AI Agent Email Sending Limits: Hourly Caps, Drafts and Allowlists
Set per-hour send caps, draft-first modes and recipient allowlists before an AI agent sends mail for your company. A plain guide for operations leaders.

An AI agent that can send email for your company should run under three controls: a cap on how many messages it can send in an hour, a draft-first mode where a person approves each message before it leaves, and an allowlist of recipients it may write to. Put these in place before the agent gets a live mailbox, then widen them only when the review record shows the agent behaves as intended. The sections below explain how to choose the values, what to log, and when drafts can safely become sends.
Key takeaways
- Start every new message type in draft-first mode, not only the messages that feel risky.
- Set an hourly cap low enough that a faulty loop is contained before anyone notices it.
- Use recipient allowlists by domain and by named contact, and hold anything outside them.
- Give the agent its own mailbox identity so its messages can be audited and suspended on their own.
- Widen permissions based on reviewed drafts and logged behavior, not on a calendar or on pressure to move faster.
Why email needs its own limits
Email is outbound and public to the recipient. Once a message is sent, the company cannot quietly correct it the way it can fix a cell in a spreadsheet. A misread instruction, a retry that never stops, or a stale contact list can produce a large number of messages within minutes, and the first sign of trouble is usually a complaint from someone outside the company.
The OWASP Top 10 for LLM Applications describes this risk as excessive agency: a model-driven system that can take consequential actions beyond what its task needs. The practical answer is not a better prompt. It is a set of hard limits outside the model that hold even when the model misunderstands. Limits give the operation a ceiling on the damage while the team learns what the agent actually does with real mail.
The three modes: read, draft and send
Most teams treat email access as a single permission. It is better to separate three modes. In read mode, the agent can see threads it has been given access to and summarize them, but it cannot create anything outbound. In draft mode, it writes replies or new messages into a review queue or a drafts folder, and a named person sends them or rejects them. In send mode, the agent sends messages directly within its limits.
The distinction matters because most early value comes from drafting. A draft saves the time of writing while keeping a human at the point where the message becomes irreversible. For a deeper comparison of these permission levels, see Read, draft and send: three different AI permissions. Send mode should be earned for each message type, one at a time, rather than granted to the whole mailbox at once.
How to set an hourly cap and a daily ceiling
An hourly cap limits how fast the agent can send. A daily ceiling limits the total. You need both, because a loop that sends slowly can still reach a large total by the end of the day. Choose the hourly number from the real volume the role needs, then add a modest margin. The aim is not to make the cap generous. The aim is to make the worst plausible failure tolerable.
A reasonable starting set of controls looks like this:
- An hourly send cap based on the role's observed normal volume, with a small margin above it.
- A daily ceiling that is lower than the total the team could review in a day.
- A per-recipient limit, so one contact cannot receive repeated messages from the agent in the same day.
- A rule that messages over the cap are queued and surfaced to a person, not silently dropped.
- An alert to a named owner when the cap is reached, with the reason recorded.
Queueing matters more than most teams expect. If the agent drops messages over the cap, the business loses work without knowing it. If it queues them, a person can see the backlog and decide whether the messages should go out, be rewritten, or be cancelled.
Building recipient allowlists that hold up
An allowlist defines where the agent may send. The simplest version lists approved domains, such as the company's own domain and the domains of current suppliers. A stricter version lists named contacts or contact groups, such as customers who have an open support case. Anything outside the list becomes a hold, meaning the message is stored with a reason and waits for approval.
Two details cause most of the trouble. First, replies from unknown senders are a common route for a loop or an injected instruction, so they should be held by default rather than answered automatically. Second, contact lists change. If a CRM sync or a spreadsheet adds new addresses, the allowlist should not expand itself. Changes to the list need an owner and an approval step. For the suppression side of this problem, including addresses that bounce or opt out, see Email Bounce Handling for AI Outreach Agents: Detect, Suppress, Log.
Step-by-step plan for the first stage
The first stage should be small and observable. The following sequence works for most operations teams that want an agent to handle a specific email task.
