Service Cloud

Knowledge Base Agent —
Self-Improving Support Documentation

Surfaces the right knowledge articles during every case resolution. Detects documentation gaps from unresolved case patterns. Drafts new articles from resolved case outcomes. Your knowledge base improves automatically with every case your team handles.

faster article retrieval

vs manual search

85%

KB hit rate

for L1/L2 cases

60%

gap detection

before customers notice

Auto

article updates

from resolved cases

Atlas Reasoning EngineSalesforce KnowledgeEinstein SearchPrompt BuilderSalesforce FlowEinstein Trust Layer
← Browse all agents

Typical build: 3-week sprint · Fixed price · Production-grade

Agentforce reasoning flow — retrieval + continuous improvement
CASECase IntriggerESemanticSearchArticleRetrievedResolvecaseUNRESUnresolvedcasesAPatternDetectionGAPGapFlagPBDraftArticleDraftReviewHumanReviewKBUpdatedKBUpdated123456RetrievalImprovement

Retrieval

Semantic search

Gap cycle

Weekly analysis

Publish gate

Human review

How it works

Every step, explained

This is the actual Agentforce configuration Kovil AI builds and deploys — not a diagram. Here is what runs inside every node.

1
Semantic Article Retrieval

Agent retrieves relevant articles using semantic search, not keyword matching

When a case enters the resolution flow, the Knowledge Base agent performs a semantic search of the Salesforce Knowledge Base using the case description and classification. Unlike keyword search (which misses articles that use different terminology for the same concept), semantic search understands meaning — a case about "my screen is frozen" retrieves articles about "display unresponsive" and "device not responding" as well. Search is filtered by product version, customer entitlement level, and article approval status. Deprecated articles are never surfaced; draft articles are only surfaced to internal agents, not to customers.

Semantic SearchEinstein SearchEntitlement-awareProduct Version FilteringKnowledge Status Filter
2
Contextual Summarisation

Relevant steps summarised in conversational language for the customer's context

The agent does not forward the knowledge article link. It summarises the relevant sections in language appropriate to the customer's apparent technical level — assessed from how they described the problem. For a technical user describing a specific error code, the agent provides precise technical steps. For a non-technical user describing general confusion, the agent provides simplified language with numbered steps and no jargon. The original article link is also included for customers who want the full documentation. All summaries are generated within the Einstein Trust Layer — no article content leaves Salesforce's data boundary.

Prompt BuilderContextual SummarisationTechnical Level AssessmentPlain LanguageEinstein Trust Layer
3
Gap Detection

Cases without matching KB articles are flagged as documentation gaps

For every case where the agent cannot find a relevant knowledge article above the confidence threshold, the case is tagged as a KB gap. The agent logs: the case type, the search query that returned no results, and the case outcome (how was it resolved if the agent or human found a way?). At the end of each week, the agent aggregates all KB gap flags and generates a gap report: which case types are missing documentation, how many cases were affected, and the suggested article topics based on the gap patterns. The report is posted to the configured Slack channel or emailed to the knowledge management team.

Gap DetectionConfidence ThresholdGap TaggingWeekly Gap ReportDocumentation Coverage
4
Pattern Analysis

Resolved case patterns analysed to identify where documentation should be updated

The agent runs weekly analysis on the previous week's resolved cases — both human-resolved and autonomously resolved. It looks for patterns: case types with unusually high resolution times (suggesting the existing KB article is unclear or incomplete), case types where agents frequently added manual notes that differ from the KB article (suggesting the article is outdated), and newly emerging case types that have no KB article but were successfully resolved (suggesting a new article should be created). Pattern analysis output is a prioritised list of KB update actions, ranked by case volume affected.

Pattern AnalysisResolution Time AnalysisArticle Freshness DetectionEmerging Issue DetectionUpdate Prioritisation
5
Article Draft Generation

Prompt Builder drafts new and updated articles from resolved case data

For identified gaps and outdated articles, Prompt Builder generates draft articles using the resolved case data as source material. For new articles, it uses the resolution steps that human agents applied to the unresolved cases as the procedure. For updates to existing articles, it generates a revised version with the corrections agents applied in practice added to the documented steps. All drafts follow your configured KB article template — title, summary, procedure steps, related articles section. Drafts are flagged as 'Draft — Review Required' and never published automatically.

