Order status & tracking
"Where is my order?" "When will it ship?" "Why is my tracking not updating?", Yiannis pulls real-time status from your order DB and explains it in your brand voice.
Operational tier · Customer support
Yiannis, your AI tier-1 support agent.
Yiannis reads incoming support tickets, resolves the routine 60–70% of tier-1 volume autonomously, and escalates the rest to a human with full context attached. Your support team stops drowning in "where is my order" and starts handling the tickets that actually need them.
What does Yiannis do?
Yiannis is an AI customer support agent. He reads inbound tickets from email, Zendesk, or Intercom, classifies them, retrieves the right answer from your knowledge base, resolves tier-1 questions autonomously (order status, account access, basic troubleshooting, refund eligibility, FAQ), and escalates everything else to a human with full context attached, the ticket, his reasoning, the customer's history, and a suggested response. Built on Claude Sonnet. Operational tier. 95% eval pass rate. Typical production performance: 60–70% autonomous resolution rate on tier-1 volume after 30 days of tuning. Cost per resolution: roughly €0.05–€0.15 after build cost.
theagency47 · Updated May 2026 The spec sheet
How Yiannis actually works.
| Job-to-be-done | Resolve tier-1 support tickets autonomously, escalate the rest with full context |
| Tier | Operational (departmental flow) |
| Underlying model | Anthropic Claude Sonnet (Haiku for simple classification, Sonnet for resolution) |
| Trigger | Event-driven on incoming ticket (webhook from Zendesk/Intercom or new email) |
| Inputs | Ticket content, customer profile, order/account history, knowledge base, brand voice document |
| Tools available | Knowledge-base retrieval, customer database read, order database read, refund tool (bounded amount), Slack escalation, draft response |
| Autonomous decisions | Ticket classification, knowledge-base citation, response drafting, refund within policy (typically < €50 or category-bounded), escalation when uncertain |
| Escalation rules | No matching KB article → human · Confidence score < 0.7 → human · Sentiment negative + repeat customer → human · Refund > policy → human · Legal/compliance flag detected → human |
| KPIs measured | Autonomous resolution rate, time to first response, time to resolution, customer-satisfaction (CSAT) score, escalation accuracy, cost per ticket |
| Eval suite | 30 test cases (60% common-case, 25% edge, 15% adversarial including prompt-injection). Pass rate 95% as of May 2026. |
| Audit trail | Every action logged with KB citations used, confidence score, and decision reasoning |
What he handles autonomously
The repetitive 60–70% of your tier-1 volume.
Account access
Password reset triggers, login issue diagnostics, account merge requests for duplicate signups. Handled inside policy with audit logging.
Basic troubleshooting
Standard product-knowledge questions resolved from the KB. "How do I export X?" "Does feature Y support Z?" "Why is the dashboard showing N?"
Refunds within policy
Yiannis can issue bounded refunds (typical limit: under €50, certain categories). Anything outside policy escalates with refund recommendation.
FAQ resolution
"Do you ship internationally?" "What is your return policy?" "Is there a free trial?", all resolved from the KB with full citation.
Smart escalation
When something is outside scope, Yiannis writes a perfect handoff message, ticket summary, customer context, reasoning trace, suggested next steps. Human picks up where he left off.
Integrations
Standard integration stack.
- Helpdesk: Zendesk (preferred), Intercom, Front, Help Scout, Gorgias, Freshdesk, Re:amaze
- Email: Gmail (Google Workspace), Outlook (Microsoft 365), when no helpdesk in use
- Knowledge base: Notion, Confluence, Document360, Helpjuice, GitBook, or custom CMS
- Customer data: Stripe (subscriptions/billing), order DB (Shopify, custom), CRM (HubSpot, Salesforce)
- Internal escalation: Slack (preferred), Microsoft Teams, in-helpdesk assignment
- Multi-channel: WhatsApp Business API, Instagram DM (via Meta), Facebook Messenger (via Meta)
Configuration deep-dive
The two design choices that make Yiannis reliable.
1. Strict retrieval-augmented generation
Every claim Yiannis makes must be traceable to a source in your knowledge base. If a customer asks "what is your return policy?" and the answer is not in the KB, Yiannis does not guess from training data, he escalates. This is the single biggest design choice in his architecture. It eliminates the hallucination class of failures entirely.
Practical consequence: your knowledge base becomes the single source of truth. We document this dependency in the engagement and build a 5-day KB ingestion phase into every deployment.
2. Bounded action authority
Yiannis can read most things but write very few. Specifically: he can draft a response (always reviewed in first 14 days, sampled thereafter), issue a refund within a configurable limit, update a ticket status. He cannot change customer account details, modify billing terms, contact other customers, send marketing communications, or alter the knowledge base itself (KB updates are a separate human-approved flow).
Practical consequence: the worst-case bug is a confused response, not a corrupted database or a leaked customer record.
FAQ
Questions about Yiannis.
What does Yiannis do?
AI customer support agent. Reads tickets, retrieves answers from KB, resolves tier-1 (order, account, troubleshooting, FAQ, bounded refunds) autonomously, escalates the rest with full context. Operational tier, Claude Sonnet, 95% eval.
How much support volume can he handle autonomously?
60–70% of tier-1 volume at production maturity (30 days of tuning). The 30–40% remainder escalates to humans on harder, more interesting tickets. Net: support team handles less volume but more meaningful tickets, faster overall.
Will Yiannis hallucinate?
No, when configured correctly. Strict retrieval-augmented generation: every claim must be sourced in the KB. If not in KB, escalates. Eval suite specifically tests for hallucination with adversarial inputs.
How is Yiannis different from Intercom Fin or Zendesk AI?
Those are SaaS chatbots configured in-product. Yiannis is custom-built for you: trained on full KB (not just site crawl), integrates with internal systems (order DB, billing), tuned escalation policy, owned by you. Per-resolution cost ~€0.05–€0.15 vs Fin's ~$40.
How much does customizing Yiannis cost?
Yiannis template customized to your business: deployable as Spark (€2,500, 14 days) or part of Workforce engagement. Operating cost €50–€300/month for Claude API at typical SMB volume.
Will Yiannis replace our support team?
No. He reduces tier-1 routine volume by 60–70%, freeing the team to handle complex tickets, build deeper customer relationships, and improve the KB. Most clients keep team size flat and grow business with it.
What about angry customers and edge cases?
Both escalate. Sentiment-detection flags negative emotion → human. Out-of-scope topics → human. Ambiguous tickets → human with confidence score and recommendation. Yiannis handles the easy 60–70%; humans handle the hard 30–40%.
Does Yiannis handle multiple languages?
Yes, given a KB in the target language. Standard configuration handles English; multi-language support adds 1 day of setup per additional language (typically +€500 each).