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

Tools & integrations

Email Bounce Handling for AI Outreach Agents: Detect, Suppress, Log

Learn how an outreach agent should detect hard and soft email bounces, suppress addresses, and log each decision to protect your sending domain.

Be Human typographic cover with the title: Email Bounce Handling for AI Outreach Agents: Detect, Suppress, Log
Conceptual illustration created with AI · Be Human

An outreach agent should handle email bounces by reading the delivery event for every message, separating hard failures (the address will never accept mail) from soft failures (the mailbox or server is temporarily unavailable), suppressing hard failures permanently, retrying soft failures on a capped schedule, and writing each decision to a log that a person can audit. Handled this way, the sending domain keeps a clean record with mailbox providers, and the team can see why every address was contacted, paused or removed.

Key takeaways

  • Classify every bounce as hard or soft from the provider's status and diagnostic text, never from the fact that the send call returned successfully.
  • Suppress hard bounces at the contact level and across every campaign, so no sequence or agent run can email that address again.
  • Retry soft bounces only on a short, capped schedule, then move the contact to a held state that a person reviews.
  • Write one log entry per decision that records the signal, the rule applied and the action taken.
  • Track bounce and complaint rates per sending domain, and let the system pause sending before the domain is damaged.

What separates a hard bounce from a soft bounce

A hard bounce is a permanent rejection. The receiving server says the address does not exist, the domain has no mail service, or the recipient is not accepting mail from that sender. A soft bounce is a temporary rejection: the mailbox is full, the server is busy, the message is too large for the moment, or a temporary DNS or greylisting condition is in place. The exact codes differ between providers, so the classification should be a maintained mapping in configuration, not a guess made by the model during each run.

There is a third category that many teams miss. Some rejections cite sender reputation, policy or content. These do not mean the recipient address is bad. They mean the domain or the message is being filtered, and retrying the same message can make the reputation problem worse. Treat these as a stop signal for that domain or campaign, and escalate them to a person.

A practical decision guide for the agent:

  • Permanent address failure: suppress the address now and close the sequence for that contact.
  • Temporary failure with a mailbox or server cause: schedule a retry under the capped rule.
  • Rejection that cites reputation, policy or content: pause sending from the affected domain or campaign and escalate.
  • Unrecognized or missing code: treat it as soft for one retry, then escalate if it repeats.

Read the delivery signal, not only the send result

A send call that returns accepted only means the message was handed to the sending service. Delivery and bounce outcomes arrive later, as asynchronous events. The agent needs a reliable way to receive them, either through webhooks from the sending provider or through periodic polling of a delivery log. The choice affects latency and failure behavior, and the trade-offs are covered in webhooks vs polling for AI workflows.

Two details matter more than they seem. First, verify that each event really comes from the provider before the agent acts on it, because a forged bounce event could be used to suppress real customers or to hide a problem. Second, expect duplicates. Providers may deliver the same event more than once, so the handler must recognize an event it has already processed and skip it. The same principle applies to every action the agent takes after an event, as described in idempotency for AI agents.

Suppress hard bounces everywhere, and make it permanent

A suppression list only works if every sender consults it. The list should sit in a shared store that the campaign tool, the outreach agent, any manual import and any integration with the CRM all read before sending. If one path ignores it, the bounce will repeat from that path, and the domain will pay for it.

Suppression should survive imports. When a new spreadsheet or CRM export contains an address that was previously suppressed, the import should not silently re-enable it. The agent should be able to add suppressions, but removing one should require a person, a recorded reason and a timestamp. This follows the same logic as limiting write access in any system: the agent gets the narrowest permission that does its job. The mailbox permissions involved are discussed in Gmail and Outlook access for agents.

Spam complaints deserve the same treatment as hard bounces. A recipient who marks a message as spam should be suppressed for the same campaign and, in most cases, for all marketing mail, because a repeat sends the signal that damages reputation.

Retry soft bounces on a capped schedule

Soft bounces are worth one or two retries, but only on a schedule that gives the temporary condition time to clear. Retrying immediately usually produces the same failure. Retrying indefinitely turns a full mailbox into a stream of attempts that look like spam. The retry interval should lengthen after each failure, and the maximum number of attempts should be fixed in configuration, so that changing it requires a deliberate edit.

When the cap is reached, the agent should not quietly drop the contact or keep trying. It should move the contact to a held state, stop the sequence, attach the last error text, and place the item in a review queue owned by a person. Scheduled retries of this kind need their own alerting so that failures reach someone, a pattern covered in scheduled AI agent tasks.

