Agents waste 40% of their time searching for solutions across email threads, shared drives and tribal knowledge — and when a senior agent leaves, their know-how leaves with them. Static KBs go stale within weeks because nobody updates them manually.

Utilyx models the knowledge base as a living workflow: the AI generates articles from resolved tickets, flags them for review, and serves them through semantic search so agents and users find answers in seconds.

The workflow is structured as a decision tree with 3 branches (Auto-publish, Review-required, Reject) and 6 steps end-to-end. Each step is executed by a named actor (Resolution trigger, AI Agent, KB reviewer, ai-agent semantic search, End user, KB owner) and every action is timestamped.

• Step 1 (Start — Resolution trigger): when an incident or request is closed, the workflow extracts the resolution text, ticket category, affected CI and diagnostic timeline. Fields: source_ticket_id, resolution_text, category, ci_affected, language.

• Step 2 (AI Agent): the ai-agent generates a draft KB article from the resolution — structured as problem statement, root cause, resolution steps, verification — and assigns a confidence score. It also detects duplicates against existing KB and merges if similarity > 80%.

• Step 3 (Routing): three branches — Branch A (confidence > 0.85, common category): auto-publish to KB with 'draft' visibility for the KB owner; Branch B (confidence 0.6-0.85, or sensitive category like security): routed to KB reviewer for validation; Branch C (confidence < 0.6, or duplicate): rejected, resolution text kept in ticket archive only.

• Step 4B (KB reviewer): the reviewer edits the draft, adds tags, screenshots and cross-references, then publishes. Review SLA is 48h; overdue drafts escalate to the KB owner.

• Step 5 (ai-agent semantic search): when an agent or end user searches the KB, the ai-agent performs semantic search (not keyword) across all articles in 7 languages, returning the top 3 matches with relevance scores. If a user-facing query matches a KB article above threshold, the self-service portal surfaces it before a ticket is created (deflection).

• Step 6 (KB owner): monthly, the KB owner runs a freshness review — the AI flags articles not accessed in 90 days for archive, and articles with declining success rates for rewrite.

Teams that adopt Utilyx report agent search time cut by 40%, 50% of new KB articles auto-generated from resolved tickets without manual writing, and 35% ticket deflection on the self-service portal because semantic search surfaces the answer before the ticket is created.

The 7-language semantic search lets a single KB serve a multilingual user base, the confidence-based routing balances auto-publish speed with review control, and the freshness review keeps the KB relevant — no stale articles polluting search results.