Am I correct in thinking you are talking about a single-turn inference using a Data Action, rather than a multi-turn conversation with an LLM?
In the past I have created a multi-turn integration to an in-house custom conversational agent using Genesys's Bot Connector, and performed bespoke classification of emails using a Data Action (backed by an LLM).
Response Times
In both instances the response times varied greatly based on the task, model & guardrails. So it would be hard to give a precise answer to this.
Security considerations
Without knowing your precise use-case this is hard to say, but here are some of our considerations:
- Model-based content safety parameters - the model we are using has the ability to tweak the sensitivity of its own guardrails around the response
- Guardrails - We use an external system on both inbound and outbound messages to flag attends to manipulate the output etc
- Online evaluations - These are passive evaluations run against the output, which produce both metrics and LLM-as-a-judge based metrics
- 'Banning' feature - Multi-turn jailbreaking can be difficult to prevent, so we limit this risk by tracking it across turns/interactions.
Customer experience improvements
This is entirely dependant on your use-case, so it's hard to offer any suggestions - other than frequently reviewing interactions .
Hopefully that was of some help.
------------------------------
Lucas Woodward
Winner of Orchestrator of the Year, Developer (2025)
LinkedIn -
https://www.linkedin.com/in/lucas-woodward-the-devNewsletter -
https://makingchatbots.com------------------------------