Separate a successful send request from evidence that a service update reached its required recipients. Use recipient-level status and explicit closure rules so an AI assistant cannot mistake transport progress for completed work.
An AI assistant drafts a site-access request, a coordinator approves it, and the email tool returns success. The service case closes. Later, a delivery failure arrives for the person who must confirm access. The workflow recorded a tool result, but never established the business outcome.
An AI email delivery-status workflow reconciles the sending attempt, transport evidence and outstanding business obligation separately. This guide proposes a design for operational messages, not bulk outreach. Its records, states and example are implementation recommendations—not a tested customer deployment or a claim that one platform supplies every control.
Define what must happen before the case can close
Start with the obligation, not the API. Does the process require an attempt to notify, acceptance by a recipient’s mail server, an explicit acknowledgment, or an authorized approval? These are different completion conditions. Name the required recipients and evidence before sending.
Microsoft Graph’s sendMail documentation says its successful response is 202 Accepted, without a response body. That confirms acceptance of the request, not completion of processing. An integration must not invent a returned message ID or a delivered status from that response.
Likewise, the Amazon SES event definitions distinguish a successful sending request, delivery to the recipient’s mail server and temporary delivery delay. Server acceptance is not evidence that a person read the message or authorized the requested work.
Write a closure predicate in business language: “The case remains open until the designated site contact confirms access for the approved visit.” Email delivery can support that process; it cannot satisfy the confirmation itself. Keep separate fields for transport evidence, acknowledgment and the case decision.
If the real decision is whether to stop tracking an unanswered request, use a separate no-response closure policy. The transport ledger here supplies contact evidence; it does not decide reminder counts, administrative eligibility or cancellation.
Keep AI outside the transport decision boundary
AI interprets the exception; rules establish its identity
Supply the assistant with permitted diagnostic text, the linked service request and known recipient context. It may propose a plain-language explanation, identify missing information or draft a reviewer note. Require evidence references and an unresolved category when the data is insufficient.
Deterministic code validates event origin, parses provider-defined fields and correlates each event to an existing sending attempt and recipient. The model must not decide that “probably delivered” overrides missing evidence, change the destination address, lift a suppression restriction or release another send.
People authorize recovery; systems execute it
A named service owner decides whether to correct a contact, use an approved alternate channel, extend a deadline or close with a documented exception. A separately authorized executor performs any resulting action. An AI diagnosis is not authority to message a newly discovered address.
The broader AI email automation workflow covers classification, drafting and approval. This delivery layer begins at the approved sending attempt and ends with evidence-backed reconciliation or an assigned exception.
Use a recipient evidence ledger, not one “sent” flag
Persist the authorized intent before submission. Keep the business message separate from its transport attempts, and track each required recipient independently. A success for a copied observer must not hide a failure for the decision-maker.
| Record | Required evidence | Control |
|---|---|---|
| Business obligation | Case ID, required recipient roles, approved content revision, closure condition and owner | Transport progress cannot rewrite the obligation. |
| Sending attempt | Stable local operation identity, provider/account scope, approved destinations, submission time and observed response | Keep acceptance, rejection and unknown outcome distinct. |
| Provider correlation | Provider message identifier where actually exposed; documented mapping to local intent | Do not equate local IDs, email headers and provider IDs. |
| Recipient event | Recipient, event identity where available, event type, provider time, receipt time and protected original reference | Preserve evidence; deduplicate according to the actual event contract. |
| Recovery decision | Reason, reviewer authority, approved destination/content and related attempt | A changed recipient or message is a new authorized intent, not a hidden retry. |
| Closure | Predicate evaluation, acknowledgment or exception evidence, decision-maker and time | Retain unresolved transport or business questions explicitly. |
SES notification contents document a provider-assigned mail.messageId distinct from the original message’s header identifier. Notifications can describe multiple recipients, and ordering and batching are not guaranteed. A delivery notification can also be followed by a bounce. Design the reducer for those documented conditions instead of assuming one ordered event per email.
If your connector exposes only a send acknowledgment and no supported delivery evidence, record that limitation. Use the approved manual evidence path or a closure condition that genuinely requires only attempted notification. Do not relabel missing telemetry as delivery success.
Make recovery decisions from evidence, not elapsed time alone
| Observation | Record | Next action |
|---|---|---|
| Request accepted | Accepted; recipient outcome not established | Await configured evidence. Do not close an acknowledgment-required case. |
| Provider reports temporary delivery delay | Delayed for the affected recipient | Check provider retry ownership and the business deadline; avoid a competing resend. |
| Recipient server accepts the message | Server-accepted evidence | Apply the actual closure condition; continue awaiting any required human answer. |
| Permanent rejection or final failure | Delivery exception with diagnostic evidence | Assign contact recovery; do not repeatedly send to the same failed destination. |
| Submission times out | Unknown submission outcome | Reconcile using supported provider records before considering another submission. |
| No correlated event by the review deadline | Evidence overdue, not proven failure | Investigate telemetry and provider records; keep the obligation owned. |
| Late contradictory evidence | Retained earlier evidence plus new exception | Re-evaluate communication-dependent closure under the agreed policy. |
The review deadline is an operational choice based on the consequence of silence and your provider’s documented behavior. It is not a universal email delivery timeout. Distinguish a provider still trying delivery from a request that your application is eligible to retry.
