AI Agents are customer-facing automation that can interpret requests, use knowledge and tools, perform permitted actions, and operate a Salted customer conversation. They participate in Live Conversations alongside customers, human agents, and external experts.
An AI Agent is not the same thing as the automation logic around the conversation. An AI Agent can be one part of that logic, alongside deterministic menus, rules, integrations, and workflows that do not require generative AI. In Salted configuration and technical documentation, this automation logic is called Your Logic.
What an AI Agent can do
Depending on the implementation, knowledge, tools, and permissions, an AI Agent can:
- understand a customer's free-text request,
- answer from approved knowledge and policies,
- ask follow-up questions and collect required information,
- present structured choices or next-step buttons,
- retrieve customer, order, account, booking, or case data,
- perform permitted business actions through connected systems,
- update conversation context and classification,
- route the interaction or ask a human for help,
- invite a specialist or external expert,
- complete eligible work and ask whether the customer's need was resolved.
The exact capabilities come from the configured tools and automation logic. The model should not be treated as having access or authority it has not explicitly been given.
Knowledge and tools
An AI Agent can use two broad sources of capability:
Knowledge
Knowledge explains products, policies, procedures, eligibility, tone, and other information the AI needs to answer correctly. It can be provided through configured content, skills, files, or connected knowledge sources.
Knowledge should be governed like an operational dependency. Conflicting or outdated instructions can produce inconsistent behavior even when the AI model is working as designed.
Tools and actions
Tools let the AI retrieve data or ask the surrounding automation logic to perform controlled operations. A tool can check an order, look up an account, create a cart, issue an allowed refund, send a payment or booking link, classify a conversation, or invoke another supported business capability.
Consequential actions should be protected by permissions, deterministic rules, limits, and human approval when appropriate. Tool availability is not the same as permission to use it in every situation.
Structured and free-text interaction
Salted can combine:
- free-text conversation,
- menus and customer-facing questions,
- dynamically selected follow-up options,
- files and supported media,
- channel-specific behavior.
A customer should not need to know whether a path is driven by a menu, a rule, or generative AI. The workflow should preserve context as the customer moves between these modes.
Working with human agents
An AI Agent can involve a person in several ways:
- request a bounded approval or decision,
- ask for human assistance when it cannot proceed safely,
- prepare context and a suggested reply for the person who joins,
- hand over an unstructured or sensitive interaction,
- remain silent while a human owns the conversation,
- continue after the human contribution when the automation logic knows how to proceed,
- remain visible to authorized supervisors, who can monitor the active conversation and proactively join or take over when intervention is needed.
This makes human involvement a configurable part of the workflow rather than a single one-way escalation path.
Operational control
A production AI Agent should operate within explicit boundaries:
- which knowledge sources it may use,
- which tools and actions it may call,
- which data it may access,
- which decisions require deterministic checks,
- which actions require human approval,
- how confidence and failure affect escalation,
- what the customer is told while waiting,
- how conversation state, reason, priority, and outcome are recorded,
- how new versions are tested before rollout.
Salted and the configured automation logic can provide the conversation and action controls. The organization remains responsible for defining the policy and system authority behind them.
Channel behavior
AI Agents can participate in supported digital and connected voice workflows, but the interaction model differs by channel. Chat can use menus and real-time replies, email requires thread-aware free text, WhatsApp can require templates, and voice has its own call lifecycle and latency requirements.
Confirm the production maturity and limitations of each channel and integration for the account.
Evaluation and improvement
AI-operated conversations can feed the same Customer Journey, Quality Intelligence, Search, Ask, analytics, and review workflows used for human interactions. Teams can inspect:
- whether the AI used correct knowledge,
- whether it called the right tool,
- whether it followed policy,
- whether escalation happened at the right time,
- whether a human correction should become a knowledge, workflow, or tool improvement,
- whether the customer outcome was actually resolved.
Technical documentation
- See How automation logic controls a customer conversation for how automation logic controls the conversation.
- See Conversations SDK for the Conversations SDK overview.
- See AI Setup for AI setup.
- See Intents for intents.
- See Conversations SDK Testing for Conversations SDK testing.
- See How AI and human agents work together in Salted for human-AI collaboration.
Availability
The packaged AI Agent offering, supported models, knowledge sources, tools, channels, testing, observability, and deployment responsibilities can vary. Confirm the current Salted-managed and developer-configured options before making a production commitment.