Salesforce Einstein Copilot was a conversational AI assistant built into Salesforce CRM, but its defining idea was bigger than chat: connect company data to a system that could select and carry out permitted business actions. Salesforce announced its public beta on February 27, 2024. The product lineage is now called Agentforce, and the older Agentforce (Default) path is no longer receiving new features.
What Einstein Copilot was
Einstein Copilot was Salesforce’s conversational assistant for CRM users. It was designed to answer questions, summarize records, draft content, interpret conversations, and help automate work in sales, service, marketing, commerce, and related workflows. Salesforce presented it as a combination of a conversational interface, a large language model (LLM), business data and metadata, and executable actions—not simply a general-purpose chatbot placed beside CRM.
At its February 2024 launch, Salesforce announced a public beta for Sales Cloud and Service Cloud. Its launch announcement described plans to extend availability to Commerce Cloud and Marketing Cloud later that year. The launch-era announcement also specified English-language support and U.S. data residency; those are historical launch details, not a statement of Agentforce’s current availability. Salesforce’s launch announcement sets out that original positioning.
What Salesforce meant by “reasoning”
In Salesforce’s description, a reasoning engine interpreted a request, considered its context and available business information, selected an action or sequence of actions, and used the LLM to help produce a response or action plan. “Reasoning” here describes a product workflow; it is not evidence of human-like thought, general judgment, or reliable performance on unfamiliar tasks.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
#1 Best Overall
A useful way to understand the system is to separate its jobs:
- Generation: Draft text, summaries, or recommendations.
- Retrieval and grounding: Find relevant CRM, knowledge, or connected business information to inform a response.
- Planning: Select and sequence available actions to address the request.
- Execution: Invoke a permitted action that may read or change data, send content, or start a workflow.
- Governance: Apply access, approval, and audit controls to what the system can see and do.
For example, Salesforce described a seller asking which product tier to recommend. The assistant could examine a customer’s existing products, consider upgrade options, and update information across Salesforce and other systems through tools such as Flow and MuleSoft. That example illustrates the intended chain from request to action; it does not establish that every organization’s data or configuration will support it.
Why actions matter more than chat alone
A text-only chatbot returns an answer. A drafting assistant can prepare a summary or email for a person to use. An action-taking agent can also invoke a configured capability—such as looking up a record, updating it, or starting a workflow. These categories overlap, but the distinction matters: an action can change business records or affect a customer, so the capability must be deliberately exposed, configured, permissioned, and tested.
Salesforce’s launch examples included summarizing records, drafting emails, querying information, updating records, closing a case, creating a sales opportunity, and supporting an add-on sale. Current Salesforce documentation lists standard action examples including answering questions with Salesforce Knowledge, drafting or revising email, extracting fields from user input, getting record details, identifying objects and records, querying records and aggregate data, searching the web, summarizing and updating records, verifying customers, and speech-to-text capabilities. Availability depends on the edition, add-on, agent type, and product. See Salesforce’s Agentforce action reference.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsIn practice, an agent cannot perform an action just because a model understands the request. The action has to exist and be available to that agent, and the user or integration must have the required permissions. Custom work may involve Salesforce Flow, Apex, MuleSoft, or APIs.
How business data grounds responses
Einstein Copilot was intended to work from business context rather than rely only on information encoded in a general-purpose model. Salesforce described grounding prompts in Data Cloud, now commonly branded Data 360. Salesforce’s current product material describes Data 360 as connecting and harmonizing Salesforce and other data, including structured and unstructured information.
Depending on the implementation, useful context may include CRM records and metadata, knowledge articles, conversation transcripts, or external data connected through Data 360. Semantic search and vector-based retrieval can help find relevant unstructured content. Flow, Apex, MuleSoft, and other integrations can connect the assistant to processes and systems beyond a CRM record. The underlying design still matters: poor-quality or conflicting records, stale knowledge, incomplete integrations, or overbroad permissions can undermine the result. Grounding can make an answer more relevant; it does not guarantee that the answer is true.
Salesforce’s current overview of the product lineage and its data approach is at Agentforce Assistant, formerly Einstein Copilot. Data 360 details are at Salesforce Data 360.
Examples by business role
Sales
- Summarize an account, opportunity, or prior customer interactions.
