Skip to content
be-human.ai
← Back to the journal

Tools & integrations

AI Agent Email Send Caps: Set Hourly and Daily Limits First

Set hourly and daily send caps, recipient checks and a hold queue so an AI agent's mistake stays small. A practical setup for business email agents.

Be Human typographic cover with the title: AI Agent Email Send Caps: Set Hourly and Daily Limits First
Conceptual illustration created with AI · Be Human

An AI agent that sends email should operate inside two fixed ceilings, an hourly cap and a daily cap, and every outgoing message should pass recipient checks in the sending layer before it is queued. These limits stop a looping, misled or manipulated agent from turning one bad decision into a large number of messages. Set them before the agent receives send access, keep the counter outside the model, and log blocked attempts as carefully as sent ones.

Key takeaways

  • Enforce send caps in the sending service, never in the prompt, so the model cannot raise its own limit.
  • Use an hourly cap to stop bursts and a daily cap to bound total exposure, with both tracked per agent identity.
  • Check every recipient against an allowlist, a suppression list and domain rules before a message leaves.
  • Hold anything above a cap or outside the rules in a queue that a named person releases or discards.
  • Log every attempt, including blocked and held ones, and change one limit at a time after review.

Why a send cap is the first control to set

A person who makes an email mistake usually makes one. An agent is different because it repeats. A retry that does not know the first attempt already succeeded, a misread row in a spreadsheet, or a customer message that contains instructions the agent follows can produce the same email many times within minutes. The mail provider accepts each message as a normal send, and the damage lands on your customers, your domain reputation and your support team at once. A cap turns an unbounded failure into a bounded one. It does not need to understand what the agent intended. It only needs to count and stop.

This is why the control belongs before the agent is trusted with anything useful. Prompt instructions can be ignored, reworded or overridden by content the agent reads. Authorization decisions belong in code that the model cannot edit, which is the general principle described in the OWASP authorization cheat sheet.

How do you choose the hourly and daily numbers?

Start from the work the human role does today, not from a number that sounds safe. Pull the sent folder for the mailbox or alias the agent will use, look at an ordinary day and a busy day, and note how many messages a person in that role actually sends. Set the daily cap comfortably above the busy-day peak so legitimate work is rarely blocked. Set the hourly cap low enough that a loop is stopped within minutes, not discovered at the end of the day.

Illustrative scenario: a support coordinator sends about twenty follow-up messages on a busy day and rarely more than forty. The team sets a daily cap of sixty and an hourly cap of fifteen. A duplicate trigger that tries to send the same follow-up repeatedly reaches the hourly ceiling within the first quarter hour, and the remaining attempts wait in the queue for a person to review.

Raise a cap only through a written change, and only after the logs show that legitimate work hit the ceiling. A cap that rises whenever someone complains about a blocked message has stopped being a control.

Which recipient checks should run before every send?

Recipient checks cost far less than cleanup after a bad send. Run them in code, in a fixed order, before the message enters any queue:

  • Allowlist for the role: the agent writes only to the customer accounts, staff addresses or partner domains listed for its role.
  • Suppression list: addresses that have unsubscribed, bounced or complained are never contacted, whatever the task says.
  • Domain rules: flag any domain the role has never written to, and block free mail domains if your business does not use them for this work.
  • Distinct recipient count per run: a single task that resolves to more than the expected number of addresses goes to a person before anything is sent.
  • Duplicate guard: the same recipient and the same template within a short window is blocked, using the same logic that prevents duplicate records in other tools, as covered in idempotent tool writes.
  • Link and attachment rules: messages with links outside approved domains, or attachment types outside the role's list, are held.

Each check should return one of three results, send, hold or reject, with a reason code. A held message is not a failure. It is a decision waiting for a human, and keeping that distinction is what keeps the queue useful instead of noisy.

Where should the counter and the rules live?

The counter and the rules belong to the sending layer, the service that actually calls the mail provider. Store the counters in a shared store keyed by agent identity and mailbox, so a restart or a second worker cannot quietly reset them. Load the limits from configuration the agent cannot write, and give change rights to a named owner. Keep draft access and send access as separate permissions. An agent can create drafts across a wider scope while sending only within a narrow one, a split that is worth reviewing in Gmail and Outlook access for agents.

The agent's identity should also stay separate from any employee login. A dedicated identity with its own scope is easier to limit, to audit and to revoke, which matters the first time a cap needs to be tightened in a hurry.

A step-by-step plan for the first rollout

Use this sequence with the operations lead who will own the role:

  • Name the role, the mailbox or alias it may send from, and the one person who approves changes to its limits.
  • Write the allowlist of recipients and domains for the first phase, and keep it short enough to review in one sitting.
  • Measure the normal and busy-day send volume of the human process from the sent folder.
  • Set the hourly and daily caps above the busy-day peak and well below the level where a mistake would hurt the business.
  • Route every send through the sending service, and give the agent draft access only until the logs show the behavior you expect.
  • Connect the held queue to a named approver, with a response expectation written down.
  • Test a deliberate loop in a sandbox mailbox and confirm that the hourly cap stops it before the daily cap is reached.
  • Review the logs at a fixed interval, change one limit at a time, and record the reason for each change.

What should you log, and who gets alerted?

Log each attempt with the agent identity, the mailbox, a template or task identifier, the recipient in a form your privacy policy allows, the reason code, the counter value at that moment and any human decision that followed. The OpenTelemetry observability primer describes logs, metrics and traces as complementary signals, and all three help here: traces show the path a task took, metrics show counter levels over time, and logs record each decision.

Alert on hitting the hourly cap, not only on errors, because a blocked loop is evidence that something upstream has changed. Send alerts to a person who can act on them, not to a shared inbox that nobody reads. Personal data in these logs should follow the same minimization rules as the rest of your records. The NIST AI Risk Management Framework is a useful reference for treating logs like this as part of how the system is governed.

How Be Human handles send caps in managed operations

In a Be Human engagement, the first step is to define the single role, its mailbox, its recipient scope and its caps, together with the approval boundary for any increase. The checks run in the sending layer, the counters and held queue are monitored, and limits are reviewed on a recurring schedule as part of permission review. When a cap is hit unexpectedly, incident response starts with the logs, and any change to a limit goes through the approver the client named, so the client keeps authority over what the agent may send.

If you are unsure which email task should go to an agent first, the free Be Human AI diagnostic can help you find one task worth handing over with clear limits.

Frequently asked questions

Should I set the hourly cap or the daily cap first? Set both on the same day, because they protect against different failures. The hourly cap stops a fast loop before it grows, and the daily cap limits total exposure if a problem is slow and steady. Start with both set low, then adjust them from observed volume rather than from guesses.

Can the model enforce its own send limit? It should not. A model can be wrong, misled by content it reads or instructed by someone outside the company, so the counter and the rules must live in code the agent cannot modify. The agent can know its limits and report when it is close to them, but the sending service has the final say.

What happens to a message that hits a cap? It should go to a held queue with a reason code, not disappear and not send anyway. A named person reviews the queue and either releases the message, asks for a revised request or discards it. Keep held messages long enough to review patterns, because they are some of the best evidence you will have about what the agent attempted.

Do send caps replace human review of drafts? No. Caps limit how much can go wrong, while draft review decides whether a particular message is right. For higher-stakes roles, keep a person approving drafts alongside the caps, and reduce that review only after the logs show the role behaves as expected.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →