Salesforce AI

Best for: The human-facing chat panel embedded in every Salesforce app. Renamed from Einstein Copilot in January 2025 to align under the Agentforce brand — same product, new name. It assists a logged-in user instead of acting autonomously.

• Side-panel chat in Sales Cloud, Service Cloud, Marketing, Commerce, Slack, and the mobile app
• Grounded in the record context the user is viewing — an opportunity, case, account, or contact
• Natural-language query of CRM data: "summarize this account," "draft a follow-up," "what's blocking this deal"
• Calls the same Agent Builder topics and actions as autonomous Agentforce agents — one definition, two surfaces
• Generative content (emails, summaries, knowledge articles) routed through the Einstein Trust Layer

Limitations: Quality of out-of-the-box answers is bounded by data hygiene — messy CRM data produces vague answers. The rename caused real confusion: Salesforce docs, training, and partner content still use "Einstein Copilot" interchangeably through 2026. Don't confuse with Einstein Copilot for Tableau, which is a separate analytics assistant.

Best for: Predictive scoring, forecasting, classification, and anomaly detection across CRM data — the non-generative AI that Salesforce has shipped since 2016. The "Einstein" brand still owns predictions and analytics; generative features have largely migrated to the Agentforce brand.

• Einstein Lead/Opportunity Scoring — ranks pipeline by likelihood to close
• Einstein Forecasting — revenue projections with confidence intervals
• Einstein Case Classification — auto-routes service cases by category and priority
• Einstein Discovery (now part of Tableau / CRM Analytics) — explainable model insights and what-if analysis
• Einstein Bots — the legacy rule-based / NLU chatbot product, being superseded by Agentforce Service Agent

Limitations: "Einstein" is now a portfolio name covering a dozen distinct features, each licensed differently — some bundled with Sales/Service Cloud, some sold as add-ons. The 2023 generative umbrella term "Einstein GPT" has been absorbed: trust-layer plumbing kept the Einstein name, generative agents moved to Agentforce. Expect to see both brands referenced in older docs.

Best for: Understanding why Salesforce thinks Agentforce produces fewer hallucinations than a raw LLM wrapper. Atlas is the orchestration layer that decomposes a user goal, retrieves grounded data from CRM and Data 360, plans a sequence of actions, and validates the output before responding.

• Implements inference-time "System 2" reasoning — iterative plan/act/reflect loops rather than single-shot prompting
• Advanced agentic RAG over Data 360, structured CRM records, and unstructured knowledge bases
• Topic classification routes the user request to the right Agent Builder topic and action set
• Salesforce reports 2x response relevance and 33% end-to-end accuracy gains in customer-service pilots vs. baseline LLMs
• Model-agnostic: works with Salesforce-hosted xGen, OpenAI, Anthropic, Google Gemini, and customer-bring-your-own models via Model Builder

Limitations: Not a separately licensed product — it ships inside Agentforce. The performance numbers come from Salesforce-run benchmarks; independent reproduction is limited. Reasoning quality still depends on the underlying LLM choice and the grounding data quality.

Best for: The data backbone that grounds every Agentforce agent and Einstein prediction. Renamed from Data Cloud in 2025; the previous "Customer 360 Data Cloud" and "Salesforce CDP" brands also fold in. Without Data 360, Agentforce has nothing to reason over beyond a single Salesforce org's records.

• Zero Copy federation — query Snowflake, Databricks, BigQuery, Redshift, and Iceberg tables in place without moving data
• Identity resolution & unified profiles — collapse duplicate customer records across systems into a canonical profile
• Vector database — embed and index unstructured content (PDFs, knowledge articles, transcripts) for RAG
• Calculated insights, segments, and activations — ship audiences to Marketing Cloud, ad platforms, or back into Sales/Service Cloud
• Consumption pricing: $500 per 100,000 credits (one fungible credit type since 2025), $23/TB/month for stored data, free ingest of structured Salesforce data
• Data 360 Starter packaged at $60K/year list for first-time adopters
• Profile-Based SKUs (active March 2, 2026): $240 / $420 per 1,000 profiles annually — predictable per-profile pricing as an alternative to consumption credits

Limitations: Cost can balloon — ingestion, profile unification, segmentation, agent grounding, and insight generation all consume credits. Forecasting spend before deployment is genuinely hard; even Salesforce ships a calculator. Zero Copy works best when your warehouse is one of the named partners; arbitrary databases require classic ingestion.

