Use case
Customer support agents
Tier-1 support spends most of its time classifying, routing and rewriting context. A Dezifi support agent collapses that work into a single pass — and escalates only what really needs a human.
What you'll learn
- How to design a support agent for classification plus drafting
- Where to draw the line between auto-resolve and human handoff
- Which Zendesk / Freshdesk hooks to wire up
- How to roll this out without surprising customers
The agent design
Support agents need accurate intent classification, conversational drafting, and tight escalation rules. Tone control is the hardest part.
- 1
LLM choice
Claude Sonnet or GPT-4o for draft quality. For pure classification, route to a cheaper model — token costs scale fast at ticket volume. - 2
Tools
Zendesk or Freshdesk (read ticket, post internal note, change status, assign), the product knowledge base for RAG, Jira for bug escalation, Slack for routing alerts. - 3
Guardrails
PII redaction on outbound drafts. Toxicity and prompt-injection checks on customer-supplied content. Refund or credit replies always require agent approval. - 4
Workflow shape
Trigger on new ticket. Branch on classified intent (billing, bug, how-to, churn risk). For churn or refund, escalate immediately. For how-to, draft a reply grounded in the KB and queue for human send-off.
Tools to connect
- Zendesk or Freshdesk — primary helpdesk. Read tickets, write internal notes, change status.
- Knowledge base — index help center articles, internal runbooks and product docs for RAG grounding.
- Jira — auto-create bug tickets when the support agent detects a regression.
- Slack — route high-priority or churn-risk cases to a CSM channel.
How to set this up in Dezifi
- 1
Connect Zendesk
Integrations → Zendesk → OAuth. Pick the brand and groups the agent should monitor. - 2
Index your help center
Knowledge → New Source → URL crawl. Point at your help center root. Schedule a daily re-index. - 3
Create the agent
New Agent → "Support Triage Agent". Attach Zendesk + the help center KB + Jira. Use Claude Sonnet as default model. - 4
Apply guardrails
Standard profile. Add a refund-detection rule that routes to approval. Add an internal-only mode initially — the agent posts drafts as internal notes, never sends to the customer. - 5
Pilot internal-only for one week
Watch Monitor traces. Check classification accuracy and draft quality in Eval. Once the team trusts the drafts, flip the policy to allow public replies for low-risk intents.
Frequently asked questions
- Will the agent reply to customers directly?
- Only when you explicitly allow it. Most teams start with internal-note mode — the agent drafts, a human sends. Once trust is high, you can enable auto-reply for narrow intents like password resets or shipping status.
- How does it handle tickets in multiple languages?
- Modern LLMs handle 30+ languages natively. The agent detects the customer language, retrieves KB content, and drafts in the same language. Set a language allowlist in the policy if you want to restrict.
- What happens when the KB doesn't contain the answer?
- The agent flags low retrieval confidence and escalates to a human. It never invents answers — guardrails block ungrounded factual claims.
- Can it create Jira bugs automatically?
- Yes. When intent is classified as "bug" with high confidence, the workflow creates a Jira ticket, links it to the Zendesk ticket, and notifies engineering in Slack.