Salted lets AI agents, human agents, and external experts contribute to the same customer conversation without treating every human intervention as a permanent handoff.
The appropriate pattern depends on customer preference, confidence, policy, risk, available expertise, and the action being performed. Automation logic controls what happens next. It can combine rules, AI, workflows, and connected systems, then tell Salted which actions to perform. In Salted configuration and technical documentation, this automation logic is called Your Logic.
Collaboration patterns
AI-led conversation
An AI agent can answer questions, collect information, use connected systems, perform permitted actions, and complete eligible work. The workflow should still provide a recovery path when the request cannot be completed safely or correctly.
Human monitoring and proactive takeover
Authorized supervisors can monitor ongoing AI-led conversations through Live Conversations without waiting for automation to request help. They can inspect the active transcript, participants, alerts, routing, and available context without joining. When intervention is needed, they can join the conversation, take over customer communication, assign another human agent, or contribute only the judgment required by the workflow.
Monitoring and participation are different states. Viewing an active conversation does not create a human engagement. Joining does, and enables the person to communicate with the customer. The workflow determines whether AI pauses, remains silent, assists the person, or resumes after the human leaves.
Bounded human guidance or approval
AI can ask a human agent for a structured decision without requiring the human to take ownership of the whole conversation. The customer does not need to see an internal approval question. After the human responds, the workflow can communicate or execute the result and release the human from the interaction.
This pattern is useful when a request is automatable but one decision requires judgment, permission, or accountability.
Human joins or takes over
A human agent can join or own the interaction when the issue requires empathy, investigation, negotiation, exception handling, or an unstructured decision. Agent Desktop preserves the transcript and available customer context so the human does not receive a cold transfer.
External expertise
A partner or specialist outside the normal Salted user population can be invited into a selected conversation. This can bring the person closest to the problem into the workflow without routing every exception through a general support team.
Return to AI
Human involvement can end without ending the AI workflow. When the configured logic knows how to proceed, AI can continue after the human contribution. This is different from a conventional one-way escalation where the bot permanently disappears.
Guidance is not the same as handoff
Use guidance or approval when the human is needed for a bounded decision and AI can continue the process.
Use handoff or takeover when a person should own the next part of the customer interaction.
Use external expertise when another organization or specialist is best placed to resolve the issue.
The workflow may move through more than one of these states during a single conversation.
Example workflow
A customer requests a consequential account action:
- AI identifies the request and gathers the required information.
- Automation logic checks the relevant policy and connected account data.
- The action falls inside a range that requires human judgment.
- Salted presents an agent-facing approval question with the relevant context.
- The human approves, changes, or rejects the proposed action.
- The human contribution ends.
- AI communicates or executes the result and continues the conversation.
- The completed interaction remains available for quality review, coaching, and analysis.
The actual policy must be enforced through deterministic controls and connected systems. A prompt alone should not be treated as a hard operational boundary.
Context available to the human
Depending on account setup and the configured automation logic, Agent Desktop can show:
- the active transcript and available customer context,
- prior customer-journey information,
- suggested or prefilled replies,
- internal notes,
- information panels and links,
- standard and custom actions,
- routing, priority, reason, and outcome information,
- invitations to other agents or external experts.
Operational controls
A production workflow should define:
- which actions AI may perform independently,
- which decisions require a person,
- how the correct agent or specialist is selected,
- what happens when an invitation times out,
- what the customer is told while waiting,
- when AI should remain silent during human ownership,
- when and how control can return to AI,
- how the final result is classified and evaluated.
Related concept guides
- See What is Agent Desktop? for the human workspace inside Live Conversations.
- See Customer conversation lifecycle in Salted for the end-to-end customer conversation lifecycle.
- See How automation logic controls a customer conversation for the event, decision, and action model behind Your Logic.
Technical implementation
This article explains the operating model. It does not replace the technical guides. Configuration, API, and SDK documentation use the exact Salted labels and contracts.
- For conversation lifecycle behavior, see Live Conversation Lifecycle.
- For implementation guidance, see Your Logic Implementation Tips.
- For response patterns, see Your Logic Response Examples.
- For agent-facing questions, see Agent Facing Questions.
- For dynamic Agent Desktop controls, see Dynamic Agent Desktop.
Availability
The collaboration primitives available in a specific deployment depend on account configuration, channel, integration, and current product maturity. Confirm the required permissions, actions, events, and channel behavior in the linked technical documentation before production rollout.