AI 智能體需要致電時，如何透過 MCP 將通話交由真人處理

Disclosure up front: I run TaskBullet, and the last section describes our MCP server. Everything above it applies no matter which tool you use.

Agents are getting good at research, email and forms, but a lot of real-world work still ends at a phone number. The dentist that only books by phone, the supplier with no online stock check, the county office that answers questions only by voice. That's where most agent workflows quietly stop and hand the job back to a person.

Here is a simple pattern for when an agent needs a call made.

1. Treat the call as a structured job, not a chat

Don't ask the agent to "call and find out about pricing." Have it produce a small spec:

• the phone number and who it belongs to
• the goal of the call in one sentence
• a short list of questions, each with an id (q1, q2, q3)
• what to do if the person can't answer (leave a callback number, ask who can, etc.)

Ids matter. They let answers come back matched to questions, so the agent can drop them straight into its next step without re-parsing a transcript.

2. Validate before you dial

Have a check step before anything costs money or reaches a real person: is the number in a supported region, are the questions answerable on a call, is anything in the spec sensitive or something the caller shouldn't promise. Catching a bad spec here is much cheaper than a wasted call.

3. Let a person make the call

Voice bots are fine for simple, predictable calls. For anything where the other side goes off-script (hold queues, transfers, "let me check with my manager"), a human caller usually handles it better. The agent doesn't need to know who dialed; it just needs the result in a predictable shape.

4. Return structured answers plus notes

What comes back should be:

• an answer for each question id (or "unanswered" with a reason)
• short wrap-up notes for anything that didn't fit a question
• the transcript or summary if you want it for audit

Now the agent can keep going: update the CRM, compare quotes, or schedule the follow-up.

5. Make it async

Calls take minutes, not milliseconds. Use a place / poll (or webhook) / fetch-result flow and let the agent work on other things in between. Allow updates and cancels while the call is queued.

How we do it

TaskBullet's MCP server (npx -y @taskbullet/mcp-server, v0.2.0) exposes this exact flow as tools: validate_call, place_call, get_call, update_call, cancel_call and get_transcript. Your agent supplies the questions; a real TaskBullet assistant makes the call and returns answers matched by question id plus wrap-up notes. US and Canada numbers. Pricing is $2.50 for the first minute (minimum), then $1 per additional minute. The same MCP connection also has a delegate_task tool for other jobs an agent can't finish on its own.

Docs: https://taskbullet.com/docs/agent-api.md
Hub: https://taskbullet.com/hub

If you've built your own version of this pattern, I'd like to hear how you handled hold queues and voicemail.