Unlocking Competitive Advantage with Data and GenAI

Relationships within vast datasets unlock possibilities, revealing insights into the past and enabling future predictions.
The volume of collected data can become overwhelming, highlighting the critical relationship between data and AI.

Love Agarwal (Worldwide Sales Leader, IBM Watson x) and Edward Caldasburg (VP of Product Management, IBM Watson X platform) introduce the topic.
The discussion takes place in an IBM research facility, showcasing a prototype AI system.

Conversations about AI inevitably lead to data, emphasizing that there is no AI without data.
Data is a sustainable source of competitive advantage.

GenAI models can effectively utilize unstructured data, which forms the majority of new data.
GenAI can analyze large volumes of language data (documents, software code) to identify patterns and connections with minimal supervision.
GenAI can enhance data management by organizing, refining, and enriching data to improve its quality and ease of consumption.

GenAI can interpret different data encoding methods in legacy applications (e.g., formats, dates, names).
It can understand and relate data across siloed systems, enabling holistic utilization of data.
Effective data management yields significant savings in time and energy, impacting business performance.
Well-managed data becomes intellectual property and a source of competitive advantage.

High-quality data is crucial for enterprises leveraging GenAI to improve their business.
Customizing or tuning large language models with enterprise data and integrating GenAI into applications drives productivity gains, enhances business performance, and creates competitive advantages.

Architectural problems, such as inaccessible or siloed data across on-premise and cloud environments, hinder data utilization.
The data silos problem cannot be solved by creating new silos in the cloud.
Approaches to solving data challenges include building a virtual data layer and querying data across multiple sources, or consolidating data onto a data lakehouse (open and cost-effective).
Overcoming psychological barriers involves shifting the mindset to turn the business into data and then the data into a business.

Data monetization, selling a version of core business data as a product, is the ultimate goal.
Ensuring data quality, security, and governance is essential to prevent business risks and regulatory exposure.

Implement traditional data quality practices, including cataloging and organizing data into a business glossary.
AI can assist in this process, but the more prepared the data, the faster the value derived from AI.

Establish thoughtful data access policies to define what information AI can access.
Redact sensitive information, such as social security numbers and personally identifiable information (PII).

Implement monitoring and enforcement of data policies, set centrally and enforced locally.
Actively monitor model inputs and outputs to ensure policy effectiveness and track changes due to real-world interactions.
Data and AI governance is critical; risk and uncertainty about data and model outputs are major barriers to AI adoption.
Companies must protect their own and their clients' data.

Quality, trusted data is essential for successful GenAI implementation.
Many clients struggle to move beyond prototypes to customizing models with their own data and deploying them across the enterprise.

Two primary methods:
- Tuning the Model: Instructing or partially retraining the model with enterprise data examples to tailor its responses. This allows the model to quickly learn and adapt to the business's specific language and structure.
- Retrieval Augmented Generation (RAG): Using a knowledge base of quality enterprise data to improve accuracy and limit responses to known facts, minimizing the risk of hallucinations.

IBM Research uses a data lakehouse architecture for training its foundation models.
- Data is sourced, cataloged, filtered, and transformed.
- The data is used to train, test, and tune large language models.
- End-to-end lifecycle governance is implemented to maintain complete lineage of datasets, data pipelines, and AI pipelines.

A data lakehouse is likened to a commercial kitchen or a well-stocked pantry.
Instead of sourcing data (ingredients) from disparate locations, a data lakehouse centralizes and organizes data for quick access.

Combines flexibility, scalability, and cost advantages of data lakes with the performance and functionality of data warehouses.
Enables flexible, open, and interoperable architectures, allowing users to choose the right tools for the job.
Supports different query engines for various tasks (data ingestion, interactive queries, document embedding).
Offers optimal price performance, essential for enterprise-scale deployment.

A key lesson learned from scaling AI adoption is to avoid building models on the sidelines of core enterprise IT.
Implement data and AI governance frameworks from the beginning to manage and track the end-to-end lifecycle.

Open source technology fosters transparency, security, and project oversight.
Transparency and security are enhanced through broad user and developer participation in finding and fixing vulnerabilities.
Project oversight involves a diverse community driving the technology's roadmap, aligning with broader community interests.

Rapid GenAI implementation can differentiate winners and losers across industries.
Preparedness is valuable, but starting with proper guardrails is essential.