Illustrative scenario: a wholesale distributor's outreach agent sends a quote follow-up to a buyer whose company mail server returns a temporary busy response. The agent logs a soft bounce and retries later that day. The next morning it retries again. After the final allowed failure, it marks the contact as held, stops the follow-up sequence, and places the item in the sales operations queue with the error text attached. Nobody has to search sent folders to learn what happened, and the buyer is not contacted again until a person decides how to proceed.

Log every bounce decision in a form a person can audit

The log is what turns bounce handling from a black box into something an operations leader can trust. Every decision should be recorded at the moment it is made, not reconstructed later. A useful entry contains the following:

  • The contact identifier and the address, stored under the same data retention rules as the rest of the outreach records.
  • The timestamp and the provider's event identifier, so the entry can be matched to the source.
  • The raw bounce class and the diagnostic text returned by the provider.
  • The rule applied, identified by name and version.
  • The action taken: suppressed, retried, held, escalated or ignored as a duplicate.
  • The actor that made the change, including the agent run identifier or the person who approved it.
  • A reference to the original message, so a reviewer can see what was sent.

The log should follow one event from the webhook to the final state. Tracing that path end to end is the same idea that the OpenTelemetry documentation describes for observability: you need to see inputs, decisions and outputs together to explain a result. Keep the log readable without special tools, because the person reviewing a complaint may not be the engineer who built the pipeline. Also limit the personal data in each entry to what the review needs, and set retention periods with your legal or privacy owner, since data protection rules in Brazil and the European Union both expect that personal data be kept only as long as necessary.

Protect the sending domain with thresholds and automatic pauses

Single bounces are normal. The risk comes from patterns: a rising share of hard bounces in a campaign, a list imported from an old source, or a sudden jump in complaints after a message change. Thresholds for these patterns should be calculated by the system, not estimated by the model. The agent can report that a rate is high, but the decision to pause sending should follow from a deterministic check against a limit that the operations owner approved.

When a threshold is crossed, the system should pause sending from the affected domain or campaign automatically, notify a named person, and keep the pause in place until someone reviews it. The review should check the list source, the recent message changes, and the domain's authentication setup for SPF, DKIM and DMARC, which a technical owner needs to confirm. The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) describes this kind of measuring and managing of risk as an ongoing activity, which fits a bounce rate that changes every week.

Before going live, test the bounce classifier on a sample of past events. The [OpenAI evals guide](https://platform.openai.com/docs/guides/evals) describes the general practice of checking agent behavior against expected outcomes, and the same idea applies here: the classifier and the rules should produce the same actions a careful operator would take on a known set of events.

How Be Human's managed approach handles bounces

In a Be Human engagement, bounce handling is defined as part of the first useful role, not added afterward. The classification mapping, the retry limits, the suppression scope and the domain thresholds are written down with the operations owner. The agent receives permission to add suppressions and to pause its own sends, but not to remove suppressions or change thresholds. Each decision is logged in the same record the reviewer uses. In managed operations, the team runs recurring evaluations on a sample of bounce events, reviews permission and suppression changes, maintains the integration when the sending provider changes its events, and responds to incidents when a domain starts to degrade. If you want to find the first outreach task worth handing to an agent, the free Be Human AI diagnostic is a practical place to start.

Frequently asked questions

Can an outreach agent decide on its own that a bounce is permanent? It can classify a bounce when the provider's code is in the approved mapping, and it should suppress that address immediately. Codes outside the mapping, or rejections that mention reputation or policy, should go to a person. The mapping itself should be reviewed by the operations owner, not changed by the agent.

Should a soft bounce ever be treated as a hard one? Yes, when the soft bounce repeats until the retry cap is reached, the contact should move to a held state and stop receiving messages. Continuing to send to a contact that keeps failing adds risk to the domain without adding value. A person then decides whether the address should be corrected, suppressed or left alone.

Does the agent need access to the email provider's settings to handle bounces? It needs to read delivery events and update the suppression and held states in its own store. It should not need administrative access to the provider's domain settings, DNS records or account billing. Keeping that access separate makes it easier to review and easier to revoke.

What should happen to a sending domain after bounces start rising? The system should pause sending from that domain automatically once a pre-approved threshold is crossed, and notify a named person. The review should check the source of the recipient list, the recent message changes and the domain's authentication setup before sending resumes. Resuming should be a deliberate action with a logged reason, not a timer that expires on its own.

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 →