What can actually be deflected (% by inquiry category)
Honest deflection is not "60% of tickets vanish by magic". It's the combination of three different rates, each with its own realistic ceiling:
| Inquiry category | % deflectable when well-designed | Comment |
|---|---|---|
| Operational FAQs (hours, policies, shipping) | 70-90% | The easiest terrain |
| Product inquiries (info, comparisons) | 50-70% | Depends on KB quality |
| Technical incident resolution | 30-50% | Only if the system escalates well |
| Complaints and grievances | <15% | Always better to escalate to a human |
| Returns and refunds | 20-40% | Legal risk — supervision is key |
The 3 mistakes that sink ROI
- Weak knowledge base. If your KB is small, outdated or inconsistent, the agent answers badly and customer trust drops fast. The KB has to be the first job, not the last.
- Badly designed human escalation. When the agent doesn't know, it has to pass to a human WITH full context. If it passes "the customer wants something" with no context, the support team hates the system within a week.
- No CSAT measurement. Without measuring satisfaction before and after, you don't know whether you saved hours or destroyed the experience. And a lost customer costs more than ten saved tickets.
Architecture: knowledge base + RAG + agent + human routing
A serious AI support system has four layers, not one. The mistake is selling "the agent" as if it were a single piece.
- Structured knowledge base (KB). Not "uploading the FAQ PDF"; it's clean, versioned content with metadata.
- RAG (Retrieval-Augmented Generation). The system that pulls what's relevant from your KB for each inquiry and passes it to the agent as context.
- Conversational agent. The LLM that responds, with a clear system prompt on tone, limits and when to escalate.
- Human routing layer. The system that decides when to hand off to a person and passes the summarized context so the customer doesn't have to repeat.
Recommended stack by company size
| Size | Stack | Monthly cost (approx.) |
|---|---|---|
| SMB / startup | Implementa AI Support ($89-169/mo) or Intercom Fin | $90-330/mo |
| Mid-market | Zendesk + AI agent + RAG over your KB | $1,700-5,500/mo |
| Enterprise | In-house platform or Salesforce + RAG over historical data | $9,000-28,000/mo |
How to measure CSAT and deflection without fooling yourself
Four critical metrics that need to be on the dashboard from day one:
- Real deflection rate — % of tickets resolved without human intervention. Trap: don't count a ticket that closed from customer silence as "deflected".
- CSAT pre and post — Net Promoter Score or equivalent, ideally by inquiry category.
- Average resolution time — before vs. after the agent, broken down by channel.
- Clean escalation rate — % of tickets passed to a human WITH context vs. without context. If the second number isn't zero, you have a design problem to fix.
Realistic timeline and cost
| Phase | Duration | Cost |
|---|---|---|
| SMB self-serve | 24-48h setup | $89-169/mo |
| SMB with custom (large KB) | 2-4 weeks | $3,500-9,000 setup + $220-550/mo |
| Mid-market integrated (Zendesk/Intercom) | 4-8 weeks | $9,000-28,000 setup + $1,700-5,500/mo |
| Enterprise with historical data | 8-16 weeks | From $28,000 |