AI 代理：已知限制與變通方法

AI 代理功能強大，但有一些平台層面的限制需要注意。請使用以下建議的變通方法，確保您的 AI 代理表現一致且可靠。

對話可見性

AI 代理人對對話的元資料、歷史與內部上下文的可見性有限。

通道與來源的元資料

AI 代理人會自動接收基本通道上下文，包括通道類型、通道名稱與通道 ID。在適當指示下，您的 AI 代理人可以利用這些資訊調整其回應長度、格式、語氣與行為，以適應不同通道。

來自 Meta 與 TikTok 的廣告歸因

當對話由 Meta 點擊開啟對話廣告 或 TikTok 訊息廣告 開啟時，您的 AI 代理人會在通道上下文旁自動接收一個廣告上下文區塊。不需要額外配置。此區塊包含：

廣告平台： Meta 或 TikTok
Ad Campaign: The campaign name
廣告群組：廣告群組或廣告組合的名稱（如有提供）
廣告名稱：單一廣告或創意名稱（如有提供）

這表示您可以撰寫指示，根據聯絡人來自哪則廣告進行調整。例如：以活動專屬訊息開場、轉派到不同團隊，或觸發不同的對話流程。

注意：廣告上下文區塊僅在該廣告進入事件發生於對話最後 20 則訊息內時才會包含。如果對話已超出此範圍，則不會包含廣告上下文。如果聯絡人在最近 20 則訊息內開啟超過一則廣告，則只會包含最近的一次廣告進入事件。請撰寫指示以處理兩種情況。

需注意的限制：

通道與來源上下文都以高層次方式提供。雖然您的 AI 代理人會自動接收結構性元資料（例如廣告平台、活動與廣告名稱），但無法查看廣告的實際內容。您仍需撰寫指示，告訴代理人根據該上下文採取何種行動，以及廣告內容為何。
AI 代理人不會自動接收來源元資料，除非對話是由 Meta 或 TikTok 的付費廣告所啟動，此情況會提供上述的廣告上下文區塊。

建議的變通方法：

如果您的使用情境需要得知聯絡人來自於付費廣告之外的來源，請使用工作流程擷取例如廣告內容等資訊，並將其儲存在聯絡人欄位。

例如：

建立一個含有對話已開啟觸發條件的工作流程
根據通道或來源新增條件
更新一個聯絡人欄位，例如 潛在客戶來源 或 活動名稱

接著，您可以指示您的 AI 代理在其指令中引用該聯絡人欄位。這可確保行為一致且可靠，即使通道或來源細節無法自動取得。

通道感知行為的範例提示

您可以透過明確指示，指導 AI 代理人在不同通道上如何調整其回應。以下是您可以使用的一些範例提示：

調整回應長度與語氣

如果此對話發生在 Instagram 或 TikTok，請將回覆保持短小（1–3 句）、友善並具對話性。 避免冗長說明。

如果此對話發生在 WhatsApp，可提供較詳細的回應，但仍保持簡潔且易讀。

如果此對話透過 Email 進行，請使用較有結構且正式的語氣。 較長的說明與清晰格式化的回應是可以接受的。

控制格式與內容

在 TikTok 回應時請勿分享連結，除非聯絡人明確要求。

透過 Email 回覆時，可在適當情況下加入連結、要點清單及分步說明。

基於通道的行為與升級

如果對話發生在 Instagram 或 Facebook，且聯絡人詢問帳單、付款或帳戶相關問題，請立即轉給人工代理人。

如果對話發生在 WhatsApp，除非聯絡人顯得沮喪，否則請先嘗試基本的故障排除再升級處理。

在提及對話時

當要指明聯絡人從哪裡發送訊息時，請說「此聊天」或「此對話」。 切勿明確提及通道名稱。

Assignment awareness

AI Agents receive an Assignment Event when they are assigned to a conversation. This helps the AI Agent understand that it has just been assigned and decide what to do next.

This applies when an AI Agent is assigned by:

A human agent
A Workflow
Another AI Agent

This means you can now instruct your AI Agent to act based on assignment events. For example, you can ask it to:

Greet the Contact when it is assigned to the conversation.
Fetch Contact data or route the conversation based on your instructions.
Route the conversation to the correct team if the Contact has clearly stated their problem.
Continue the handoff when AI Agent 1 assigns the conversation to AI Agent 2.

