Amazon Connect voice AI · ComparisonAmazon Connect 语音 AI · 方案对比
Four Amazon Connect paths in my clinic demo我的诊所演示里的四条 Amazon Connect 路径
All four keys use the same Amazon Connect phone line and the same Polly voice. The real difference is where the brain of the dialog lives: on my own server, in Lex rules, in a Bedrock LLM with a knowledge base, or in a Connect AI agent built in Agentic CX.四个按键用的是同一条 Amazon Connect 电话线路，声音也都是 Polly。真正的区别在于对话的“大脑”放在哪里：在我自己的服务器上、在 Lex 的规则里、在 Bedrock 的大模型加知识库里，还是在用 Agentic CX 搭建的 Connect AI agent 里。
Side-by-side comparison对比总表
| Key 1按 1Bring my own agent接入自研 agent | Key 2按 2Self-service booking自助预约 | Key 4按 4Bedrock knowledge baseBedrock 知识库问答 | Key 5按 5Agentic CXAgentic CX | |
|---|---|---|---|---|
| In one line一句话说明 | Connect handles the call and speech; my own agent runs the whole dialogConnect 只做电话和语音，对话全交给我自己的 agent | Lex runs a rule-based transaction flow用 Lex 做规则化的事务流程 | An AWS-managed LLM answers from a knowledge base用 AWS 托管的大模型 + 知识库回答问题 | A Connect AI agent built in the Agentic CX designer, no code在 Agentic CX 设计器里搭建的 Connect AI agent，无代码 |
| Speech recognition (ASR)语音识别（ASR） | Lex V2 | Lex V2 | Lex V2 | Amazon Connect agentic voice |
| Text-to-speech (TTS)语音合成（TTS） | Polly | Polly | Polly | Polly (inherited from the main flow)Polly（继承主流程的声音设置） |
| Who understands intent谁理解意图 | My agent (5-class intent classifier)我的 agent（5 类意图分类） | Lex NLU (intents + slots)Lex NLU（intent + slot） | The prompt in the Bedrock Flow; Lex has no intents here and only turns speech into textBedrock Flow 里的 prompt；这里 Lex 没有设计意图，只负责语音转文字 | The Agentic CX AI agent (LLM)Agentic CX 的 AI agent（大模型） |
| Where answers come from答案从哪里来 | My RAG: pgvector + reranker + Redis cache我的 RAG：pgvector + reranker + Redis 缓存 | Booking data in DynamoDBDynamoDB 里的预约数据 | Bedrock Knowledge Base | Q&A knowledge base; small talk on Nova 2 LiteQ&A 知识库，闲聊用 Nova 2 Lite |
| Middle layer中间层 | Lambda adapter (Python) → HTTP → /ai_sendLambda 适配器（Python）→ HTTP → /ai_send | Lambda code hook (Python)Lambda code hook（Python） | Lambda (Python) → Bedrock FlowLambda（Python）→ Bedrock Flow | No Lambda不需要 Lambda |
| Multi-turn state多轮对话状态 | Lex session + my MongoDBLex session + 我的 MongoDB | Lex session (slots filled one by one)Lex session（slot 逐个收集） | Lex session | Managed by the Agentic CX AI agentAgentic CX 的 AI agent 自己管理 |
| Transfer to a person转人工 | My agent returns TRANSFER → queueagent 返回 TRANSFER → 进队列 | Yes, a transfer exit in the flow有，flow 里有转人工出口 | The Flow prompt returns a TRANSFER keyword → queueFlow 的 prompt 返回 TRANSFER 关键字 → 进队列 | LLM-based conditions → a person or the main menu大模型判断条件 → 转人工或回主菜单 |
| Barge-in打断 | Supported by LexLex 支持 | Supported by LexLex 支持 | Supported by LexLex 支持 | Supported by agentic voice, on by defaultagentic voice 支持，默认开启 |
| How predictable结果是否可预测 | Medium RAG retrieval + LLM answer中 RAG 检索 + 大模型生成 | High every step is a rule高 每一步都是规则 | Medium LLM answer中 大模型生成 | Lower most flexible, hardest to control exactly较低 最灵活，但最难精确控制 |