Follow the workflow retry discipline: recover the failed step only when its execution outcome and retry eligibility are known. A local stable key helps prevent duplicate dispatch within your application; it does not create provider-side idempotency where none is documented.
Process events safely and preserve late evidence
- Validate the incoming event. Apply the provider’s authentication and source restrictions before interpreting its content. For SNS delivery to HTTP endpoints, AWS requires signature verification and recommends rejecting unexpected topic identifiers.
- Persist before reducing state. Store a protected event reference and ingestion result. Validate schema and correlation fields. Quarantine unknown accounts, message IDs or recipients rather than joining by subject text.
- Apply the event once logically. Deduplicate repeated event delivery using a contract-appropriate identity. A state update must not automatically send another email as a replay side effect.
- Evaluate evidence, not arrival order. Retain provider and receipt timestamps. Do not downgrade a known outcome merely because an older progress event arrives later; do not discard a genuine later failure either.
- Reconcile overdue attempts. Use available, authorized provider records or an operator investigation. Where no supported lookup exists, keep the uncertainty visible instead of claiming a universal status-query API.
- Assign closure separately. Evaluate all required recipients and any business acknowledgment. Record why the condition passed or why an authorized exception was accepted.
For Graph specifically, Microsoft’s send-mail process overview explains that much of transport occurs after the method returns and that non-delivery reports can arrive later. Confirm what your integration can actually observe; do not assume Graph exposes the same event model as SES.
Restrict diagnostic access and retention. Bounce content may include addresses, excerpts or other personal information. Give the assistant only the fields needed for the exception, and treat free-text diagnostics as untrusted input, not instructions.
Walk a two-recipient failure through to closure
Illustrative example: case R-462 requires a site contact to confirm an access window. An approved email also copies the operations coordinator. The local attempt is A-1; transport evidence is stored separately for each recipient.
- The sending request is accepted. R-462 remains “Awaiting access confirmation,” not “Completed.”
- The copied coordinator’s mail server accepts its copy. The required contact’s delivery is delayed. The aggregate status shows a required-recipient exception rather than a single green success.
- A final failure arrives for the contact. The assistant summarizes the diagnostic with its source reference. It cannot substitute another address or send a new message.
- The service owner verifies an alternate contact through an approved record and authorizes revised communication. A-2 links to A-1 and the recovery decision; A-1’s history is retained.
- The verified contact explicitly confirms the window. The owner records that confirmation against the agreed closure condition. Delivery evidence alone did not establish access.
If late evidence arrives after closure, review whether the closure depended on the failed communication. An independent, verified confirmation may still satisfy the obligation. Otherwise reopen or escalate under the recorded policy. Neither automatic reopening of everything nor silent dismissal of late failures is a sound default.
Where Workhint fits
Workhint’s Workflow Engine describes routing approvals to defined roles and requiring details, files, notes or approvals before work progresses. Those documented controls are relevant to assigning a delivery exception and deciding what evidence a reviewer needs.
The workflow automation software solution describes configurable forms, conditional routing, human decisions and status changes. That is an operational-review connection—not proof of a native SES or Graph telemetry integration. Verify event ingestion, identifier mapping, recovery actions and evidence access in your proposed implementation before relying on them.
Test the closure rule, not just the send tool
Build permitted fixtures for accepted-without-delivery, a required recipient failing while a copied observer succeeds, duplicate and out-of-order events, delivery followed by bounce, an unknown submission outcome, forged event input and an uncorrelated report. Include missing telemetry and a verified acknowledgment that arrives through an approved alternate channel.
For each fixture, specify the expected recipient evidence, case status, reviewer and allowed next action. These are proposed acceptance tests, not measured results. Track false closures, overdue unknown attempts, event-correlation failures, duplicate dispatches and time to an owned exception decision. Measure AI diagnostic accuracy separately from transport-state correctness.
Frequently asked questions
Does a successful send API response prove delivery?
No. Interpret the response using that provider’s documented contract. Graph’s accepted response is explicitly not completion, and SES distinguishes send and delivery events.
Should AI decide whether a bounce deserves a retry?
AI may explain the diagnostic. Rules and the responsible owner determine retry eligibility, provider retry ownership and authorization. Unknown execution outcomes require reconciliation, not another model-generated send.
Can an email delivery event close a service request?
Only when the agreed closure condition is genuinely satisfied by that evidence. A request requiring acknowledgment or approval remains open until the required decision or an authorized exception is recorded.

Leave a Reply