Glossary
Conversation AI vs Scripted Bot
Scripted bots follow a fixed decision tree — press 1, press 2, type the keyword. Conversation AI uses an LLM to understand intent in natural language and respond contextually, holding real back-and-forth dialogue.
The first generation of "chatbots" was scripted — keyword matchers and decision trees that broke the moment a user said anything off-script. Patients hated them, abandoned them, and learned to type "agent" to escape. Conversation AI is a different category: an LLM reads the patient's actual message, understands intent regardless of phrasing, and responds in natural language while pulling from a knowledge base for accuracy. The PatientCopilot Conversation AI framework is built on Goals, Prompts, and Actions — the agent has a clear objective, contextual instructions, and concrete tools it can use to act on the patient's behalf.
Why scripted bots fail
Real patient language is messy. "Can I push my Thursday thing?" "Do u guys take BCBS?" "Need to talk to someone about my back" — none of those map cleanly to a menu. Scripted bots either misroute, dead-end, or force the patient through 4 levels of menus they did not want. Patients hang up or ghost.
What Conversation AI does differently
It reads the message as-written, identifies the underlying intent ("reschedule appointment," "ask insurance question," "request callback"), pulls relevant practice context from the Knowledge Base, and either responds directly or takes an action (book, reschedule, send a form). When it cannot handle something — emotional context, complex insurance, anything ambiguous — it hands cleanly to a human with the full conversation context attached.
The Goals / Prompts / Actions framework
Goals: the high-level objective for the agent (e.g., "help the patient reschedule"). Prompts: the contextual instructions and constraints (tone, what to avoid, when to escalate). Actions: the concrete tools the agent can use — query KB, send SMS, book appointment, escalate to human. Together they keep behavior consistent without forcing rigid decision trees.
When to still use scripted logic
Hard compliance gates (opt-in/opt-out, HIPAA acknowledgment, age confirmation) should be deterministic, not AI-judged. Use scripted nodes for those, and let the LLM handle everything in between.
Related reading
Ready to Get Started?
Contact us today and take the first step. Free consultations available.