We deployed AI customer-support agents across 18 client environments in the last year — fintechs, D2C brands, SaaS companies, and a logistics platform. Here's what the numbers actually look like once the demo dust settles.
The deflection rate isn't 80%. It's 60–70%.
Vendor pitch decks love the 80% deflection number. In production, on real ticket distributions, we consistently see 60–70% — and that's the right ceiling to plan around. The remaining 30–40% are genuinely hard tickets where you want a human anyway.
Cost stack, per resolved ticket
- LLM inference: ₹2 – ₹6 per ticket (with smart caching and a small model fallback)
- Vector DB + retrieval: ₹0.30 – ₹0.80
- Tooling + observability: ₹0.50 fixed, plus engineering time amortised
Compared with ₹35 – ₹80 per ticket for a human agent, the unit economics are obvious — once you're past the build cost.
Where humans still win
Refund disputes, account closures, regulated industries, anything emotional. Don't fight this — design the handoff well, and CSAT actually goes up because the human gets pre-loaded context.
The honest payback period
For a brand doing 8,000+ tickets a month, payback is typically 4–6 months. Below that volume, you're better off with a templated chatbot and a small support team.