# Set up AI email customer support with a daily send limit

> Let an agent answer routine support questions and send on its own, held in check by a daily limit, a recipient cap and a block on automated mail.

Updated: 2026-10-27 · Fact-checked: 2026-10-27 · Canonical: https://commandmail.app/en/guides/ai-email-support · Deutsch: https://commandmail.app/de/anleitungen/ki-kundenservice-per-e-mail.md

An agent can answer customer questions in your support inbox and send the replies itself, without you approving each one. To keep that from getting out of hand you need limits that do not depend on the agent behaving: a fixed number of emails per day, a cap on recipients, and no replies to machines. In Command Mail the server checks these limits on every send the agent attempts.

Command Mail is our product. This guide covers the setup with your own agent in Claude, ChatGPT, Claude Code or Codex. If you would rather see every reply before it leaves, [Reply to emails with AI](https://commandmail.app/en/guides/reply-to-emails-with-ai) is the better route. This page is about the case where the agent really should send on its own.

## What this is good for

Autonomous replies pay off for questions that come in often and have a settled answer: opening hours, delivery times, where to find an invoice, how to reset a password, whether something is in stock. The agent knows your offers, agreements and tone only as far as you give it the information and sources.[^1] When information is missing, it saves a draft for you.

Keep a person on refunds and anything involving money, cancellations, legal questions, and anything where a wrong promise gets expensive. A good standing instruction tells the agent exactly which topics it answers itself and which ones it only drafts.

## What you need

- A Command Mail account on Pro. The free plan does not send; sending starts with Pro, which the plan summary lists at €19 per month for the first 100 paying Pro accounts and €29 after that, with a 7-day trial for €1.[^2]
- A support address that is allowed to send from Command Mail: your own domain, or an address with a sending route. An address that only arrives by forwarding receives right away but sends only once you give it a sending route under Settings → Connections. With SMTP you send through your mail provider as that address. With the relay, a Command Mail relay address is the sender and replies reach your support address through Reply-To.
- An agent that reaches Command Mail over MCP at `https://commandmail.app/mcp`, and a way to run it on a schedule. You set up the schedule in the agent, not in Command Mail.[^1]

## Setup, step by step

1. **Connect the agent with the right permissions.** On the consent screen, switch on `mail:read` and `mail:send`. `mail:send` allows sending without your approval and starts switched off, so you have to turn it on yourself. Add `mail:draft` as well: without it the agent cannot park hard cases as drafts. A sorting agent should never get `mail:send`.[^1]
2. **Set the support mailbox to `auto`.** Under Settings → Automation & rules. Only on `auto` may an agent send directly. On any lower level the server refuses with `autonomy_blocks_send`, and the reply belongs in a draft.[^3] The level is per mailbox, so your personal mailboxes can stay on `suggest`.
3. **Set the daily limit.** Also under Automation & rules. The default is 50 agent emails per day, and anything from 0 to 500 is allowed. Start low, say 20, and raise it once the log looks clean.
4. **Write the standing instruction.** Keep it short and specific, for example: "Every 30 minutes, check support@. Answer only questions about delivery times, invoices and password resets, using the wording in our FAQ. Save everything else as a draft with a note on what is missing. Never quote a price that is not in the FAQ."
5. **Watch the first run.** From a second address, send yourself one typical question and one the agent should leave alone. Afterwards Sent shows what went out, and Drafts shows what was held back.

## How each limit works

We traced these limits in the shared send function of the application code (commit `9cab1eb`, `outbox.ts`), which direct agent emails pass through.

**A daily limit for the account.** The count is agent emails per UTC day across your whole account, not per mailbox. A slot is reserved before the message is handed to the delivery service. Once the day is used up the agent gets `daily_send_limit` back and nothing goes out. A send with an uncertain outcome keeps its slot; only a send that was definitely refused gives it back. According to the documentation you change the number in the app's settings. Mail you send yourself and approved drafts are not counted, so a limit of 0 blocks only direct agent sending.

**At most 20 recipients per message.** To, Cc and Bcc together. Above that the server refuses, so a single message cannot get around the daily limit.[^2]

**One key per message.** Every agent email carries its own `sendKey`. If the agent repeats a request with the same key after losing the response, no second message is sent. That protects against duplicate replies when a run crashes and starts over.[^1] Only a send that was definitely refused frees the key again. When the outcome is uncertain, the agent checks the delivery status and does not retry with a new key. Keys not tied to a stored message are cleared after 30 days.

