Navigating the AI Frontier: A Guide for Salesforce Admins on Building Reliable Agents

The rise of AI has brought with it a wave of excitement and, for many, a degree of uncertainty. For Salesforce Admins, the question isn’t if AI will impact their roles, but how they can leverage it effectively. Building reliable AI Agents is a game-changer, offering the potential for monumental gains in productivity and customer experience. But what’s the best approach, and what skills do #AwesomeAdmins need to master to excel in this new era?

We’ve been building AI Agents for Agentforce since it launched, and we’ve written a lot about the approach which we are using. This approach has been adopted by Salesforce and they can be found speaking and writing about it.  Here is a session from London TDX on Salesforce+.

Is It Even an Agent? Differentiating AI Tools

Not everything that feels like AI is a full-fledged Agent. It’s crucial to understand the spectrum of AI capabilities within Salesforce. AI isn’t something completely new. It’s a different way of engaging with your data.

Workflow: It’s not an AI Agent, but can be delivered more simply using automation. Some of our earliest AI Agents should have been a workflow, but we built them to get practice on well-understood use cases with a well-defined business process, e.g. booking PTO.

Workflow with AI: At its simplest, AI can enhance existing workflows. This might involve a prompt template behind a button that, when clicked, generates an email or performs a few simple actions. Think of it as “almost as an AI Agentic workflow”.

True AI Agent: A true AI Agent possesses the ability to reason, plan actions, and execute them autonomously based on natural language input and its instructions. These AI Agents are designed to handle complex tasks and can even make decisions, like a sales coach analyzing a customer transcript and providing personalized coaching.

The key distinction lies in the AI Agent’s capacity for reasoning and autonomous action. Before diving into AI Agent development, assess whether a simpler AI-enhanced workflow or a full-blown AI Agent is truly needed for your use case.

The Elements Agent Finder can now look at your business processes and identify agent opportunities, and even categorize them, and write the business requirement. Watch Xavery Lisinski, Chief Product Officer, Elements.cloud, demo Agent Finder, live from the campground at TDX25.

Scoping for Reliability: Agents and Topics

Building an AI Agent isn’t difficult. Building a reliable AI Agent is. The foundation of a reliable AI Agent lies in precise scoping. Salesforce’s Agentforce defines an Agent as a “domain” e.g., “Customer Support”, which contains “Topics” that are the “Jobs To Be Done” (JTBD). A JTBD is more granular, e.g. “Take inbound support request, triage is based on complexity, find potential solution by looking at knowledge base, and then create a case and route it.”

Start Narrow, Think Big: When implementing AI Agents, begin with a narrowly scoped use case, ideally with just one Topic. This approach simplifies experimentation, accelerates deployment, and makes data governance and monitoring more transparent, ultimately lowering technical, business, regulatory, and reputational risks. Once you’ve successfully deployed the initial AI Agent and gained experience, you can gradually extend its capabilities by adding more Topics and Actions. Read this article on architect agents and topics.

Designing Topics: Instructions and Actions

The core of a Topic is its Instructions and Actions, which guide the Agent’s behavior. “An AI Agent is equipped with resources-data, automated workflows, and even guardrails that tell it when to pass things off to a human,” and these are defined in the Topic.

Process-Led Design is Key: As Jack Lavous, CIO at Elements.cloud, emphasizes, “Implementing AI Agents is much like any other technology-enabled business transformation. Time spent planning will help eliminate painful rework and get you a working AI Agent far faster.” Skipping the design phase, particularly mapping out the AI Agent design, will lead to frustrating iterations and a lack of confidence in deploying the AI Agent into production.

Engaging Stakeholders accelerates deployment: Using the diagram to design the AI Agent ensures that the AI Agent is delivering what the stakeholders need. “It is far easier to explain to them what the AI Agent does by showing the diagram vs a long list of instructions.” This approach gives them the confidence to agree to deploy the AI Agent.

Explicit Instructions: Think of the Agent as an “eager, fresh-faced new hire”. It has no inherent context or common sense, so your instructions must be far more explicit than they would be for a human. Design the Topic in a diagram that describes the overall flow that then generates explicit, unambiguous instructions. The diagram is the change log as the AI Agent iterates. I built an AI Agent that calculates what to pay each band member after a gig: Agents read every word literally. The AI Agent calculated round-trip mileage because it was told to “calculate the mileage,” not just the distance from the band member’s home to the gig, which was the human’s implicit assumption. This underscores the need for clear, crisp, and unambiguous natural language in your instructions. You are programming an AI Agent with natural language.

Single Instruction for Order: A crucial discovery is that to get consistent and accurate results, it’s best to combine all instructions and guardrails into a single instruction within the Topic. This maintains control over the order in which the AI Agent applies them, dramatically reducing testing time.

Actions as the Agent’s Toolbox: Actions are how the Agent interacts with Salesforce and external services (via Data Cloud) to achieve its goals. These can be existing Apex, Flow, or Prompt Templates, or newly created ones. Actions should be narrowly scoped for predictable behavior and ideally reuse existing metadata. Clear documentation of these actions is vital, as the Agent “is essentially reading your documentation to understand the capabilities of an action”. 

Crafting Effective Prompts: A Design Approach: Prompt templates are a type of Action and, much like instructions, require a thoughtful, structured approach. A prompt template essentially is instructions. Consider the logical sequence of steps. If you think about those different steps, then the instructions naturally drop out of it. This process-led thinking helps ensure reliability, just as it does for designing the broader AI Agent.

Here is a step-by-step guide using the Elements.cloud FREE Agent Designer capability.

Salesforce Admin Skills: Bridging the Gap to Agentblazer

Salesforce Admins possess many foundational skills essential for building successful AI Agents, making them prime candidates to become “Agentblazers”.

Existing Strengths: Admins already understand their Salesforce Org, including data structures, objects, and fields. They are proficient in building Flows for automation. This understanding of metadata is critical because AI Agents leverage existing functionality.

Skills to Enhance: While the core is there, Admins need to enhance their skills, particularly in:

• Business Analysis: Understanding the business process thoroughly is paramount. This includes the ability to translate natural language business needs into precise, explicit instructions and action descriptions for the AI Agent.
• Agent Design Thinking: Approaching AI Agent development with a structured design mindset is vital for building reliable AI Agents. Instructions are the programming language for AI Agents.
• Data Governance and Quality: Since “Poor data is junk food for AI”, a strong grasp of data quality and transparent data flow is more critical than ever.
• Prompt Engineering: Learning how to write effective and unambiguous prompt templates is a new, crucial skill.

Tools: There are tools in the Salesforce Platform that Admins are already familiar with:

• Salesforce Setup for managing objects and fields
• Flow Building for building and managing Flows
• Change Sets for promoting metadata 

But there are new tools to master:

• Process Mapping / Configuration Mining to understand where agents can be deployed
• Agent Designer to be able to design, document, and govern Topics, Instructions, and Actions
• Metadata Dictionary and Dependencies to understand the impact of changes and document metadata
• DevOps tooling to make the iteration of agents more efficient than Change Sets

By embracing a process-led approach, focusing on meticulous design, and leveraging their existing Salesforce expertise while honing new AI-specific skills, #AwesomeAdmins can confidently lead their organizations into the age of AI Agents, delivering transformative results and cementing their role as indispensable strategic assets.