Prompt BuilderArticle DraftResolution-based ContentKB TemplateDraft StatusReview Gate
6
Review, Approval & KB Update

Human review required before any article is published or updated

All AI-generated article drafts are queued in a dedicated Knowledge Management review flow. The knowledge manager sees: the draft article, the source cases it was generated from, the gap or outdatedness signal that triggered it, and a diff view for updates to existing articles. Review typically takes 10–15 minutes per article (reading and approving) versus 2–3 hours for an agent to write from scratch. Once approved, the article is published to the correct audience (internal, partner, or customer-facing). Article performance metrics (how often it is retrieved, CSAT on cases it resolved) feed back to the KB agent to improve future retrieval ranking.

Human Review GateKnowledge Manager WorkflowDiff ViewArticle PublicationPerformance Feedback LoopRetrieval Ranking
Tech stack

Every tool in the agent

Atlas Reasoning Engine

Gap + pattern detection

Runs weekly pattern analysis on resolved cases to identify documentation gaps, outdated articles, and emerging issue types.

Salesforce Knowledge

Article repository

The native Salesforce KB that the agent retrieves from, improves, and updates — maintaining a single source of truth for all support documentation.

Einstein Search

Semantic retrieval

Powers semantic article retrieval — understanding meaning rather than keywords so the right article surfaces for every case description.

Prompt Builder

Article drafting

Generates new article drafts and updated article versions from resolved case data — using your KB template and following your documentation standards.

Salesforce Flow

Review workflow

Manages the knowledge management review queue — notifying reviewers, tracking draft status, and publishing approved articles to the correct audience.

Einstein Trust Layer

Data security

Ensures case data used for article generation stays within Salesforce. No customer information leaves the platform during KB improvement processing.

Slack

Gap reporting

Delivers weekly KB gap reports and update priority lists to the knowledge management team in the configured Slack channel.

What we build

A 3-week sprint. Production ready.

Kovil AI scopes, builds, tests and deploys this Agentforce configuration end-to-end. You do not touch Agent Builder until your KB is surfacing the right articles and improving automatically.

  • Knowledge Base agent with semantic retrieval and contextual summarisation
  • Gap detection system with confidence threshold and weekly gap reporting
  • Resolved case pattern analysis running on weekly schedule
  • Prompt Builder article drafting templates following your KB format
  • Knowledge manager review workflow with diff view for updates
  • Slack integration for weekly gap reports and update priority lists
  • Article performance feedback loop for retrieval ranking improvement
Sprint timeline3 weeks
Week 1Retrieval + summarisation
  • Semantic search configuration, entitlement filtering, contextual summarisation Prompt Builder
Week 2Gap detection + patterns
  • Gap tagging logic, confidence threshold, pattern analysis scheduling
Week 3Draft + review flow + deploy
  • Article draft generation, knowledge manager workflow, Slack reporting, production deployment
FAQ

Common Questions

Will the agent publish articles automatically?

Never. All AI-generated drafts require human review and approval before publication. The agent creates drafts flagged as 'Draft — Review Required' and queues them for the knowledge manager. The human reviewer reads the draft, checks it against the source cases, and approves or edits before publishing. The AI removes the blank-page problem and reduces writing time from hours to minutes — but a human always makes the final call.

How does it know when an existing article is outdated?

The agent compares two signals: resolution notes added by agents (if agents are consistently adding the same correction that is not in the article, the article is likely outdated) and resolution time for cases that matched the article (if resolution time is significantly longer than average despite the article being retrieved, the article may not be solving the problem). Both signals feed into the weekly pattern analysis and surface the article for review.

Can it handle multiple languages?

Semantic search and gap detection work across all languages configured in your Salesforce Knowledge instance. Article draft generation via Prompt Builder supports English natively and can generate drafts in other languages if configured. Multi-language KB management — maintaining parity across language versions — can be added as a separate configuration during the sprint.

What is the typical KB coverage improvement after deployment?

Within 90 days, clients typically see a 30–50% improvement in KB article hit rate (percentage of cases where a relevant article is found), driven by: new articles filling identified gaps, updated articles that were previously unhelpful, and improved retrieval ranking. The hit rate improvement directly correlates with autonomous resolution rate — more relevant KB articles means more cases the resolution agent can handle without human intervention.

A knowledge base that improves itself.

Book a 30-minute discovery call. We'll audit your current KB coverage against your case type distribution and scope a 3-week implementation that closes your documentation gaps automatically.

Browse other agents

3-week sprint · Fixed-price · Production-grade · Post-launch support included