**No replies to machines.** Before the agent answers a message, the server checks whether it was sent automatically: `Auto-Submitted` under RFC 3834; `Precedence: bulk`, `junk` or `list`; mailing-list headers; an empty `Return-Path` on bounces; the usual out-of-office subject lines; and no-reply senders.[^4] The agent also does not write on its own to no-reply addresses or to addresses whose messages stored in this account were all recognised as automatic. In those cases it gets `automated_sender` back.[^1] If the message being answered is itself recognised as automatic, Command Mail also refuses the agent's reply draft; reply yourself in the app if needed. Recognition works from typical markers. It stops the common loop of two autoresponders answering each other, but it is not complete protection.

**The level is read at send time.** If you switch the mailbox back to `suggest` while a run is going, that holds from the next send attempt. Together with a daily limit of 0, this blocks further direct send attempts.

## Decisions to make on purpose

`mail:read` reads across the whole account, not just the support mailbox. A support agent you connect can see every mailbox on the account. If that does not suit you, put support in a separate Command Mail account.

With `mail:read` and `mail:send` together, an agent may also permanently delete mail and empty Trash and Spam in mailboxes on `auto`.[^2] A support agent has no need for that. Rule it out explicitly in its instruction and keep an eye on the log.

Emails an agent sends directly get the signature you set up in Command Mail, if you have one. On approved drafts Command Mail adds no signature; the approved content goes out as it is. If you use a signature, tell the agent not to sign off with a name of its own, or the name appears twice.

Customer email is untrusted text, and a message can contain instructions aimed at the agent. Command Mail flags suspected passages and marks email content as untrusted. That is not a reliable defence. What the agent can do in the worst case is set by its permissions, the mailbox level and the limit. More in [Prompt injection in email](https://commandmail.app/en/learn/email-prompt-injection).

## Check that it works

Emails the agent sends directly appear under Sent; approved drafts are also recorded in the activity log. If a send reports that the message went out but no copy was stored, or that the outcome is uncertain, check the delivery status and do not send again. In the first week, read the sent replies and the drafts it held back every day. The drafts show which questions the agent was right to leave alone and which belong in your FAQ.

The agent sees its own state through `capabilities`: which mailboxes may send directly, what the daily limit is and whether it has been reached today. A well-set-up agent calls it at the start of every run and stops when the limit is hit.

## Troubleshooting

**`autonomy_blocks_send`.** The sending mailbox is not on `auto`. Change it, or have the agent save drafts.

**`daily_send_limit`.** The limit for this UTC day is used up, or it is set to 0. At 0, direct sending stays blocked the next day too; otherwise the count resets at midnight UTC.

**`automated_sender`.** The message or the recipient was recognised as automatic. If it is a person after all, send the reply yourself in the app.

**Sending is not included in the plan.** The account is on the free plan.

**The reply comes from the wrong sender, or not at all.** Check that the support address has a sending route. A forwarded-only address sends only once SMTP or the relay is set up; with the relay, the relay address is the sender and your support address goes in Reply-To.

## When a help desk fits better

If you run a support team with several people, assignments, SLAs and reporting, a help desk with a built-in AI agent is the better home. Command Mail does not replace a ticketing system.

Front's Autopilot triages conversations, hands them over with a draft or resolves them itself, across email, portal, SMS, WhatsApp and Slack. It is billed per outcome, and a resolution costs $0.89. An email counts as resolved when no teammate replies within 72 hours. Front says only English is optimised so far.[^5] Triage costs $0.05 and a handoff $0.39. Front's seat prices come on top, from $25 per user per month billed annually.[^6] Its pricing page also lists external AI agents such as Claude or Codex working in Front.

Intercom's Fin works with Intercom's own help desk or with others such as Salesforce and HubSpot, and handles email, tickets and chat among other channels. Fin charges $0.99 for resolutions and other defined outcomes and $9.99 for sales qualifications, with a minimum of 50 outcomes a month when used with another help desk.[^7]

Help Scout has a free plan with five users and one inbox. Its separately billed Envoy chatbot answers questions from your knowledge base for $0.75 per resolution; the pricing page does not describe an agent that sends on its own from an email inbox.[^8]

The built-in AI agents of these products are billed by outcome type, depending on the vendor. In Command Mail the agent you connect yourself does the work, you pay for Pro and for your agent; the daily limit caps the number of emails the agent sends directly.