| Effort开发量 | Medium: adapter + silence, timeout and retry handling中：适配器 + 静音、超时、重试处理 | Medium to high: bot design, slots, caller check, conditional write中高：bot 设计 + slot + 身份验证 + 条件写入 | Low to medium低到中 | Low: configured in the designer低：在设计器里配置 |
| Good for适合 | Reusing an AI backend you already have已有自己的 AI 后端，想复用 | Bookings, order changes, balances: tasks that must be exact预约、改单、查余额这类必须准确的事务 | A fast FAQ line without running your own RAG快速上线 FAQ 问答，不想自己维护 RAG | Natural, human-like talk without writing code想要接近真人的自然对话，又不想写代码 |
How each path works每条路径怎么工作
Key 1 · 按 1 · Bring my own agent接入自研 agent
Lex V2 → Lambda → /ai_send
Lex V2 only turns speech into text. A Python Lambda adapter sends every utterance over HTTP to the same agent behind my in-house VoIP demo, which decides the intent and when to hand off. It handles silence, retries on backend timeout, and stays warm with an EventBridge schedule.Lex V2 只负责把语音转成文字。Python Lambda 适配器把每一句话通过 HTTP 发给我自己的 agent，也就是自研 VoIP 演示背后的同一个 agent，由它判断意图、决定什么时候转人工。适配器还处理静音和后端超时重试，并用 EventBridge 定时预热。
Key 2 · 按 2 · Self-service booking自助预约
Lex V2 slots → Lambda code hook → DynamoDB
Lex NLU finds the intent and fills the slots. A Python code hook verifies the caller once per call, then books, checks or cancels the appointment. A conditional write keeps one booking per patient, health numbers are masked in logs, and the Lambda role can reach only that one table.Lex NLU 识别意图并收集 slot。Python code hook 每通电话只验证一次来电者身份，然后预约、查询或取消。用条件写入保证每个病人只有一个预约，日志里的健康卡号做了脱敏，Lambda 的 IAM 角色只能访问这一张表。
Key 4 · 按 4 · Bedrock knowledge baseBedrock 知识库问答
Lex V2 → Lambda → Bedrock Flow → Knowledge Base
Lex has no intents here; it only turns speech into text. The Bedrock Flow holds the prompt: it answers from the Bedrock Knowledge Base and returns a TRANSFER keyword when the caller needs a person.这里的 Lex 没有设计意图，只负责语音转文字。意图判断写在 Bedrock Flow 的 prompt 里：从 Bedrock Knowledge Base 里找答案，需要转人工时返回 TRANSFER 关键字。
Key 5 · 按 5 · Agentic CX
Agentic CX block → Connect AI agentAgentic CX 区块 → Connect AI agent
An Agentic CX block in the contact flow hands the call to an AI agent built in the Agentic CX designer. Amazon Connect agentic voice does speech recognition and barge-in; the voice is Polly, set in the main flow. The agent answers from a Q&A knowledge base, makes small talk on Nova 2 Lite, and uses LLM-based conditions to send the caller to a person or back to the main menu.contact flow 里的 Agentic CX 区块把通话交给在 Agentic CX 设计器里搭建的 AI agent。语音识别和打断由 Amazon Connect agentic voice 负责，声音是主流程里设置的 Polly。agent 从 Q&A 知识库回答问题，用 Nova 2 Lite 闲聊，用大模型判断条件，把来电者转给人工或送回主菜单。
Key takeaway核心结论
From key 2 to key 5, the dialog moves from predictable to flexible. Key 2 follows rules at every step. Keys 1 and 4 use an LLM grounded in a knowledge base. Key 5 lets an AI agent run the whole conversation.从按 2 到按 5，对话从“可预测”走向“灵活”：按 2 每一步都是规则；按 1 和按 4 用的是基于知识库的大模型；按 5 让 AI agent 掌控整个对话。
Real contact centers usually mix them: anything that touches money or bookings goes through Lex rules, and open questions go to an LLM.实际的呼叫中心通常会混合使用：涉及钱和预约的事务走 Lex 规则，开放式问题交给大模型。
See also: Twilio voice AI options compared →另见：Twilio 语音 AI 方案对比 →