- Draft a personalized follow-up message or query call transcripts.
- Support recommendations about next steps or product tiers.
- Update CRM records or support forecasting and closing plans through configured actions.
Service
- Find relevant knowledge, summarize a case, or draft a response.
- Verify a customer and update case or customer records.
- Start a follow-on service or sales process when the configured workflow permits it.
Marketing and commerce
Salesforce describes potential uses such as creating campaign briefs and content, drafting email campaigns, personalizing promotions, supporting storefront work, and generating product descriptions or SEO metadata. These are vendor-described capabilities, not independently established productivity results. Salesforce’s product page provides its current examples.
Financial services, healthcare, and other industries
Salesforce’s examples include capturing customer details, looking up transactions, requesting fee reversals or provisional credits, updating patient or member information, and preparing outreach. These workflows may affect financial or health-related outcomes. They need industry-specific validation, restricted permissions, and appropriate human review; general AI controls are not a substitute for compliance and operational review.
Trust controls—and what they do not promise
Salesforce says Agentforce is integrated with the Einstein Trust Layer and respects standard Salesforce access controls. Its Trust Layer documentation describes measures including zero-data-retention handling with third-party LLM providers, personally identifiable information (PII) masking, toxicity scoring, protection against unauthorized access, customer-configured masking, and audit and feedback data stored in Data 360 for reporting and alerts. The exact data path and controls can vary by feature and configuration. Salesforce’s descriptions are available in its Trust Layer documentation and data-usage documentation.
These measures are not guarantees of correct answers or safe outcomes. They do not fix a badly configured permission model, prevent an ill-designed action from doing something undesirable, or replace approval policies for financial, legal, medical, or customer-impacting decisions. Nor should one assume that every Salesforce AI feature uses the same model, retention policy, or data path.
Recommended Free Tools
Current name and product lifecycle
Current naming: Salesforce renamed Einstein Copilot for Salesforce to Agentforce. Salesforce said the change of name did not alter functionality at the time. Its product page identifies Agentforce Assistant as formerly Einstein Copilot. See the Salesforce release note.
There is also an important lifecycle distinction. Salesforce says that starting June 17, 2025, Agentforce (Default) would receive no new features or improvements and would not be available in new environments. Salesforce recommends that existing implementations migrate to Agentforce Employee. An organization may still have the older path, so administrators should confirm which agent type it runs before planning new work. Salesforce’s Agentforce considerations explain lifecycle and implementation details.
Current platform constraints to plan around
Salesforce’s current documentation says the reasoning engine supports OpenAI GPT-4o for reasoning calls, with Anthropic models through Amazon Bedrock available as an alternative provider in supported scenarios. Model choice is not universal: the documentation distinguishes the reasoning engine from custom actions and uses such as prompt templates, so bring-your-own-model support should not be assumed for every operation.
The same documentation gives these engineering limits:
- Agent actions time out after 60 seconds.
- Reasoning-engine requests time out after 30 seconds.
- Action outputs exceeding 65,000 characters are truncated.
These limits can matter for slow external services, long-running flows, large responses, or plans with several dependent steps. Agentforce should not be treated as a batch-processing or long-running transaction engine without designing around those constraints. For work that may outlast a request, a workflow may need to hand off to an asynchronous process and report its status separately.
Availability, licensing, and usage costs
The launch-era availability and licensing statements should not be used as a current quote. Salesforce’s current documentation lists Lightning Experience availability across Enterprise, Performance, Unlimited, and Developer Editions, while required add-on licenses vary by agent type. Confirm the exact combination of Salesforce edition, cloud, agent type, Data 360 entitlement, region, language, and action licensing with Salesforce before committing to a deployment.
AI usage is not necessarily one flat per-user charge. Salesforce documents consumption-based, hybrid, and business-metrics-based billing, with metering that can include prompts, actions, conversations, Einstein Requests, or Flex Credits depending on the product and commercial arrangement. Its documentation does not establish one universal public price for the Einstein Copilot/Agentforce Assistant use case. Review the applicable contract and estimate usage from the intended workflow rather than assuming a single pricing unit. See Salesforce’s AI usage and billing documentation.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Implementation risks worth testing
Before enabling actions for users, test how the agent behaves in realistic and failure-prone cases—not just clean demonstrations. Salesforce notes that agents are optimized for specific topics and may perform poorly on open-ended or underspecified requests.
