Just discovered this. If your AI suggestions are wasting agents' time on high-priority or complex tickets, you can shut it off per ticket type.
Found it buried in the admin settings. You define a trigger for specific ticket types (like Priority = 1 or Category = "Billing Dispute") and set an action to disable the AI agent.
Example trigger setup:
```json
"trigger": {
"if": {
"all": [
{"field": "priority", "operator": "is", "value": "1"},
{"field": "status", "operator": "is", "value": "open"}
]
},
"then": [
{"action": "disable_ai_agent"}
]
}
```
Our deflection rate on P1 tickets dropped 5% after enabling AI suggestions, but resolution time increased by 20%. Agents were wasting cycles dismissing irrelevant suggestions. Turning it off for those tickets brought resolution time back down.
Check your deflection vs. resolution time metrics by ticket type. If the AI is adding friction, not value, for certain workflows, this is a necessary config.
cost per transaction is the only metric
Interesting find. The 20% resolution time increase is actually a textbook example of what happens when you inject a synchronous suggestion layer into a high-stakes workflow. The agent's cognitive overhead from dismissing irrelevant suggestions isn't just time wasted, it's context switching cost. That metric alone justifies the per-ticket-type toggle.
One thing I'd add: check whether the trigger evaluation itself adds any measurable latency. Freshdesk's rule engine is probably fine, but if you're running this on a high-volume queue with a complex "if" block more than a few fields deep, you could see a few milliseconds of overhead per ticket load. That's noise for most, but for P1 tickets where every second counts, it's worth validating with a before/after trace on the API response time at the agent's browser.
Also, I'd recommend logging the number of times the AI agent is suppressed per ticket type. If you see a suppression rate over 80% for a given category, the AI model is clearly misaligned with that workflow and you should feed that back to the training data pipeline rather than just hiding the symptom.
Every microsecond counts.