Industry executives and experts share their predictions for 2023.  Read them in this 15th annual VMblog.com series exclusive.

Delivering Improved Business Outcomes with Scalable, Long-Lasting, Human-Centered AI Systems in 2023

By Dr. Sanjay Rajagopalan, Chief Design & Strategy Officer, Vianai Systems

In 2018, Gartner estimated that through 2022, 85% of AI projects would "deliver erroneous outcomes due to bias in data, algorithms, or the teams responsible for managing them." Even if they perform well during initial development and testing, AI models can quickly become error-prone, drifting away from the regime where they were trained to perform well and ultimately becoming untrustworthy. Further, even when they perform well from a technical perspective, they fail to deliver the expected business outcomes.

In 2023, many productive applications of these latest technologies will enter the enterprise, where AI-led transformation initiatives are already increasingly important. 2023 will continue to drive AI adoption, but to sustain the current momentum and harness its full potential, businesses will also need to better understand the technology, its boundaries and limitations.

Improving AI Monitoring & Continuous Operations

AI/ML models are not like traditional business applications. The performance of these models is often tightly coupled to the data-regime prevalent during the training phase, which can change (or drift) in production. Hence, the behavior of a model in production can be very different from the behavior of a model during testing.

When it comes to AI/ML, this becomes a particularly acute problem with an added dimension of complexity. AI and machine learning applications are, by nature, learning systems in that they model past behavior patterns contained in data. These systems fundamentally behave differently as the data on which they operate changes, and this can be a big problem if the data-regime shifts beyond certain limits.

Traditional applications run on the assumption that we can codify an idea in the form of an algorithm or business logic, and then the application can perform the task acceptably well in perpetuity. This may not hold true for AI/ML applications, even for short durations. As the data changes, the model logic needs to adjust. Data-driven systems can become obsolete much more quickly than the typical enterprise applications that users are accustomed to.

To counter this in the new year, enterprises need to enable continuous monitoring of both inputs to and output from models over an extended period to understand how they behave in the real world and what risks they might introduce over time. This includes things like uncertainty or bias in the data used to train models, and the shifting of run-time inference data or model outputs beyond certain boundaries. It also includes key steps in the continuous operations of systems to immediately identify root causes of problems in order to fix them, retrain models when and if needed, optimize resource consumption, and validate that models continue to perform as intended over a long period of time. Understanding the limits and setting thresholds that trigger retraining or re-modeling must be agreed upon collaboratively by all stakeholders before putting a model into production and managing such a dynamic landscape at scale.

Bringing Generative AI to the Enterprise Through a Human Lens

AI has made impressive progress in its ability to create visual and written artifacts. In 2023, we will also see enterprises incorporating Generative AI and Large Language Model capabilities into their business workflows and systems. The high-level of excitement around these new technologies (e.g. MidJourney, Stability.ai, ChatGPT etc) show a desire for AI technologies that augment human creativity and support evidence-based decision-making. However, deep within the systems of today are unfathomably large-scale mechanisms which use opaque AI techniques to re-constitute artifacts from the things they have ingested, in most cases, retaining the same biases and prejudices present in the sources, and sometimes "hallucinating" about the real world.

For this reason, AI will not replace human creators who can bring critical judgment into decision making regarding the proper use of these artifacts, but will rather augment them. Incorporating human-machine co-operation as a foundation in the design of systems can create intelligent "human-centered AI" systems that can improve and amplify the creative process while still allowing for human judgment to be applied. The best AI will be transparent, observable, and explainable, always making their users aware of their strengths and limitations. The feedback from humans will naturally improve the AI's performance and outputs - but humans will always be necessary to bring trust and credibility.

Closing the Gaps

There is a gap in tools and solutions that facilitate clear ownership, communication, and collaboration in long-duration, iterative, at-scale processes involving Business Sponsors, Data Scientists, Machine Learning Engineers (MLE) and Machine Learning Operations (MLO) teams.

Model development tools, which are mainly aimed at data science professionals or "citizen" data scientists, are plentiful but highly fragmented. Development toolchains are well entrenched within their respective communities of developers and are challenging to disrupt, despite their many shortcomings. Irrespective of how models are developed, when they are taken into production, the rubber hits the road (and stays on the road for a while). A model that takes a few months to build and test may live on in operation for years and even decades. Few companies pay attention to what it takes to keep models in operation, running smoothly throughout their extended lifetime.

In 2023, new tools need to provide enterprises with a state-of-the-art foundation for succeeding with human-centered AI applications at scale. This could mean including advanced features for model management, continuous operations, and observability. Systems must be designed, managed and monitored with the right tools in order to recognize the limitations of AI, operate it continuously and sustainably, and take advantage of the significant potential. In the enterprise, this means things like quickly reading, analyzing or parsing through massive amounts of data and then serving it up to decision-makers. It means performing repetitive tasks at scale based on human input, providing new levels of efficiencies or other business outcomes. Building such systems will need a design-led approach to problem definition and solution design.

AI in 2023 and Beyond

In the future, we will certainly see more widespread productive use of AI in the enterprise - with sophisticated predictive and prescriptive models, and with large language and generative visual capabilities integrated into enterprise systems and workflows. This will become possible with an expansion of human-centered AI, bringing human creators and AI power together - but it will not happen without a robust framework for model management, observability and continuous operations.

AI leaders and developers will need to focus on creating AI that is not only beneficial, but also safe and trustworthy. In 2023, shifting priorities to creating an AI with humans at its core will allow us to create intelligent systems that can improve the world around us, and bring us to the unimaginable heights which neither humans, nor AI alone, can reach.

##

ABOUT THE AUTHOR

Dr. Rajagopalan has extensive experience and background in Human-centered Design as it relates to the Enterprise - in products, processes, organizations and services.  He is currently Chief Design & Strategy Officer at Vianai Systems, a Palo Alto, Calif., based company with a mission to bring human-centered AI to life in enterprises worldwide, and help companies realize the full potential of their AI investments. Previously, Sanjay was the SVP and Head of Design & Research at Infosys, and the SVP of Design and Special Projects at SAP. He holds a Ph.D. in manufacturing and design from Stanford University.