## Next steps

Start with a narrow list of topics and a low limit. Whatever runs cleanly for a week can grow. The permissions and what each level allows are in [Permissions and autonomy](https://commandmail.app/en/docs/permissions-and-autonomy). Why approvals are often the better choice is covered in [Human in the loop for email](https://commandmail.app/en/learn/human-in-the-loop-email).

[Start Command Mail for free](https://commandmail.app/login?signup=1) and connect the support inbox with read and draft permissions first. Sending comes with Pro.

## Sources and review date

Reviewed: 2026-10-27. Vendor pages and Command Mail's public pages were retrieved on that date. Statements about the limits on agent sending also rest on the application code at commit `9cab1eb`: the send function, the daily limit, send keys and the recognition of automatic mail. This guide does not claim a send test of our own.

## Frequently asked questions

### How many emails can the agent send per day?

By default 50 directly sent agent emails per UTC day for the whole account. You set the number in the app anywhere from 0 to 500, and 0 blocks direct agent sending. Mail you send yourself and approved drafts do not count.

### Will the agent reply to out-of-office notes and newsletters?

Not on its own. When Command Mail recognises an automatic message, for example under RFC 3834, with Precedence bulk, from a mailing list, as a bounce or from a no-reply address, the server refuses an agent reply. Recognition works from typical markers and can miss individual cases.

### What happens when the limit is reached?

The server refuses every further direct agent email for that UTC day, and nothing is sent. With `mail:draft`, the agent can save the reply as a draft for you to approve instead.

### Do I need a paid plan?

Yes. The free plan does not send. Sending, including autonomous sending by an agent, starts with Pro.

### Can a customer trick the agent with a crafted email?

That risk comes with any agent that reads outside mail. Command Mail marks email content as untrusted and flags suspected prompt injection, but it cannot reliably stop it. Give a support agent only the permissions and information it needs for routine questions.

## Sources

1. [Command Mail: documentation, section Let your agent reply](https://commandmail.app/en/docs), accessed 2026-10-27
2. [Command Mail: plans, agent permissions and autonomy levels](https://commandmail.app/llms.txt), accessed 2026-10-27
3. [Command Mail: permissions and autonomy](https://commandmail.app/en/docs/permissions-and-autonomy), accessed 2026-10-27
4. [RFC 3834: Recommendations for Automatic Responses to Electronic Mail](https://www.rfc-editor.org/rfc/rfc3834), accessed 2026-10-27
5. [Front: Autopilot channels, languages and billing](https://help.front.com/en/articles/4890624), accessed 2026-10-27
6. [Front: pricing](https://front.com/pricing), accessed 2026-10-27
7. [Intercom: Fin AI Agent pricing](https://fin.ai/pricing), accessed 2026-10-27
8. [Help Scout: pricing and the Envoy AI assistant](https://www.helpscout.com/pricing/), accessed 2026-10-27

[^1]: [Command Mail: documentation, section Let your agent reply](https://commandmail.app/en/docs), accessed 2026-10-27
[^2]: [Command Mail: plans, agent permissions and autonomy levels](https://commandmail.app/llms.txt), accessed 2026-10-27
[^3]: [Command Mail: permissions and autonomy](https://commandmail.app/en/docs/permissions-and-autonomy), accessed 2026-10-27
[^4]: [RFC 3834: Recommendations for Automatic Responses to Electronic Mail](https://www.rfc-editor.org/rfc/rfc3834), accessed 2026-10-27
[^5]: [Front: Autopilot channels, languages and billing](https://help.front.com/en/articles/4890624), accessed 2026-10-27
[^6]: [Front: pricing](https://front.com/pricing), accessed 2026-10-27
[^7]: [Intercom: Fin AI Agent pricing](https://fin.ai/pricing), accessed 2026-10-27
[^8]: [Help Scout: pricing and the Envoy AI assistant](https://www.helpscout.com/pricing/), accessed 2026-10-27

---

Command Mail: one inbox for you and your AI agent. Start free: https://commandmail.app/login?signup=1 · Connect your agent: https://commandmail.app/en/agents · MCP: https://commandmail.app/mcp