Salesforce Agentforce Specialist key terms

Salesforce's platform for creating and running autonomous AI agents that interpret a request, decide on steps, use company data for context, and execute business tasks through configured topics and actions.

A configured AI assistant inside Salesforce that has a defined job, a set of skills it may use, and behavior guidance, and that can plan and carry out multi-step work on behalf of users or customers.

The reasoning component behind Agentforce agents. It reads the user's request, picks the most suitable area of expertise, plans the steps, selects and runs tools, then drafts the reply.

A bounded area of responsibility given to an agent, such as order status or returns, packaged with a description for matching, behavior guidance, and the specific actions allowed within it.

A single callable capability that lets an agent do real work, for example running an autolaunched Flow, invoking Apex, executing a prompt template, or searching knowledge articles.

Plain-language rules and guidance that shape how an agent behaves, including tone, boundaries, escalation rules, and formatting, written at the agent level and refined for each area of expertise.

A short natural-language summary attached to each topic that the reasoning engine compares against the user's request to decide which topic should handle the conversation.

A prebuilt customer-service agent template that answers questions using knowledge and CRM data, performs service tasks such as case updates, and hands off to human reps when needed.

Salesforce-delivered starting configurations for common roles, such as service, sales development, or coaching, that include ready-made topics and actions you copy and adapt.

The practice of feeding an agent relevant, current, trusted business information at request time so its answers reflect real records and documents instead of only what the model learned in training.

A technique where the system first looks up relevant records or document chunks, inserts that retrieved content into the prompt, and then has the model answer using that supplied evidence.

Salesforce's data platform that ingests, unifies, and stores structured and unstructured data, and supplies the harmonized profiles, metrics, and search indexes agents rely on for context.

A standardized, harmonized object definition in Data Cloud that maps data from many sources onto shared fields, so agents, insights, and segmentation work from consistent semantics.

A configured connection that continuously brings data from a source system, such as a CRM object, cloud storage, or connector, into Data Cloud for storage and later harmonization.

A Data Cloud structure that makes large volumes of text searchable, enabling fast retrieval of the most relevant passages so an agent can answer from source material.

A Data Cloud runtime component that searches an index or data graph for the records or passages most relevant to a request and returns them as context for a prompt or action.

A Trust Layer feature that pulls the specific CRM and Data Cloud data relevant to the current request and merges it into the prompt just before the model is called.

A metric computed inside Data Cloud from harmonized data, such as total spend or average handling time, that can be handed to an agent as a ready-made fact.

The Salesforce setup surface for assembling an agent: naming it, describing its job, adding areas of expertise with guidance, attaching tools, assigning its runtime user, and previewing conversations.

A versioned, reusable prompt with declared inputs and a chosen model, authored so an agent or Flow can generate consistent text grounded in the values passed in.

The Salesforce tool for writing, versioning, testing, and activating prompt templates, including merging record, Flow, or Data Cloud values into the prompt text.

An autolaunched Flow exposed as an agent capability. The Flow performs the logic, integrations, and record changes, while the agent supplies inputs and consumes the returned outputs.

A dedicated Salesforce user account that an agent runs as. Its profile, permission sets, and sharing rules decide which records and operations the agent can actually reach.

Setup tooling for registering, testing, and selecting the large language model used by prompts and agents, including connecting an external model provider of your choice.

A deployment surface, such as enhanced web chat, in-app messaging, SMS, or a messaging app, that lets end users hold real-time conversations with a deployed agent.

Agentforce tooling where you define scenarios with sample user messages and expected outcomes, run them in batches against an agent, and inspect scores and failures.

A saved scenario pairing a representative user message with the area of expertise and tools expected to handle it, used to check that an agent behaves correctly and keeps behaving correctly.

Agentforce telemetry that captures how each conversation unfolded, including which area of expertise was chosen and which tools ran, and writes that trace to Data Cloud for analysis.

The percentage of agent conversations that end without being handed to a human. It is a headline measure of self-service effectiveness, usually read alongside satisfaction and escalation reasons.

An in-builder simulated chat where a builder sends test messages to a draft agent, sees which path was taken, and adjusts guidance or tools before activating.

Salesforce's security architecture for generative AI, wrapping each model call with grounding, sensitive-data masking, zero retention, harmful-content scoring, injection defenses, and audit logging.

A Trust Layer step that finds sensitive values such as names, emails, or card numbers in a prompt, replaces them with protected placeholders before sending, and restores them in the reply.

A contractual and technical guarantee that the external model provider does not keep prompts or completions, so customer data never persists outside Salesforce for training or reuse.

Trust Layer scoring of model output for abusive, hateful, or otherwise harmful language, with configurable thresholds that can flag or block a response before the user sees it.

Protections that spot attempts to hijack an agent through crafted user input or poisoned retrieved content, and then block, sanitize, or refuse the request instead of following it.

Retained logs of agent interactions and model calls recording prompts, responses, and related metadata, giving administrators and compliance teams evidence of what was sent and returned.

Designing an agent's user, permission sets, and data access so it reaches only what its assigned duties require, limiting damage if the agent is manipulated or misconfigured.