- Ambiguous requests: Require clarification when the intent, record, or desired outcome is unclear.
- Wrong record resolution: Test names shared by multiple accounts, contacts, opportunities, or cases; add confirmation where a mistaken match could matter.
- Permission mismatch: Check what the invoking user and any integration credentials can access, and audit action permissions.
- Bad or conflicting data: Include stale records and contradictory knowledge in tests; plausible wording does not prove the underlying facts are reliable.
- Overpowered custom actions: Restrict inputs and add approvals for actions that can edit broadly, send messages, or affect customer outcomes.
- Partial execution: Verify that the user and audit trail can show which steps completed when a multi-action plan fails midway.
- Timeouts and truncation: Exercise slow integrations and large outputs against the documented limits.
- Streaming and safety checks: Salesforce says some Trust Layer checks apply only to the final response; if a hallucination is detected after streaming begins, the response may be deleted and regenerated.
- Deactivation: Plan for how ongoing conversations are handled if an agent is turned off.
- Model and product changes: Monitor provider availability, routing, limits, and agent lifecycle because they can change independently of the surrounding interface.
- Regulated decisions: Keep domain-specific validation and human approval where law, policy, or customer impact requires them.
For external-system work, Salesforce Flow and MuleSoft are possible integration surfaces, but adding them brings design and maintenance work. See Salesforce Flow and MuleSoft.
Who is it a good fit for?
Agentforce is most compelling when Salesforce is already central to the work and the goal is to do more than generate text—such as finding information, updating CRM records, or invoking governed workflows. The case is stronger when the organization has usable data, reliable knowledge content, administrators or architects to configure actions, and a clear approach to approvals and monitoring.
It may be a poor fit for a company that does not use Salesforce as a system of record, needs a low-cost standalone chatbot, lacks the staff to maintain Salesforce configuration, or needs long-running work beyond the documented request limits. It is also a harder fit if the organization requires every model to be independently selected or fully self-hosted, or needs predictable per-user budgeting for a heavily consumption-based workload.
How it compares with alternatives
These options are platform-fit comparisons, not claims of feature or pricing equivalence. Start with the system where the work and permissions already live:
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Quick Recap
| Option | Most natural fit | Key question |
|---|---|---|
| Salesforce Agentforce | Sales or service work centered on Salesforce CRM and its actions. | Do your data, licenses, integrations, and governance support the workflow? |
| Microsoft Copilot Studio or Microsoft 365 Copilot | Organizations centered on Microsoft 365, Teams, Power Platform, and Azure. | Can the required Salesforce records and actions be integrated and governed appropriately? |
| Google Gemini for Workspace | Organizations standardized on Gmail, Docs, Sheets, Meet, and Google Workspace. | Will separately connecting Salesforce data and CRM workflows add too much friction? |
| ServiceNow AI | IT service management, employee workflows, and service operations built on ServiceNow. | Is the target work in ServiceNow rather than a Salesforce sales process? |
| HubSpot Breeze | Organizations already centered on HubSpot CRM and marketing automation. | Does it meet the needs of a highly customized, multi-cloud Salesforce environment? |
| A custom LLM/API stack | Organizations that need more control over model, hosting, orchestration, or user experience. | Can your team build and operate retrieval, permissions, tool execution, monitoring, audit, and safety controls? |
Questions to resolve before a pilot
- Which agent type is deployed, and is it the older Agentforce (Default) path?
- Is the needed capability included in the current Salesforce edition, or does it require an add-on?
- Will the workflow be metered by actions, prompts, conversations, Einstein Requests, Flex Credits, or another contractual measure?
- Does the intended grounding workflow require Data 360?
- Will custom Flow, Apex, MuleSoft, or API actions be needed?
- Which actions require human confirmation, and how will partial completion or failure be handled?
- Can administrators audit the relevant prompts, data access, and executed actions?
- Are the model provider, language, region, and compliance requirements supported for this specific configuration?
- Would a platform already central to the organization’s work—or a custom stack—be a better fit?
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
