re an optimist, then that means 190 million inferencing calls a year—and if you’re a pessimist, it means 19 billion inferencing calls. In either eventuality, all we’ve done is apply agentic AI to the contact centre of a midsized bank—and so you could easily imagine that the total number of inferencing calls that large enterprises will be making in the next few years will be orders of magnitude greater.
When the multiplier is potentially measured in tens or even hundreds of billions, we need unit costs to be low. And that’s why we support three generative AI design patterns:
- In-database-management-system execution of task-specific language models in the 100 million to 250 million parameter range
- In-platform execution of more complex models, where the accelerated compute infrastructure and optimized inferencing algorithms from our partners at NVIDIA, combined with our own Open Analytics Framework, provide a demonstrable price: performance improvement over central-processing-unit-based inferencing for larger and more complex models
- Application programming interface (API) based integration with the foundation models provided by our hyperscaler partners, through tight integration with Amazon Bedrock, Azure OpenAI Service, and Google Gemini
By supporting these AI design patterns, we allow our customers to choose the right approach for the right business problem and optimize for cost, accuracy, performance, throughput, and latency, as appropriate.
Regardless of which design patterns we chose, many or most of the AI applications we imagine for our enterprise customers involve the vectorization of unstructured data like images, audio, and text—and the reprocessing of that data with similarity and dissimilarity searches (for example, to support retrieval augmented generation workflows).
Role 2: Support the storage, management, and processing of vector embeddings
When they’re choosing a vector store platform today, customers often feel that the choice in front of them is: a dedicated solution that performs, but does not scale; or an MPP platform that scales, but does not perform.
Teradata’s platform is already a scalable and performant vector store. That’s the basis of our ability to support AI for CX POCs. Later this year, we’ll introduce even better support for the storage, management, and processing of the millions and billions of vector embeddings we see in our large enterprise customers’ futures.
Crucially, managing those vector embeddings on our platform will enable customers to integrate these new datasets with all the other data they already manage in their Teradata environments. We believe this will be critical for enabling our customers to convert signal to action. It’s one thing to be able to infer that an email represents a customer complaint; it’s quite another to know what to do about it. Knowing what to do requires everything else that we know about the customer. In the organisations we serve, that data is either in or available from Teradata’s platform.
So, contact us today if you need:
- A scalable and performant end-to-end AI/ML platform
- To operationalize models at scale and cost effectively, regardless of where they have been trained
- To dramatically improve data scientist productivity and time to market for new analytics with an enterprise feature store
- A scalable and performant vector store to support your AI initiatives
- To offload expensive and underperforming Apache Spark and Databricks workloads
- To leverage new generative AI tools for advanced “AI for CX” applications
