Artificial Intelligence FAQ

In the past, AI development entailed designing and building highly-complex, one off models that were fragile, error-prone, difficult to understand, and even more difficult to support. With the emergence of open source AI function libraries such as TensorFlow and PyTorch, this landscape has changed dramatically. Today, many complex AI problems can be addressed using functionality from these libraries, turning custom development from an esoteric exercise into a component assembly process – what we refer to as component integration.

Without consistent, accurate, timely data, even the most sophisticated AI model or powerful use case doesn’t work.

• Reliable data sources (internal or external) must be identified, in some cases contracted for, and integrated into your data infrastructure.
• Data must be normalized, cleansed, and in some cases transformed for consistency, otherwise models will not perform.
• Data must be received, processed, and available by the time it is needed.
• Most importantly, your data infrastructure must work. Every time.

Some AI solution providers tout their ability to develop a minimum viable product (MVP) at a very low cost. Typically, an MVP includes using your test data set to implement a simplistic base model. Unfortunately, a minimally viable solution doesn’t include the cost and effort to implement a resilient data infrastructure; ramp the model to reflect real world data volumes, the cost of data (internal and external); ongoing testing, validation, and retraining; careful curation of data/model/result sets; compliance requirements for explainability and ethical behavior; and don’t forget user adoption.

The premise and promise of AI means that some people’s jobs are going to change. When users can’t accept and support the changes that AI brings, projects fail. Change readiness, stakeholder management, communications, training, and reinforcement are all important factors to consider.