- Step one: name the single task, such as drafting replies to delivery-window questions from existing suppliers.
- Step two: give the agent its own mailbox identity, separate from any employee's account.
- Step three: enable read access only to the threads the task requires, and nothing else.
- Step four: enable draft mode, with every draft going to a named reviewer.
- Step five: set the hourly cap, the daily ceiling and the per-recipient limit, and assign an owner to the queue.
- Step six: define the allowlist and set all other recipients to hold.
- Step seven: review every draft and every hold for a fixed period, recording the reviewer's decision and the reason for any rejection.
Do not move to send mode until the review record shows the drafts are accurate, the holds are sensible and the reviewer rarely needs to rewrite a message for reasons the agent could have avoided. Keep the review record. It is the evidence for every later decision about permissions.
Illustrative scenario: an operations agent drafts replies to supplier questions about delivery windows. In the first stage, every draft goes to the purchasing coordinator. A message from an address that is not on the supplier list is held, with the sender and the subject line recorded. After several reviews, the coordinator notices that the agent often adds an unconfirmed date. The team tightens the instruction and adds a rule that any message containing a date must be drafted, never sent, until the date is confirmed in the order record. Send mode is not considered for that message type until the date problem has been absent in review for a sustained period.
What to log, review and retire
Every outbound attempt should leave a trace: the draft or message, the recipient, the mode, the approver if one exists, whether the cap or allowlist affected it, and the reason for any hold or rejection. The OpenTelemetry documentation on observability describes this kind of structured record as the basis for understanding what a system did and why. For agents, the record also shows whether the agent's behavior is drifting over time.
Review the log on a regular cadence, not only after an incident. Look for three patterns: repeated holds to the same unknown domain, drafts that reviewers rewrite in the same way, and any attempt to exceed a limit. Each pattern points to a specific change, either to the instruction, the allowlist or the cap. Retire a permission when the task changes or the mailbox is no longer needed. Unused send permissions are a quiet risk because nobody is watching them.
How Be Human handles email limits
Be Human designs these controls as part of the agent's operating context rather than as a setting someone remembers to change. A first engagement defines the email task, the mailbox identity, the read scope, the draft and send modes, the caps and the allowlist, along with the success criterion that decides when a mode can change. Managed operations include monitoring of the send queue, recurring permission reviews, allowlist maintenance when contacts change, and controlled improvements to the agent's instructions based on the review record. The customer's team keeps control of which messages can leave the company, and the agent's limits are written down where that team can read and change them.
When drafts can become sends
Moving from draft to send is a business decision, not a technical one. Make it per message type, with a named owner who signs off on the change and a written record of the evidence. Keep the cap and the allowlist in place after the change, because a send-enabled agent still needs a ceiling. Recheck the decision when the agent's instructions change, when the mailbox receives a new kind of inbound message, or when the business relationship with a recipient changes.
If you are not sure which mailbox task should start in draft mode, the free Be Human AI diagnostic will help you find the first task worth handing to an agent.
Frequently asked questions
What is a good hourly send cap for an AI agent? A good cap is based on the observed normal volume for that specific role, with a modest margin above it. It should be low enough that a faulty loop is contained within a short period and reviewed by a person. Start conservatively, then adjust only from the log of real sends and holds.
Should an AI agent ever send email without human review? It can, but only for message types that have been reviewed in draft mode long enough to show consistent accuracy, and only inside a cap and an allowlist. Even then, the business should keep a named owner, a way to suspend sending, and a record of each message. Sending without review should be the result of evidence, not the default.
How do recipient allowlists differ from spam filters? A spam filter decides whether incoming mail is unwanted. A recipient allowlist controls where the company's agent is permitted to send outbound mail. The allowlist is written and owned by the company, and anything outside it is held for a person to decide, which is a stricter control than filtering.
What happens when the agent reaches its limit? The message should be queued and a named owner should be alerted, with the reason recorded. Messages should not be dropped silently, because lost work is harder to notice than a visible backlog. The owner then decides whether to release, rewrite or cancel the queued messages.
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.


