A SaaS company we worked with rolled out a support agent. Within three weeks, CSAT dropped seven points. The agent was answering 60% of incoming tickets correctly. The problem wasn't the 60%. It was the other 40% — every customer who got the wrong answer also got the impression that the company didn't care enough to put a human on it.
Support agents that ship are not measured on deflection rate. They're measured on how well they respect the moments a customer wants a human. The escalation contract is the design.
What customers actually want
Customers contacting support are usually in one of three modes:
- Quick answer — "Where do I update my billing email?"
- Diagnosis — "Why isn't my export working?"
- Recovery — "I was charged twice and I'm upset."
Mode 1 is where agents shine. Mode 2 is where they help if they have the right data access. Mode 3 is where agents ruin reputations if they try to handle it. The distinction is not about question complexity. It's about emotional load.
A support agent that's smart about emotional load — that recognises a frustrated customer in tweet two and routes immediately to a human — earns trust. One that doesn't makes every interaction worse than the human-only baseline.
The escalation contract
Every support agent we ship has an explicit escalation contract: the agent answers if and only if it's confident, the topic is in scope, and the customer's tone is neutral. Otherwise it hands off, with context.
Concretely:
- Confidence threshold — the agent self-reports a confidence score. Below threshold (we typically start at 0.85), it routes to a human.
- In-scope check — every agent ships with an explicit list of topics it handles. Anything outside the list routes immediately, no try-and-see.
- Tone gate — sentiment cues like "this is the third time", "I want a refund", all-caps, exclamation points → automatic escalation.
- Context handoff — when escalating, the agent posts a one-paragraph summary of the conversation into the human agent's queue. The customer never has to repeat themselves.
That last item is the difference between agents customers tolerate and agents they hate. Repeating yourself to a human after struggling with a bot is the worst part of the experience. Don't do it.
Tone is a feature
Most support agents sound like product manuals. The good ones sound like a competent teammate. The voice difference is small in characters and big in CSAT.
Compare:
- Manual voice: "To update your billing email, navigate to Settings > Account > Billing and click 'Edit email address'."
- Teammate voice: "You can update it in Settings → Account → Billing. Hit 'Edit email' and you're done. Want me to walk you through it?"
The second one is two characters longer and reads like a person who's done this before. It also opens a follow-up loop, which catches the customers who actually need diagnosis but started with a how-to question.
The voice has to be calibrated. Once. Everywhere. With eval cases that catch drift. The voice eval is doing more work than most teams realise.
Where the deflection number goes wrong
Vendors love the deflection rate. It's the number on the slide. But deflection rate without satisfaction-rate context is a vanity metric. We've seen agents at 70% deflection with 4.2 CSAT and we've seen them at 35% deflection with 4.8 CSAT. The 35% one is doing better business.
The question is: of the tickets the agent handled end-to-end, how many customers were satisfied? Of the tickets it escalated, how many customers felt the handoff was smooth? Those are the two metrics. Deflection-rate-as-headline is a trap.
How to start
- Pick one ticket type. The most common, lowest-emotional-load category. Something like "how do I do X in the product?"
- Build the agent with a strict escalation contract.
- Eval it weekly against a curated set of 100-200 real tickets, scored by your own support team.
- Roll out at 10% of traffic for two weeks. Measure CSAT, escalation rate, repeat-contact rate.
- If CSAT holds, expand the in-scope list one topic at a time. If it drops, find out why before adding scope.
The teams that move fast and break this end up with a deflection rate they can't show their CEO without context. The ones that move with discipline end up with a real defection: cost down, CSAT up, support team focused on the harder questions.
Close
Support agents work when they're polite about their limits. The escalation contract is the design that makes them polite. Build it first, measure both deflection and satisfaction, and resist the temptation to widen scope before the eval data says you should.
Related reading
- The agent maturity curve — where support agents typically sit.
- Agents in healthcare: scribe yes, nurse no — the line drawn for clinical agents, restated for support.
- Your AI agent should plan like a kitchen brigade — escalation as a station-level decision.
We build AI-enabled software and help businesses put AI to work. If you're shipping a support agent, we'd love to hear about it. Get in touch.