KEY TAKEAWAYS
A Google Cloud Next ’24 panel framed AI project maturity in five stages: Prompt, RAG, Tune, Ground, and Agents. Moving from simple prompts to retrieval, custom tuning, externally verified grounding, and task-performing agents increases contextual accuracy, consistency, and autonomy—and with them, potential ROI and operational efficiency for B2B implementors.
The Prompt stage uses a simple, low-context request, such as generating a product description from a product title alone.
RAG retrieves relevant enterprise information so Gen AI responses go beyond the model’s original training set.
Tuning customizes a foundational model on specialized enterprise data so outputs match an organization’s content style and consistency.
Grounding extends RAG with externally verifiable sources such as manufacturer sites or Google search to improve factual accuracy.
At the Agents stage, apps can perform tasks, make decisions, interact with their environment, and learn from mistakes with minimal human intervention.
In our recent article, we talked about the phases of your Generative AI adoption, namely discovery, proof-of-concept, implementation, and scaling. In a panel discussion during Google Cloud Next ’24, Ali Arsanjani, Head of AI CoE at Google; Kaushal Kurapati, SVP of Product Management AI & Search at Salesforce; and Seth Siciliano, Head of horizontal ISV partnerships at Google, talked about the maturity level of your AI project. These levels, from Prompt to Agent, offer a broad overview of the project’s technological sophistication and provide a practical roadmap for implementors. Let’s delve into each of these five stages:
- Prompt: At this level, the app uses a simple prompt without a lot of context to get the response. An example would be “Generate a product description for a product titled Acme Chrome Modern Kitchen Faucet.”
- RAG: Retrieval Augmented Generation (RAG) finds and uses pertinent enterprise information to generate a detailed, and contextually relevant response from a Gen AI model. RAG allows models to move beyond their training sets. A simplistic example would be to automatically locate specifications of the product and examples of similar products and provide it in the prompt above to create product descriptions.
- Tune: At this stage, the app is using a custom-tuned foundational model. This involves training the AI on a specialized enterprise dataset or adjusting its parameters to optimize performance for particular applications. Now the model speaks like your existing content teams and provides consistent product descriptions.
- Ground: Grounding extends your RAG with externally verifiable data sources, grounding it in reality. It could be manufacturer sites or Google search. It provides much more data for the models to work with. Extending our example, grounding would check your specs with the manufacturer’s websites and similar products to ensure accuracy.
- Agents: At the final maturity level, apps or agents can perform tasks, make decisions, and interact with their environment with minimal human intervention. They can even learn from their mistakes. Now, our catalog agent app can scan product catalogs and PIM sources, correct factual inaccuracies, and spot typos for our content creators to review and edit.
By recognizing and understanding the various tiers of AI maturity, we not only gauge the complexity of AI initiatives but also establish a clear roadmap and anticipate their potential applications. Each stage promises increased return on investment (ROI) and operational efficiency. It’s more than just navigating through AI or maturity levels; it’s about leveraging AI to enhance our capabilities and optimize performance.
Ready to dive deeper into the future of AI? Let’s connect and explore possibilities together.