What it is
Deflection is the share of tickets AI closes entirely on its own, with no human involved. A modern support agent doesn't just answer FAQs — it takes actions (check an order, process a refund, change a plan) against your data and your APIs.
Where it came from
The wave was driven by Sierra (Bret Taylor, ex-Salesforce/OpenAI), Decagon, Cresta and Intercom Fin. Old decision-tree chatbots gave way to LLM agents with access to the knowledge base and the company's systems.
Why it took off
Customer numbers speak for themselves: Duolingo — 80% deflection, Substack — 90%+, Notion — 96.6% auto-resolved; ClassPass moved from "16 hours across 5 days" to full 24/7 and cut costs by 95%. The 2026 inflection: per Salesforce, 66% of service teams already run AI agents (up from 39% in 2025).
How to apply it now
- Start with your top 20 recurring ticket types — those drive most of the deflection.
- Wire in an up-to-date knowledge base (the agent is only as good as the freshness of your docs).
- Set up seamless escalation to a human with the full conversation context.
- Measure more than the resolution rate — track CSAT too, since savings shouldn't tank satisfaction.
What to watch out for
A bad knowledge base means confidently wrong answers. You need fallbacks, a log of "asked for a human", and guardrails on sensitive topics (money, cancellations, personal data). High deflection with no CSAT check is a trap.