Support AI Agents: 57% Fewer Tickets and an 18-Point CSAT Gain
Deflection is easy to fake and expensive to get wrong. Routing to specialised agents — and knowing when to hand to a human — did more for satisfaction than any single model choice.
- Client
- Customer support organisation
- Industry
- SaaS support
57%
Fewer inbound tickets
+18
CSAT improvement
<5s
Response time
~45%
Lower operational cost
The problem
Support volume was growing faster than the team, and the shape of it was frustrating: a large proportion of inbound tickets were questions the documentation already answered, arriving because finding the answer was harder than asking a person.
The team had looked at generic chatbots and rejected them, correctly. A bot that deflects a ticket by failing to answer it does not reduce cost; it moves the cost to a second, angrier contact and takes the satisfaction score with it.
The design
Rather than one assistant that knows everything badly, we built specialised agents for the domains that actually behave differently: billing, onboarding and technical troubleshooting. Each has its own grounding corpus, its own tone constraints and its own integrations.
In front of them sits a routing layer that classifies intent and hands off, including the most important route — to a human, with the conversation context attached, when the question is outside what the agents should be answering.
Grounding and speed
Every agent answers from retrieval over the live documentation and knowledge base, using embeddings and vector search, with citations attached to the response. Customers can see where the answer came from, which both builds trust and gives them somewhere to go when the answer is nearly right.
Responses land in under five seconds. That number is a product requirement, not a technical one: past roughly that point, users abandon the chat and open a ticket anyway, and every architectural decision was made to stay inside it.
The part that needed care
The escalation threshold was tuned harder than anything else in the system. Set it too eager and the agents hand off trivia, saving nothing. Set it too reluctant and customers get stonewalled by a bot, which is the failure mode that destroys satisfaction scores.
We instrumented both directions — unnecessary escalations and conversations that should have escalated and did not — and moved the threshold on the evidence rather than on instinct.
Billing was deliberately given the tightest constraints of the three. Being wrong about an onboarding step costs a minute; being wrong about someone's invoice costs the relationship.
Results
Inbound tickets fell 57%, and — the number that matters more — CSAT rose 18 points over the same period. Deflection that people resent shows up as satisfaction falling while ticket count falls, and that is not what happened here.
Operational cost dropped roughly 45%, driven by the ticket reduction rather than by reducing the team's ability to handle the hard cases. Agents handle the repetitive body of work; people handle what needs judgement, with context already gathered.
What we'd flag
These results depended on documentation that was reasonably current. Retrieval-grounded support agents are a magnifying glass on your knowledge base — where it was stale, the agents were confidently stale, and fixing the source was the only remedy.
We would also resist any pressure to push deflection higher from here. The remaining tickets are disproportionately the ones where a person should be involved, and squeezing them would trade a real satisfaction gain for a marginal cost one.
