All articles

Agentforce

Agentforce proactive messaging: design the outbound workflow

Plan event-triggered Agentforce messages with channel eligibility, recipient controls, duplicate suppression and confirmed delivery states.

A delayed delivery can justify a status update without justifying an unrestricted customer conversation. Define the event and permitted message before building the agent.

Salesforce documents proactive Agentforce Service messaging for Bring Your Own Channel for specified messaging and contact-center channel configurations. That is a scoped product capability, not evidence that every Agentforce deployment can initiate messages on every channel.

Define who may receive what

Start with one event, such as a confirmed shipment delay, and the authoritative record behind it. Specify the intended recipient, channel, message purpose and conditions that prevent sending.

Check channel eligibility and your organization’s approved consent, preference and suppression rules before each send. Legal requirements depend on jurisdiction and message purpose; have the responsible policy owner approve the workflow rather than treating an agent prompt as the policy.

Minimize the information in the initial message. If a recipient replies asking for account-specific details or an action, perform the required identity and authorization checks for that next step.

Separate drafting from dispatch

The agent can prepare a message using approved facts. A deterministic send action should check the current record state, recipient eligibility and message identifier immediately before dispatch.

If the event has been resolved since drafting, suppress or revise the message. If the same event arrives again, recognize it rather than sending a duplicate. Keep the event identifier separate from a conversational session ID.

Distinguish queued, accepted by the provider, delivered where observable and replied-to states. Provider acceptance is not proof that a customer received or read the message.

Test the awkward transitions

Exercise a repeated event, a revoked preference, a stale address, a channel failure and a timeout with an uncertain send result. Decide whether each case is retried, reconciled or escalated.

A retry must not generate a fresh message identity for the same logical notification unless that behavior is intentional. Record enough status for an operator to resolve uncertainty without exposing unnecessary message content in logs.

Test replies as well as first messages. A customer may ask for an unrelated action, disclose sensitive information or request a human. The channel’s escalation path should preserve useful context without granting the agent new authority.

Begin with a limited event and a monitored recipient group. Expand after delivery handling, suppression and handoff work under realistic conditions. The customer-verification guide covers the next boundary when a conversation reaches protected account actions.

Keep reading

All articles