Large language models are good at understanding what people mean. They are less good at following a twelve-step eligibility process the same way every time. For many business processes that inconsistency is unacceptable, and it is exactly the gap Agent Script is designed to close.
Reasoning and rules in the same agent
Agent Script lets you combine LLM reasoning with deterministic logic inside Agentforce Builder. The agent still converses naturally, interprets intent and handles unexpected phrasing, while the steps you script (checks, conditions, required actions and their order) run exactly as written.
Where scripted logic pays off
Use it wherever a mistake has real consequences: identity verification before account changes, eligibility and compliance checks, refunds above a threshold, or any process an auditor may ask you to explain. If the business would write the rule down in a procedure manual, it probably belongs in script.
Where to leave room for reasoning
Not everything should be scripted. Answering open questions from knowledge, summarizing a case history or drafting a reply all benefit from the model's flexibility. Over-scripting makes agents brittle and frustrating to talk to.
A side benefit: speed
Because a scripted path doesn't require the agent to re-evaluate every topic and action at every turn, structured flows can also reduce latency, which matters most in voice conversations.
How we approach it
We map each agent job into three layers: what the agent may decide, what it must always do, and when it must hand off to a person. The middle layer becomes Agent Script, backed by scenarios in Agentforce Testing Center, so every change to the rules is verified before it reaches customers.