When you need help, you want answers fast — no endless searching or waiting. An AI agent that’s set up with the right knowledge is perfect for the job.
In October 2024, we launched Agentforce on Help to make it easier for customers to get answers – fast. Before that, our Help Portal worked just like most support sites: tons of content, search tools, videos, and case submissions. It did the job, but we kept hearing the same feedback:
- There’s too much information.
- It’s hard to find the answer quickly.
We knew we needed something better. Enter Agentforce — a digital teammate designed to deliver instant, relevant answers.
Of course, just like any new employee, Agentforce had a lot to learn. Some lessons were smooth, while others gave us valuable insights. Here’s what we got right, what we had to rethink, and the key takeaways from our journey.
D-I-Why? Deploy AI agents faster with Agentforce
Building and deploying autonomous AI agents takes time. Agentforce, the agentic layer of the Salesforce platform, can reduce time to market by 16x compared to DIY approaches — with 70% greater accuracy, according to a new Valoir report.
1. Set Up Your AI Agent Like You Would a New Hire
When you onboard a new employee, you don’t just throw them into the deep end and hope they figure it out. You give them the right resources. You guide them.
That’s exactly how we approached Agentforce.
We started by pulling 740,000 pieces of content from our Help Portal – product documentation, release notes, and knowledge articles – and hydrated it into Data Cloud. This became Agentforce’s foundation for responses.
But content alone wasn’t enough. Agentforce needed to understand how to apply that knowledge in real-world conversations. So, using Agentforce Builder, we created topics and actions that would enable it to function as a true digital teammate.
One of these topics, global instructions, set key parameters: defining that Agentforce is part of Salesforce, the voice and tone it should use, and how to structure its responses.
With these parameters in place, Agentforce could interpret, prioritize, and refine information dynamically — ensuring every response was relevant, accurate, and aligned with customer needs.
Pro Tip: Data is just the starting point. Continuous learning and iteration are what turn a digital assistant into a true teammate.
2. Smart Guardrails, Not Roadblocks
Setting up your AI agent isn’t just about what it should do – it’s just as important to define what it shouldn’t do.
We set guardrails to ensure Agentforce only answered questions related to Salesforce and knew when to redirect users to official resources or seamlessly hand off to a support engineer when needed – because AI works best when it enhances, not replaces, human expertise.
But here’s what we didn’t expect: if the guardrails were too strict, Agentforce would just…go silent.
For example, we initially restricted discussions on competitor products. Makes sense, right? But when a customer asked, “How do I integrate a competitor with Salesforce?” — Agentforce simply stopped responding. Not because it lacked content, but because it was blocked from addressing the topic entirely.
So, we adjusted our approach. Instead of outright blocking topics, we configured Agentforce to recognize when a question fell outside its scope and guide users to the appropriate Salesforce resources. We also reinforced its role as a Salesforce support engineer, ensuring every interaction stayed aligned with our brand and connected customers to the right help when they needed it.
Pro Tip: Overly strict rules can make an AI agent unhelpful. Focus on guiding conversations rather than blocking them outright.
The One Factor That Defines AI Agent Success
AI agents are about to change everything. But bad data is their junk food. Nourish them with clean, complete, connected data. Here’s how.
3. Start Small, then Scale Fast
When we first launched Agentforce on Help, we kept it small – rolling out to a small group of authenticated users.
Why? Because, honestly, we wanted to be intentional in our approach. We weren’t sure how it would perform in the real world. Would it get things wildly wrong? Would it make up answers? Would screenshots of bad responses end up on Reddit? We didn’t know.
We had tested it thoroughly, but there’s nothing like real-world use to show you where things can go sideways.
Once we saw how Agentforce handled real conversations, we made adjustments in Agentforce Builder, and expanded access step-by-step.
As we scaled and usage grew, we saw gaps in certain areas – like developer-related questions. The content wasn’t there. So we expanded its knowledge base, adding Developer Portal content, Slack content, and more. Now, Agentforce could answer developer questions much better, giving more precise and relevant responses.
Today, Agentforce handles hundreds of thousands of conversations per week — and we’re still modifying it based on feedback.
Pro Tip: Don’t launch at full scale right away. Start with a limited audience, listen to feedback, and refine your prompts and actions before expanding.
Keep Refining, Keep Improving
Agentforce, like any new employee, is always learning. The biggest takeaway? Iteration is everything.
The best thing we did was stay hands-on after the launch — updating knowledge sources, and grounding our AI in customer data. With the right data, smart configuration, and a thoughtful rollout strategy, AI can become a powerful customer success asset.
Now it’s your turn. Ready to set up your AI agent? Start by refining your data sources, testing responses, and iterating based on real conversations. The more you iterate, the better it works — for you and your customers.
