If you’ve ever struggled to understand what your AI agent actually did in a conversation, or why it missed the mark, you’re not alone. AI solutions promise better experiences, but too often, Salesforce Admins and agent builders are left in the dark without clear ways to monitor performance or improve the customer journey.
Poor agent performance isn’t just frustrating — it risks your brand reputation. Admins need visibility to catch issues early, fix misconfigurations, and build trust.
While these solutions hold the promise of automating tasks and elevating customer experiences, there has historically been a significant lack of clear visibility into agent-customer interactions. This makes it challenging to truly grasp customer needs and efficiently evaluate agent effectiveness at scale, as manually analyzing hundreds of thousands of interactions is simply not efficient.
At Salesforce, we recognized this critical demand for comprehensive post-conversation visibility as a key factor in improving agent performance and optimizing agent effectiveness. While our previous Utterance Analysis feature laid a foundation for understanding user-agent interactions, it was limited in scope, highlighting the need for more than static reports and dashboards.
To address this, as part of the announcement of Agentforce 3, we developed Agentforce Optimization as part of the Studio experience – a powerful, interactive, and intelligent tool designed to provide in-depth visibility into AI agent interactions. This solution empowers administrators to monitor performance, identify configuration gaps, and proactively address issues, ensuring agents deliver accurate, efficient, and high-quality responses. It achieves this by surfacing meaningful requests from conversations, enabling the quick identification of trends, and supporting business decisions based on actual customer input.
We’ve all been there. You go to your mobile/TV/internet provider’s website and chat with a bot. Well, chatting is a strong word, as essentially what we did was click endless buttons, exploring different routes only to understand that what we wanted support with wasn’t supported by that experience.
“When you talk, you are only repeating what you already know. But if you listen, you may learn something new.”
The Dalai Lama’s wisdom applies surprisingly well to the bots vs. agents debate.
Reflect on that as you design a bot experience: You spend days outlining customer journeys, defining expected requests, mapping user flows, and building decision trees. But in doing so, how often do you learn something new from your customers, like their true pain points, their unexpected questions, or their genuine needs?
That’s the real difference between constructing a bot and implementing an agent. A bot just repeats what you programmed it to say, while a large language model (LLM)-based agent has the potential to listen to what your customers are asking, revealing insights that might otherwise be missed.
I get it — transitioning to LLM-based agents can be intimidating. You have to closely monitor outputs, ensuring the agent behaves as intended and avoids pitfalls. But these challenges are solvable. And the upside? Giving your customers a voice unlocks opportunities to shape your business strategy based on real customer conversations, not just hypotheses.
Agents changed everything. And by everything I also mean us as consumers.
Not only do I get quick answers to what I came for; while I’m interacting with the agent, I can ask hundreds of other things as well.
All of this and more in one session. Since the experience is much better, customers can interact much more. This obviously creates a lot of new requirements and requests for a company. How can I determine whether a session was a good interaction or a bad one?
Also, consider this hurdle: A simple question or request can be asked in multiple different ways — how would we know it’s all the same request without having to read through thousands of sessions?
When we built an exploration and monitoring tool for Agentforce agents, that listening approach was exactly what we had in mind. By surfacing meaningful requests from thousands of agent-customer conversations, we help you quickly spot trends and base business decisions on actual customer input.
Imagine an AI agent experience where customers can ask diverse questions in a single session, from account inquiries to technical issues and future service needs. While this enhances the customer experience, it presents a new challenge: how to effectively assess the quality of such multifaceted sessions and identify specific areas for improvement. For instance, if an agent successfully answers two questions but fails on a third, how do you evaluate that overall experience?
That’s where Agentforce Optimization comes in. It’s a powerful tool designed to give admins and agent builders in-depth visibility into AI agent interactions. Agentforce Optimization helps monitor performance, identify configuration gaps, and optimize agent effectiveness. It enhances the categorization, analysis, and utilization of interaction data by allowing users to define personalized labels and categories, thereby enabling targeted analytics and actionable insights.
In Agentforce Optimization, we’ve been dividing agent-user sessions into particular moments. Each moment is basically a user intent being answered by the agent. Within a single session, users can have one or multiple moments. Differentiating and evaluating each one separately helps us understand an agent’s performance much better, and also aggregates similar requests across multiple sessions.
Agentforce Optimization provides an interactive way to access and drill down into agent interactions, built upon several key concepts.
Agentforce Optimization is powered by the Agentforce Analytics Foundations Semantic Data Model (SDM), which acts as a unified source of truth for metrics, definitions, and calculations. This model enables seamless querying and integration with the UI. Once agent sessions are broken down into meaningful “moments” and evaluated for quality, these moments can be labeled and grouped using intelligent logic.
Out of the box, Agentforce Optimization applies a clustering algorithm to group similar moments and assign meaningful tag names. This allows you to identify patterns in customer requests, spot performance gaps, and prioritize areas for deeper investigation.
Looking ahead, we’re working on enabling standard and custom tags — a major enhancement planned for Q4. This feature will be built around Prompt Builder, introducing a specialized prompt template for Agentforce tagging. It will let you define your own tagging logic, backed by a schema that ensures structured and validated output from the LLM.
This feature will bring more tailored insights by supporting both predefined (enum) and dynamic (free-text or numeric) values. It will allow you to generate your own logic for tagging interactions, moments, and sessions, providing user-defined custom tags to extend functionality and analytics capabilities. Each data artifact will be tagged with both a tag name (for example, “Competitor mention,” “CSAT”) and a corresponding value (for example, “Walmark”, “High”).
Agentforce Optimization empowers administrators and agent builders with a range of capabilities to enhance agent performance and derive valuable business insights.
Imagine analyzing your customer interactions with Agentforce Optimization.
Agentforce Optimization is continuously evolving. Future enhancements include:
Ultimately, Agentforce Optimization ensures your AI agents deliver accurate, efficient, and high-quality responses, elevating customer experiences. It’s a partner in driving your AI agent’s success. By offering deep insights and actionable solutions, it empowers you to create agents that not only meet but also exceed customer expectations.
Start exploring today and take the first step toward building an AI agent that truly delivers.
Your customers deserve it — and so do you. Let’s make your AI agent the best it can be.
Ready to take the next step? Sign up for Agentforce today and unlock the full potential of agents for your organization.
As Salesforce Admins, we’ve dived headfirst into Agentforce. We’re building agents in Trailhead, attending workshops like Agentforce NOW, doing hands-on sessions with our Community Groups and at conferences like TDX, and coming up with ideas of how to put agents to work for our individual business need, even bringing that to fruition through events like […]
Today on the Salesforce Admins Podcast, we talk to John Demby, Director of Solution Engineering at Tableau. Join us as we chat about Pulse for Salesforce, Tableau Einstein, and how easy it is to get started. You should subscribe for the full episode, but here are a few takeaways from our conversation with John Demby. […]
Editor’s Note: This post was updated on April 3, 2025, with the latest information and resources. As Salesforce Admins, you’re well-versed in delivering Salesforce features to your users and stakeholders. Salesforce Foundations makes it even easier for you to get the most out of Salesforce by adding cross-department features built into your existing CRM. Salesforce […]
