Salesforce Agentforce 360 for Public Sector AP key terms

Salesforce's platform for building, deploying, and managing AI agents that combine large language models with business data, logic, and actions so they can complete tasks instead of only answering questions.

A configured software worker that takes a goal, decides which steps and tools to use, carries them out, checks the outcome, and then replies or hands off, without a person directing each individual step.

An operating model in which an organization runs many AI agents alongside employees, delegating routine work across service and back-office processes while people stay accountable for exceptions and judgment calls.

The Agentforce component that reads a request, selects the right topic, plans and invokes actions, reviews what comes back, and repeats that loop until it can give a final answer.

A model trained on very large text collections that generates and interprets language; in Agentforce it reasons over a request and drafts responses but does not by itself know your agency's records.

The setup interface where an administrator defines an agent's topics, actions, instructions, and channels, then tests and activates it for users or constituents.

The classification of an agent by who it serves and how it runs, for example a service agent answering constituents on a messaging channel versus an employee agent working inside Salesforce on staff requests.

A named job an agent can handle, such as checking permit status; it bundles a description, the instructions for that job, and the actions the agent may use when a request matches.

The plain-language description that tells the reasoning engine when a topic applies; clear, specific wording improves routing accuracy and reduces the chance of the agent choosing the wrong topic.

A discrete capability an agent can invoke, such as querying a record, running a flow, or calling an API, defined with inputs and outputs so the reasoning engine can call it reliably.

An action backed by a Salesforce Flow, letting an agent run declarative automation such as creating a case or updating a record while respecting the org's existing logic and sharing rules.

An action implemented by an Apex class exposed as an invocable method, used when the logic needs code, complex calculations, or callouts that flows and standard actions handle poorly.

An action that sends a governed prompt template plus merged record data to a model and returns generated text, used for summarization, drafting, or classification within a topic.

The natural-language directions that set an agent's role, tone, rules, and boundaries; they shape behavior but must be paired with well-scoped topics and actions to stay dependable.

The typed definition of the values an action accepts and returns, described with names, types, and descriptions so the reasoning engine knows when to call the action and how to read its result.

A named value the agent can store and reuse during a conversation, such as a verified contact identifier or case number, so the person is not asked for the same detail twice.

Salesforce's data platform that ingests, harmonizes, and unifies information from many sources into a single profile, which agents then query to ground answers in current, trusted records.

A configured connection that brings data from a source system into Data 360 on a schedule or near real time; streams land raw data that is then mapped into the shared data model.

The storage representation of ingested data in Data 360, the raw table created by a data stream before any harmonization into a shared, business-aligned model.

A harmonized object in Data 360 that maps source data to a common business definition, such as a unified individual or case, so records can be joined and used for grounding.

The rule set that decides which records from different source systems describe the same person or entity, linking them into one unified profile using match and reconciliation rules.

A metric computed by Data 360, such as totals, counts, or recency, stored on a unified profile and available to an agent so answers reflect computed facts rather than raw fields.

A pattern where the agent first retrieves relevant records or document passages from a trusted source, then passes them to the model so the answer rests on that content rather than model memory.

A search method that compares numeric embeddings of text to find passages with similar meaning, letting an agent locate relevant policy documents even when the user's wording differs from the document.

Supplying an agent with authoritative data, such as records, knowledge articles, or retrieved passages, at the moment it answers, so responses are tied to verified sources and can be cited.

Salesforce's industry product set for government that adds data models and tools for licensing, permitting, inspections, grants, case management, and constituent service on top of the core platform.

A dedicated, tightly restricted Salesforce environment for public sector bodies, offering US-based data residency, stricter personnel controls, and compliance commitments suited to sensitive government data.

The declarative toolkit in Public Sector Solutions with OmniScripts, Flexcards, Integration Procedures, and DataRaptors, used to build guided service processes that agents can also invoke as actions.

A Public Sector Solutions feature that evaluates expression sets against record data to produce consistent decisions, such as eligibility or fee determination, without hand-written code in every place.

The Public Sector Solutions capability for issuing regulated licenses and permits, tracking applications, scheduling and recording inspections, and applying fees, conditions, and renewals.

The service capability where constituent requests become cases handled by agency staff, with queues, entitlements, and service console features that an agent can create, look up, and update.

The set of controls between Salesforce and a large language model that masks sensitive data before it leaves, checks responses, blocks harmful content, and logs each transaction without retaining the prompt.

A contractual and technical setting under which the model provider does not store the prompt or completion, so agency data is not used for training or kept after the response is returned.

Replacing or removing sensitive values such as names, identifiers, or health details before text goes to a model, then restoring them in the response, so the model never sees protected information.

An attack in which untrusted content or user input carries instructions meant to override an agent's rules, countered by separating instructions from data and limiting what the agent can do.

The Agentforce tool for building test suites of utterances and expected outcomes, running them against an agent version, and scoring topic selection and action accuracy before release.

A Shield capability that captures detailed logs of user and system activity, including agent runs, so agencies can audit who asked what, which actions executed, and whether policy was followed.

The permission model that gives an agent its own user and permission set, limiting the records and actions it can reach so it never exceeds what the requesting person is allowed to see.

A designed handoff where the agent recognizes low confidence, a sensitive request, or an explicit ask for a person, and transfers the conversation with context attached to a human queue.