Example prompt:

When you are assigned to a conversation, greet the Contact and check their latest message. If they have clearly stated their issue, route the conversation to the correct team. If not, ask how you can help.

Limitations to be aware of:

The Assignment Event only applies when an AI Agent is assigned, not when a human agent or team is assigned.
It does not apply when an AI Agent is unassigned or removed.
The event is only available within the AI Agent’s 20-message context window.
If multiple events happen in the same conversation, only the most recent event is available to the AI Agent. For example, if a Workflow is configured to assign an AI Agent when a conversation is reopened, the Assignment Event is kept because it happens after the Reopened Event.

Conversation reopen awareness

AI Agents receive a Conversation Reopened Event when a closed conversation is reopened by an incoming Contact message.

This helps AI Agent understand that the Contact is returning to an existing conversation instead of starting a new one.

This means you can now instruct your AI Agent to act based on reopened conversation events. For example, you can ask it to:

Identify when a Contact has replied to a closed conversation.
Respond differently to simple acknowledgement messages such as “thanks”, “ok” or “got it”.
Continue helping if the Contact sends a genuine follow-up question.

Limitations to be aware of:

The Reopened Event does not apply when a snooze period expires without a new incoming Contact message.
If multiple events happen in the same conversation, only the most recent event is available to the AI Agent. For example, if a Workflow is configured to assign an AI Agent when a conversation is reopened, the Assignment Event is kept because it happens after the Reopened Event.
The event is only available within the AI Agent’s 20-message context window.

Example prompt:

When a conversation is reopened, check if the Contact's message is a simple acknowledgement, such as "thanks", "ok" or "got it". If so, close the conversation. If it is a genuine follow-up question, continue assisting the Contact.

透過外部工具指派

當 AI 代理透過外部工具 — 例如開發者 API、n8n 或 Make.com — 被指派到對話時，該指派動作本身不會觸發 AI 代理回覆。

這是預期的行為。 Respond.io 限制此行為，以防止 AI 代理在透過自動化同時被指派多個對話時批量回覆。

一旦指派，該對話會顯示為指派給 AI 代理。然而，AI 代理不會主動發送訊息。 AI 代理會在聯絡人發送訊息後回應。

注意: 對此限制沒有變通方法。

內部上下文與訊息歷史

AI 代理人無法查看內部註解，且只能存取對話中最近的 20 則訊息。

因此，AI 代理人在做決策時不應依賴通道來源、內部備註或較舊的對話上下文。

建議的變通方法：

將重要的上下文儲存在聯絡人欄位或生命週期階段，並指示 AI 代理人改為參考那些欄位。

例如：

使用工作流程或人工流程，讓像 方案類型、客戶狀態 或 問題類別 等欄位保持最新
使用生命週期階段（例如：新潛在客戶、合格潛在客戶、客戶）來指導代理人的回應方式
在你對 AI 代理人的指示中，明確參考這些欄位（例如：「如果生命週期是客戶，則優先支援相關回應。」）

這可確保 AI 代理人即使在無法存取較舊訊息或內部註解的情況下，仍能使用可靠的結構化資料做出一致的決策。

Replied-to messages

AI Agents cannot see the message a Contact is replying to. When a Contact uses the reply feature to quote an earlier message, the AI Agent receives only the new incoming message — not the original message being replied to.

This applies regardless of how recent or old the replied-to message is. The AI Agent has no visibility into the quoted message, even when it falls within the most recent 20 messages.

As a result, the AI Agent may give incomplete or contextually wrong responses when a Contact replies to an older message and assumes the AI Agent understands the full thread.

建議的變通方法：

Write your AI Agent instructions to ask the Contact to restate their request when the message lacks context. When a human agent takes over, review the replied-to message manually before responding.

Note: Support for passing the replied-to message to the AI Agent is being considered. Vote for or follow the reply-to message support feature request to track progress.

外部能力

AI 代理主要在 respond.io 內運作，但現在可使用 AI 代理動作：發送 HTTP 請求 與外部系統互動。這使 AI 代理能呼叫外部 API 以擷取即時資料或在第三方系統觸發動作。

進一步了解 AI 代理動作：發送 HTTP 請求。

訊息類型與附件

AI Agents can send files and images as rendered attachments directly in the conversation. When an AI Agent’s response includes a valid file or image URL, it delivers the file as an actual attachment — the same way human agents and Workflows send files today.