Best for: Authoring, testing, and reusing prompts across the platform without hardcoding LLM calls. Prompt Builder lets admins (not developers) define a prompt template once and call it from Lightning record pages, Flows, Apex, the Agent Builder, and the Agentforce Assistant.

• Template types: Sales Email, Field Generation, Flex (free-form), and Record Summary
• Merge fields pull live data from CRM records, Data 360 objects, related lists, and Apex methods — the prompt is grounded automatically
• Built-in playground for prompt testing with side-by-side model comparison
• Versioning, activation states, and language localization
• Outputs run through the Einstein Trust Layer (masking, audit, toxicity check) by default

Limitations: Prompt templates only call the model providers Salesforce has wired in (OpenAI, Anthropic, Google, Salesforce-hosted models). Free-form Flex templates have a learning curve for non-technical admins. Some advanced output formats (structured JSON, streaming) are easier to build directly through the Models API in Apex.

Best for: Two related capabilities inside Einstein Studio. Model Builder registers external LLMs (OpenAI, Anthropic, Google Vertex, Azure OpenAI, AWS Bedrock, or your own SageMaker / Databricks endpoint) so they can be used by Prompt Builder and Agentforce. The custom predictive Model Builder trains classification and regression models on Data 360 data with no code.

• BYOLLM — connect any OpenAI-compatible endpoint and use it inside Salesforce with the Trust Layer applied
• Predictive Model Builder — pick a Data 360 object, choose a target field, and train a model (churn, conversion, defect, ticket category)
• Hyperparameter tuning, automatic feature engineering, and explainability scores out of the box
• Deploy predictions back into CRM as fields, into Flows as decision inputs, or into Agentforce as agent actions
• Integrates with SageMaker, Databricks, Vertex AI, and OpenAI for hosted-model registration

Limitations: BYOLLM still routes through Salesforce's billing and Trust Layer — you can't entirely escape Salesforce-side costs. Predictive Model Builder is good for tabular CRM data but not a replacement for a real ML platform on complex pipelines. Studio is bundled with Data 360, so you'll need that license to use most of its features.

Best for: Letting regulated and security-conscious organizations actually deploy generative AI inside CRM. Every prompt and response from Prompt Builder, Agentforce, and the Agentforce Assistant flows through the Trust Layer — this is the defensible-architecture story Salesforce sells to legal and security teams.

• Dynamic data masking — PII (names, emails, phone numbers, account numbers, credit cards) is replaced with tokens before the prompt leaves Salesforce; the model sees masked data only
• Demasking on return — tokens are replaced with the original values inside Salesforce so the user sees a complete, useful response
• Zero data retention — contractual commitment that partner LLM providers (OpenAI, Anthropic, Google, AWS) delete prompts and responses immediately and never train on them
• Toxicity, bias, and safety scoring on every response before it reaches the user
• Full audit trail — original prompt, masked prompt, response, scores, and user feedback all logged for compliance review
• Secure LLM Gateway encrypts requests in transit and applies prompt-injection defenses

Limitations: Masking is pattern-based — novel PII formats or context-dependent sensitive data can slip through; review your data classification before relying on it for regulated workloads. Zero retention applies to the LLM provider but Salesforce still logs the audit trail. The Trust Layer is opaque — you trust Salesforce's implementation rather than verifying it directly.

Best for: The bridge between deterministic automation (Flow) and probabilistic reasoning (Agentforce). Any Flow can be exposed as an agent action — the LLM decides when to call it, the Flow decides how to execute it. This is how admins extend Agentforce without writing Apex.

• Auto-Launched Flows become callable agent actions with one click in Agent Builder
• Inputs and outputs are mapped to natural-language descriptions that the reasoning engine matches against user intent
• Einstein for Flow — AI-suggested next steps, auto-generated formula expressions, and natural-language flow creation ("build a flow that emails the account owner when an opportunity stalls")
• Prompt Builder templates can be invoked as a Flow action, embedding generative AI inside any business process
• 2026 best practice: build many small Flows the agent composes, instead of one monolithic Flow

Limitations: Flow remains synchronous and bounded by Salesforce governor limits — long-running or external-facing work belongs in Apex or external systems triggered via MuleSoft. Natural-language descriptions of inputs/outputs need careful authoring; vague descriptions cause the reasoning engine to call the wrong action.