Mailgun’s AI email pitch puts delivery plumbing first

An AI agent can decide what to send in a flash. Getting that message to a person is a separate job. In a post about email infrastructure for AI products, Mailgun puts login messages, isolated sending reputations and machine-readable replies at the center of its pitch to developers.

The adoption numbers explain the sales push: Mailgun’s Email Impact Report says 79% of senders already use AI or plan to, while 74% of high-volume senders already use some form of AI in production. Those are different measures—not evidence that three-quarters of large senders have handed their email programs to autonomous agents.

What happened

Mailgun is positioning its delivery infrastructure as a supporting layer for AI applications, from adding a large language model to an existing software workflow to running agents that act without waiting for a prompt. The post describes capabilities and performance claims, rather than announcing a dated product launch.

Its main architectural proposal is separation. Mailgun says its Multi-Tenant Engine uses subaccounts to isolate sender reputations, application programming interface (API) keys and dedicated internet protocol (IP) addresses. The intended benefit is to keep one high-volume autonomous campaign from damaging delivery of verification messages and one-time passwords, or OTPs.

That distinction matters because the post puts promotional campaigns and account-access messages inside the same AI infrastructure discussion. They have different consequences when delivery goes wrong: a delayed campaign is not the same product problem as a user being unable to finish signing in. Mailgun presents separate sending infrastructure as a way to contain that exposure.

Mailgun also describes the return path. It says its inbound processing captures replies and turns Multipurpose Internet Mail Extensions (MIME) data into structured JavaScript Object Notation (JSON) payloads. That gives an application a more usable representation of incoming email, allowing a model to process a response and continue a workflow.

For speed, the company says 99% of messages are delivered in 20 seconds or less. Despite a section heading that invokes sub-second delivery, the numerical claim offered is the 20-second measure. The post does not provide a detailed measurement methodology, so readers should not treat it as a universal promise about when a recipient sees a message.

The background

Mailgun’s argument connects AI productivity to the less glamorous work of email deliverability. Its report says 46% of teams using AI in their email programs experienced increased speed and efficiency. Faster generation or decision-making, however, measures a different part of the workflow from transporting the finished message.

The company draws another number from Sinch’s AI Production Paradox report: 84% of AI engineering teams said they spend more than half their time building safety controls, retries and infrastructure instead of core product features. That statistic groups several kinds of engineering work together; it does not establish how much of that time is specifically spent on email.

Autonomous sending adds a control question that ordinary drafting tools do not answer. Mailgun describes agents choosing when to send, what to send and whom to contact. Once those decisions move into software, subaccount boundaries become relevant to limiting the consequences of a particular tenant’s or workflow’s activity—not simply organizing billing or credentials.

Reply processing introduces a separate distinction: structured data is easier to consume, but it is not automatically trustworthy. Our earlier coverage of hidden instructions targeting AI email assistants explains why incoming message content deserves its own safeguards. Mailgun’s MIME-to-JSON description addresses formatting; it does not establish protection against malicious instructions embedded in replies.

The performance evidence also needs careful reading. Mailgun says 54% of senders that integrated AI saw significant year-over-year improvement and says they outperformed non-AI users by 15%. The post does not explain whether that comparison means a relative percentage or percentage-point difference, and the figures alone do not establish that AI caused the improvement.

What people are saying

The source is a vendor argument, not a collection of customer reactions. Author Phil Kellum, a marketing manager and copywriter at Sinch Mailgun, calls the engineering burden a “guardrail tax.” His position is that managed delivery services can free developers to work on models and product features rather than rebuilding communications infrastructure.

To support the commercial case, Mailgun cites Forrester figures: an average $594,000 revenue increase tied to improved email performance, a 65% lift in open rates and a 21% lower bounce rate associated with form-level email validation. The post does not supply the study design, customer sample or assumptions needed to apply those numbers to another business.

The open-rate figure needs particular restraint. As our coverage of what email open tracking actually measures explains, pixel loads are not a clean count of human attention. A reported lift in opens should not be read as an equivalent increase in people reading, purchasing or successfully completing an AI-driven task.

Mailgun’s support proposition is more concrete: its Technical Account Manager team handles guided IP warming and proactive reputation monitoring. The company also advertises a 99.99% system uptime guarantee. System availability, message delivery speed and inbox placement describe different outcomes; the uptime figure does not promise all three.

What happens next

There is no rollout deadline in the post. Mailgun’s next step is commercial: inviting developers to discuss a custom high-volume plan. It also says 74% of enterprises reported rolling back live AI agents because of operational failures, but provides no breakdown showing how many failures involved email.

For teams evaluating the pitch, the useful next step is a workflow-level test rather than assuming the headline percentages will transfer. Measure account-access messages separately from campaigns, check what subaccounts actually isolate, and test how replies enter the model. Those checks map directly to the delivery, separation and inbound-processing capabilities Mailgun describes.

The scope of outsourcing deserves equal attention. Mailgun promotes compliance support alongside delivery tooling, but its post does not demonstrate that purchasing infrastructure removes the need for application-level safety controls. Developers still need to decide which sending actions an agent may take and what should happen when a reply or delivery attempt goes wrong.