In my opening post explaining the scope and mission of AT&T’s Chief Data Office (CDO), I highlighted the immense data flows we manage at AT&T, what we have done to establish a modern and connected approach to data, and some of the applications of AI deployed in recent years to drive actionable business value. In this post, I’ll explore what we call the “democratization” of data and AI – making tools accessible to business leaders and employees who aren’t necessarily trained data scientists or technical experts. Data is for everyone.
As AT&T’s North Star for data, analytics and AI, CDO is acting as a catalyst to spread reusable capabilities across AT&T to support greater data-driven decision making at all levels of the company. We have empowered business managers with self-serve access to “single version of truth” datasets using business intelligence tooling. We are also extending low/no-code AI creation capabilities across the firm. This unlocks a larger segment of analytic talent beyond just our code-savvy cohort within AT&T to create optimized and responsible AI solutions. CDO is focused on analytic skill development and employee growth to mobilize this larger segment of analytic talent that we call our “citizen data scientists.” We are revolutionizing how employees across AT&T navigate the data and AI lifecycle: from find and getting data, to engineering the data for machine learning, to creating, deploying, monitoring and governing machine learning models used in artificial intelligence.
To support the front end of the data and AI lifecycle, CDO created two capabilities to speed up the process.
First, CDO built a centralized data intelligence platform that inventories all of AT&T’s data and takes advantage of metadata search capabilities that speeds up the process of finding information scattered around the company. This platform provides a single enterprise-wide entry point to AT&T’s 160+ million data assets, regardless of where they are stored across the company’s legacy systems and reference catalogs. The functionality, using best in class functionality like Collibra, includes a simple-to-use, natural language search interface, making it easy for users to find and filter a wide range of reusable customer and operational business intelligence and AI building blocks (i.e., reusable software code and visualization tools), In addition to enabling the search across the company, we have enabled collaboration at all levels of expertise in data analytics so that we can both accelerate analytics and get broader input on our overall business and technical strategy.
Second, to get a deeper look at those data and AI tools within our catalog, AT&T and Silicon Valley startup H2O.ai co-created an AI Feature Store which is a computation and storage platform filled with production-ready machine learning features for model development. The Feature Store makes these AI building blocks discoverable, usable, shareable and reusable across the enterprise. A visual interface enables users to browse the feature catalog to review available features, identify the most popular ones, and determine their fit for reuse in new models. In many cases, features used for past models can be reused to immediately start new ML projects – significantly shortening time-to-market. Plus, users can see detailed data and source code on each feature as well as each model’s current and past performance. Because the Feature Store is constantly updated with pipeline code tweaks or newly arrived data, it also automatically validates and sends new feature recommendations and updates to users – an industry first.
With these data sets and features in hand, we turn to the cloud where we utilize cutting edge tooling to develop the best machine learning models for the job. We collaborated again with H2O.ai to start the journey by automating model development, using the latest systems to suggest the best AI given the data and features at hand. Our crew of data scientists can customize and manipulate the features, data and models in notebooks powered by Databricks. Models can be improved even further collaborating and competing with our internal AI crowd-sourced competition platform, where data scientists compete to make the best fit model given the business case and the data.
We publish the best models and integrate them into software to enable the AI to make decisions at scale – sometimes thousands and thousands per day! Models published are automatically documented in a standardized way so that at any moment, we know what our AI decides. All AI models are monitored and governed. We not only check the data coming into the AI for scoring is of the right quality, but also make sure the model does not drift and continues to provide relevant decisions. And of course we use the latest techniques to be sure that the data, features and models that we use are diverse enough to represent all situations our customers may encounter. We are always making sure that our AI is making the best decision it can for our customers.