The Short Answer
An enterprise knowledge base is typically built for humans, who can fill in gaps, identify outdated documents, and ask a colleague. An agent cannot do any of these three things. Therefore, training should focus less on adding content and more on removal, decision-making, and tagging.
The practical process involves five steps: reviewing coverage and validity, archiving conflicting and outdated records, reorganizing article structure, tagging for retrieval, and assigning ownership with SLAs for updates. The fifth step determines if the investment will last beyond six months.
How the retrieval layer consumes this content is explained in Grounding and RAG in Agentforce.
Step 1: Focused Review
Don't review the entire knowledge base. Take the list of scenarios the agent will handle and generate a list of real questions—from actual received inquiries, not imagination. For each question, check: Is there an article? When was it last updated? Who owns it? Is the answer still correct today?
The result yields four categories: covered and valid, covered but outdated, covered by several conflicting versions, and not covered at all. The third category is the most dangerous, as it causes the agent to give different answers to the same question.
The gap found in the fourth category isn't necessarily a problem—it becomes the writing backlog and, in the meantime, a list of topics that are routed to a human.
Step 2: Archive Before Writing
The hardest instruction for organizations to accept is to delete. But an outdated article remaining in the knowledge base causes more damage than a missing one: missing leads to escalation, outdated leads to a confidently incorrect answer.
A simple rule of thumb: any article without an owner that hasn't been updated beyond its defined period for its subject matter is removed from the index. It can remain in an archive for documentation purposes but is not accessible to the agent.
In a conflict between two versions, the decision rests with the content owner, not the technical team. This is a point that requires a business decision, and skipping it brings the problem back during testing phases.
Step 3: Article Structure
A good structure for an agent is also a good structure for a reader, so this isn't duplicate work. The title is phrased as a question or scenario using user-facing language, not internal jargon. The first paragraph provides a short, self-contained answer. Details are then presented in sections with subheadings.
Two rules directly impact retrieval quality: eligibility conditions and exceptions should be in a separate, clearly marked section, not interwoven into a sentence; and tables should remain small and self-contained, because a table cut off mid-way produces a partial answer.
What to avoid: long articles that cover five topics. It's better to split them into five focused articles—retrieval is more accurate and maintenance is easier.
Broader knowledge management principles in Service Cloud are detailed in Knowledge Management in Salesforce.
Step 4: Tagging and Audience Separation
Minimal tagging required in almost every organization: product or service line, market or country, language, target audience, approval status, and validity date. In a global organization, the lack of market and language tagging is the number one cause of incorrect answers—the agent will confidently pull a policy from a different country.
Audience separation is a security decision, not just a classification. Content written for agents sometimes contains discount margins, objection handling scripts, and competitive information. For a customer channel, the default should be a whitelist: only content explicitly marked as customer-approved is included.
A document that is mostly permissible but has one sensitive paragraph is not an edge case—it's common. The solution is to split the document, not mark the entire thing as internal.
Step 5: Ownership and Maintenance
This is the step that determines if the outcome will endure. For each knowledge area, assign a named owner, set a review frequency, and define what happens when an article passes its expiration date—automatic removal from the index, not an alert that no one reads.
The feedback mechanism is what makes maintenance effective. Any conversation that ended in escalation due to a missing source is added to the content gap list, and this is the best priority list for new writing. It's better to write five articles that were actually needed than fifty that seemed important.
Budget: Content maintenance is an ongoing expense, not a project. Organizations that treat it as a one-time task see a decline in accuracy within two quarters.
Index Entry Thresholds
| Criterion | Minimum Threshold | Why |
|---|---|---|
| Ownership | Named owner for each article | Without an owner, there's no one to update it |
| Validity | Within the defined review period for the subject | Prevents quoting canceled policies |
| Uniqueness | Single source for each topic | Prevents conflicting answers |
| Audience | Marked as internal or customer-approved | Prevents exposure of sensitive content |
| Structure | Question title and short answer in introduction | Improves retrieval accuracy |
Scenario: 1,400 Documents Reduced to 190
An IT service provider wanted to connect an agent to a SharePoint folder with 1,400 documents. Initial review showed that only about 30% had been updated in the last three years, and about 200 were drafts or working versions.
Instead of a sweeping cleanup project, the team mapped the 25 scenarios the agent was supposed to cover. From the existing repository, 190 relevant documents were found; of these, 40 were conflicting and required a decision from three subject owners. The selected documents were split into focused articles and tagged by product and audience.
The pilot launched with 190 items instead of 1,400, and answer accuracy was significantly higher than the initial attempt. The rest of the repository remained in the archive and was gradually incorporated based on the gap list accumulated from actual escalations.
Risks and Preventive Actions
| Risk | How it is detected | Preventive Action |
|---|---|---|
| Importing the entire repository | Low accuracy and difficulty in identifying the cause | Import by scenario, not by repository |
| Articles without owners | Silent obsolescence | Ownership as a condition for index entry |
| Conflicting versions | Different answers to the same question | Business decision and archiving |
| Mixing internal and external content | Exposure of sensitive information to the customer | Whitelist for external channel |
| Maintenance as a one-time project | Decline in accuracy within two quarters | Ongoing budgeting and SLA for updates |
Content Readiness Metrics
| Metric | Definition | Frequency |
|---|---|---|
| Scenario Coverage | Percentage of scenarios with an approved article | Monthly |
| Validity | Percentage of articles in the index within their validity period | Monthly |
| Topical Duplication | Number of topics with more than one source | Quarterly |
| Gaps from Escalations | Number of new topics required but unavailable | Weekly |
| Average Update Time | How long it takes to correct an article found to be incorrect | Monthly |
When support is needed in training the knowledge base and establishing the maintenance model, Agentforce and AI Service is the practical next step.
Knowledge Training Checklist
- ☐ A list of real questions from actual inquiries has been created
- ☐ Each question has been categorized: valid, outdated, conflicting, or missing
- ☐ Articles without owners have been archived or assigned owners
- ☐ Conflicts have been resolved by the content owner
- ☐ Articles are structured: question title, short answer, sections
- ☐ Eligibility conditions and exceptions are in a separate section
- ☐ Tagging for product, market, language, audience, and validity exists
- ☐ Internal content has been separated from customer-approved content
- ☐ Review frequency and automatic removal upon expiration have been set
- ☐ A mechanism converting escalations into a writing backlog has been established