Image files (.jpg, .jpeg, .png) render as inline previews
Document files (.pdf, .docx, .pptx) appear as file chips with filename and type
All attachments appear in the conversation’s Attachments tab

How to enable it:

Add file or image URLs to your AI Agent’s Knowledge Sources or Instructions.
Add this instruction to your AI Agent prompt: “Send image/file as attachment whenever there’s a valid image/file URL from Knowledge source or Instructions.”

Using the Make HTTP requests action? Image or file URLs returned in the API response can be sent as attachments the same way — see Sending images or files from the API response.

Limitations:

Supported file types: .pdf, .docx, .pptx, .jpg, .jpeg, .png — other URL types stay as plain text
Login-gated URLs (private Google Drive, Dropbox) are not supported in this release
Maximum of 5 attachments per AI Agent response — additional URLs are dropped
AI accuracy may vary with complex URLs in Knowledge Source (known limitation)
Videos, voice notes, stickers, and other rich media remain unsupported

For unsupported file types or login-gated assets, use a Workflow step or assign to a human agent.

Best practices to reduce hallucinated or wrong attachments:

Use simple URLs — avoid long query strings or redirect chains.
Give each item a unique, clearly labeled URL — near-identical URLs confuse the AI (e.g. image001.png, image010.png).
Avoid URL shorteners, redirects, or login/CAPTCHA-protected links — these can’t be rendered as attachments anyway.
Add the trigger instruction only in your agent’s Instructions, not inside the Knowledge Source — mixing the two can cause unpredictable behavior.

Receiving unsupported messages

When a Contact sends a message type that respond.io cannot process, such as a message reaction or other unsupported format, the AI Agent receives an [Unsupported message] signal in the conversation history.

This means your AI Agent is now aware when an unsupported message is received and can respond accordingly. You can also write instructions to customize how your AI Agent handles these cases.

Example prompts for handling unsupported messages:

Ask the Contact to resend as text:

If you receive [Unsupported message], reply: "It looks like I wasn't able to receive your last message. Could you send it as a text message instead?"

Ignore and do nothing:

If you receive [Unsupported message], do not reply. Wait for the Contact's next message.

Acknowledge without mentioning the limitation:

If you receive [Unsupported message], reply: "I'm not able to view that message type. Feel free to describe what you need and I'll be happy to help!"

模型設定

AI 代理人不支援手動的溫度設定，因為較新的 AI 模型不再提供溫度控制。

若你想調整 AI 代理人的「創意程度」（類似於溫度），最佳作法是在你的指示中直接要求。例如，你可以根據使用情境引導 AI 表現得更有創意或更嚴謹。

高創意（較高的「溫度」）：

以客戶的語氣溫暖且富有創意地回應。

低創意（較低的「溫度」）：

僅使用提供的公司知識來源和對話記錄來回答客戶的問題。 不要猜測或新增資訊。

這能讓你更好地控制語氣與一致性，而無需專門的溫度設定。

Silent mode (background tasks only)

In some use cases, you may want your AI Agent to perform background tasks — such as updating Contact Fields, adding Tags, or updating Lifecycle stages — without sending any message to the Contact.

By default, AI Agents generate a reply in every turn. To suppress all outgoing messages, add the following prompt at the top of your AI Agent's Instructions:

# CORE RULE: NO-REPLY POLICY (MANDATORY)- You MUST NOT generate or send any user-facing message under any circumstance (no greetings, no questions, no confirmations).- If the system requires a response payload, return an empty response only (e.g., response: ""). Do not include any other text.

When to use this:

Your AI Agent is only used to classify conversations (e.g., update Tags or Contact Fields based on message content)
Your AI Agent performs routing or assignment logic before handing off to a human agent or another AI Agent
Your automation requires an AI Agent step for decision-making, but the Contact-facing reply is handled elsewhere

Important considerations:

Place the no-reply instruction at the very top of your Instructions to ensure it takes priority over all other prompts.
Always test this in a test conversation before publishing, as the AI Agent may still occasionally generate a response depending on the complexity of your other instructions.
If your AI Agent also performs actions (e.g., Assign Conversation, Close Conversation), those actions will still execute normally — only the text reply is suppressed.
When an AI Agent is assigned to a conversation, the message composer is disabled for human agents. This means you cannot use a silent AI Agent for background tasks while manually replying to the Contact at the same time.