Designing Knowledge for an Agentforce Agent
A common axiom in the world of AI is, “A model is only as good as its data.”
….
That’s great and all, but like - what does that actually mean?
So. Many. Models.
Your data model feeds into your large language model. Poor data—biased, noisy, or outdated—leads to flawed outputs, no matter how advanced the model is, because AI learns and generalizes from patterns in that data. For example, Agentforce in Salesforce struggles if Knowledge articles are vague or misaligned with customer needs, as Agentforce relies on that data to reason and respond. Good data is the foundation; the model just builds on it.
What this tends to mean for Agentforce (or other AI implementations) is businesses have yet another reason to improve the quality of upstream data. This prompts (pardon the pun) organizations to face challenges they’ve always had, long before “Agentic” became a buzzword - data modeling, process engineering, governance policies and the like. And why those all merit consideration on their own, one interesting consideration now is how to not just optimize your data - but how to do that specifically to for an Agentic audience.
And in this case (the puns are just coming naturally today, I guess), specifically how to develop your Salesforce Knowledge base such that a Service Agent is grounded as successfully as possible.
Analogy Alert!
In some ways - at least, in the context of this metaphor - an Agentforce Service Agent is more like a talented and energetic entry-level employee than a software program. When initially hired, there’s a ton of aptitude and ability to consume information, but little experience or intuition. To bring this new teammate up to speed, training and enablement would happen in a variety of different ways.
Now, human beings have had to come up to speed for a long time, and so we’ve gradually learned some things … about how we learn some things… so that that learning process is easier. We’ve even got idioms for some, like “pictures are worth a thousand words,” but you’re probably familiar with others as well:
Bullet points break up paragraphs (see what I did there?)
Whitespace helps make information visually easier to digest
The Rule of Three helps your point land (again, see what I did?)
These are best practices because they resonate with the particular wiring of human brains, but that is now betting the question: what are the best practices for Agents, considering the exact wiring of how LLMs actually work?
I’m So Glad You Asked
Large Language Models (LLMs) work by transforming vast amounts of text into a kind of superpower for predicting and generating human-like responses. They’re trained on this text soup using a process where they break words and phrases into tiny pieces (called tokens) and build a massive, searchable index—think of it as a super-smart filing system that tracks how words hang out together in real life.
This indexing lets them quickly retrieve patterns, so when a customer asks, “Why’s my order late?”, the LLM can piece together a reply based on what it’s learned about shipping delays, apologies, and polite phrasing—all without truly “understanding” the situation.
Enough Already. Give Me the Best Practices!
So then back to our specific question.
Based on the underlying architecture of generative AI, how exactly should I adapt my Knowledge Base to empower by Agentforce Service Agent and create an amazing customer experience?
Focus on Single-Topic Articles
Why it helps Agentforce: LLMs can process long documents, but breaking content into discrete, single-topic articles improves retrieval accuracy and reduces ambiguity. When each article addresses one specific issue or question, Agentforce can more precisely match customer queries to the right content.
How to implement: Review your existing articles and split multi-topic ones into separate pieces. For example, instead of one article covering “Troubleshooting Solar Panel Installation Errors,” create individual articles like “Fixing Solar Panel Wiring Issues” and “Resolving Solar Panel Power Output Problems.”
Salesforce tip: Use clear, descriptive titles and leverage the “Title” field as a primary indicator of the article’s focus.
Leverage Structured Fields for Clarity
Why it helps Agentforce: Structured data (e.g., fields like “Question,” “Answer,” “Issue,” “Resolution”) provides a clear framework that Agentforce can parse more effectively than unstructured text blobs. This helps the AI understand intent and context faster.
How to implement: Define custom record types and fields in Salesforce Knowledge to standardize article layouts. For FAQs, use “Question” and “Answer” fields; for troubleshooting, include “Issue,” “Environment,” and “Resolution.”
Salesforce tip: Keep customer-facing and internal content in separate fields (e.g., “Customer Response” vs. “Internal Notes”) so Agentforce can ground its replies in the appropriate context without mixing sensitive or irrelevant details.
Use Data Categories Hierarchically
Why it helps Agentforce: Well-organized data categories enable Agentforce to narrow down its search scope, improving response speed and relevance. Hierarchical categories also mirror real-world business logic, making it easier for the AI to align with your org’s structure.
How to implement: Create Data Category Groups like “Region” (e.g., North America > USA > California) and “Product” (e.g., Solar Panels > Residential > Model X). Assign articles to multiple categories as needed for flexibility.
Salesforce tip: Control visibility via profiles or permission sets to ensure Agentforce only pulls articles relevant to the user or customer context.
Incorporate Keywords and Synonyms
Why it helps Agentforce: Agentforce uses natural language processing to match queries to content. Including keywords and synonyms in your articles enhances its ability to find the right information, even if customers phrase questions differently.
How to implement: Add a “Keywords” or “Search Terms” field to each article and populate it with variations (e.g., “solar panel not working,” “solar unit failure,” “panel outage”). Include these in the article body too.
Salesforce tip: Configure search settings in Knowledge to prioritize results based on keyword relevance.
Keep Content Concise and Clear
Why it helps Agentforce: Overly verbose or complex articles can confuse the AI’s reasoning engine, leading to less accurate responses. Concise, straightforward language ensures Agentforce can quickly extract and apply the core information.
How to implement: Use short sentences, active voice, and avoid jargon unless it’s industry-standard and defined. Aim for a readability level that’s accessible to both humans and AI.
Salesforce tip: Use article templates to enforce consistent, concise formatting across your knowledge base.
Regularly Audit and Update Articles
Why it helps Agentforce: Outdated or incorrect information in your knowledge base leads to outdated or incorrect AI responses. Agentforce can’t update articles itself (yet), so maintaining freshness is on you.
How to implement: Set up a review schedule using the “Next Review Date” field and automate reminders with Salesforce Flow. Retire or revise articles that no longer apply.
Salesforce tip: Use the Lightning Knowledge Feedback package to let agents flag outdated content directly.
Ground in Trusted Sources
Why it helps Agentforce: Agentforce performs best when grounded in verified, high-quality data. Linking articles to CRM data, case histories, or external trusted sources (via Unified Knowledge) boosts its reliability.
How to implement: Integrate with Data Cloud for real-time customer insights or reference resolved cases in articles. Avoid speculative or unverified content.
Salesforce tip: Enable Unified Knowledge to pull in third-party sources if applicable, ensuring Agentforce has a broader, accurate knowledge pool.
Test and Refine with Real Queries
Why it helps Agentforce: Simulating customer interactions helps you identify gaps or weaknesses in your knowledge structure that might trip up Agentforce.
How to implement: Run test cases in a sandbox environment with common customer questions (e.g., “Why isn’t my solar panel charging?”). Check Agentforce’s responses and tweak article content or structure as needed.
Salesforce tip: Use Agent Builder to monitor and adjust Agentforce’s action plans based on test outcomes.
Take Baby Steps
We know this might sound like a huge lift, especially if you’ve got a large Knowledge base that doesn’t sound a whole lot like what we laid out.
Call us!
And also, start small by focusing on your most frequent customer queries—say, the top 10 issues your agents handle. Restructure those articles first, applying these practices, and then scale up. For example, if “Solar Panel Installation FAQs” is a high-traffic topic, break it into single-topic articles, assign relevant categories (e.g., “Product > Solar Panels” and “Region > USA”), and add keywords like “setup,” “install,” and “configuration.”
Also, regularly check Agentforce’s performance metrics (e.g., response accuracy, case deflection rate) to ensure your changes are paying off.
