When your enterprise is handling transactions between 25 million sellers and 182 million buyers, supporting 1.5 billion listings, manual decision-making processes just won’t cut. Such is the case with eBay, the mega commerce site, that has been employing artificial intelligence for more than a decade. As Forbes contributor Bernard Marr points out, eBay employs AI across a broad range of functions, “in personalization, search, insights, discovery and its recommendation systems along with computer vision, translation, natural language processing and more.”
As part of a massive operation with so much experience with AI, Mazen Rawashdeh, CTO of eBay, has plenty to say about the current state of enterprise AI. He recently shared his views on AI’s progress across the business landscape, and where work is still needed.
How far has AI moved beyond proofs of concept?
Rawashdeh: The technology behind AI has progressed way beyond proofs of concept in many organizations. AI is at the front and center of technology strategy and execution, driving compelling customer experiences, improving business growth, and managing and reducing risk across almost every industry — finance, healthcare, transportation, security, e-commerce. In a way, it is beginning to touch several aspects of human life in a practical manner. Computer vision, natural language processing, recommender systems, and anomaly detection capabilities, for example, are fundamentally shaping the future of commerce in general, and e-commerce in particular.
Is AI being narrowly applied to specific tasks, or are there broader applications underway?
Rawashdeh: AI is currently being applied both wide and deep across industries. For example, solutions are deployed in production at scale for specific tasks such as language translations, intelligent searches, personalized experiences, fraud detections, recommender systems, across e-commerce industries.
These are foundational capabilities and quickly becoming table stakes; however, AI is emerging and aspiring to have broader applications when it is leveraged to augment human tasks. For example, a combination of AI and human evaluation is being used for fraud detection of prohibited and counterfeit items in the e-commerce industry. As AI is deployed to manage more human tasks, it raises the critical policy, regulatory and ethical considerations that need to evolve as well.
What are the structural roadblocks that inhibit AI efforts and utilization?
Rawashdeh: In order to democratize AI in an enterprise, there has to be an effective and efficient enterprise-to-enterprise machine learning platform that helps the full machine learning lifecycle along with providing higher level AI services, including computer vision, natural language processing and personalization, in easy-to-use modalities. Building these capabilities and services is not an easy undertaking and requires a strong commitment of support from executive leadership, along with an internal open source engineering model and the mindset to develop it collaboratively.
The fundamental roadblocks to successful adoption of AI at the enterprise level is as much about culture as it is about technology. Companies that establish a culture where AI is blended as part of the unified strategy, design and development process, have a higher chance of successful adoption of AI, and in turn, a greater return from that AI. When AI is thought of an ecosystem across the organization — business, policy, product, technology, experience — then the ROI can be maximized.
What kind of infrastructure is providing the best support for broader AI initiatives at the enterprise level?
Rawashdeh: There are three key pillars to build successful AI initiatives in any enterprise from a hardware and software infrastructure perspective.
First is to have an easy discoverability, transformation and cleaning framework for data;
second is to have an extensive high-performance compute, storage, network to train, validate, and deploy complex machine learning and deep learning AI models; and
third is the availability of a control plane for AI that includes various software frameworks and utilities for end-to-end management of AI modeling lifecycle from exploration, training, experimentation, learning and iteration.
What changes are required within the data infrastructure to support scaling AI?
Rawashdeh: Data infrastructure and the teams and processes behind scaling AI need to provide a ‘data as service’ type capability for any successful deployment. This enables data scientists and developers in an enterprise to discover, create, manage, deploy and share best-of-breed ‘data features’ in a quick and seamless self-service manner.
To support AI scaling, the data infrastructure should look beyond traditional data warehouses or extract transform load, to provide simplistic and appropriate AI specific abstractions for data discovery, data preparation, model training and serving. For AI to be effective, the infrastructure should provide data for models in batch as well as real-time.
Most importantly, AI is an iterative, continuous learning process, requiring automated and continuous feedback data for model iterations. The data infrastructure should evolve to support such a continuous feedback cycle from AI systems and human-in-the-loop.