Problem: As LLMs were first introduced to the world, my team noticed dips in traffic to our content sites. Following the hypothesis that LLMs were starting to replace Google searches for customers looking for solutions to their Salesforce challenges, we started testing some of the top searches ourselves.

The results were concerning. Since LLMs didn't refresh as often we released updates to our content, it was sometimes giving responses based on outdated versions of the content from several months earlier.

At the same time, my team was brainstorming solutions in response to customer feedback that our content wasn't as prescriptive or thorough as it should be around specific use cases.

We weren't the only ones thinking about how to take advantage of generative AI. We found ourselves having similar conversations with our partner teams across the organization that converged on generative AI as the solution. If we were thinking more about how generative AI can address these challenges in our own business, we knew our customers were as well.

Solution: My team prepared our findings in a presentation to platform product leaders, and proposed a high-level plan to address the situation. 

Then we got to work. 

People & Processes: We formed a dedicated workgroup and worked closely with the development team, providing content leadership and expertise throughout the process. 

Customer Experience: Our goal was to establish our content site as the premier destination to take advantage of new generative AI capabilities and receive a faster, more efficient support experience with accurate, timely responses.

• Agentic Workflow Design: The LLM was not only grounded in our authoritative content consisting of more than 750,000 assets, but also customer history and product usage data to provide the most relevant, personalized answers based on each user’s permissions, setup, and customizations.

• Content Optimization: We applied emerging AI-powered content creation tools to prepare the content that would fuel the LLMs. For example, we implemented custom vision models to automatically generate alt text for images, saving the team hundreds of hours. 

• Evaluation & Iterative Enhancement: We started with a pilot of 200 users. My team was responsible for developing the test plan to ensure an enhanced support experience. We also continuously improved the experience by analyzing both agent responses and transcripts from human interactions to identify content gaps, conflicting information, and product-related issues.

Results: When we first launched Agentforce on Help, we saw a resolution rate of more than 80% without human intervention, allowing for the support team to handle only the more complex issues.