Salesforce built Agentforce for agents that act on CRM and business data, not just chatbots that answer questions about it. The platform runs on the Atlas Reasoning Engine for planning and Data Cloud for grounding, and Salesforce’s own roadmap keeps adding reach beyond raw CRM records release after release, from Intelligent Context to a Data 360 MCP Server. Where that reach still stops, at metric definitions, cross-system lineage, and the other SaaS tools in a typical enterprise stack, is the gap Atlan’s context layer closes, alongside catalog vendors like Alation, Collibra, Informatica, and dbt Labs.
By 2026, Salesforce’s own metrics reported more than 18,000 companies running Agentforce agents in production, and the business reached an estimated $800 million annual revenue run rate. Salesforce also added Intelligent Context and a Data 360 MCP Server to ground those agents in unstructured data beyond raw CRM records. Atlan connects to that same MCP-native ecosystem through its own MCP server, adding the lineage, ownership, and business-term definitions that sit outside Data Cloud’s reach.
- Atlas Reasoning Engine: the cooperative-agent-swarm reasoning core behind every Agentforce action
- Data Cloud grounding: RAG-based retrieval plus Zero-Copy access into Snowflake and BigQuery
- Intelligent Context: Agentforce 360’s mechanism for ingesting unstructured data (PDFs, tables, images)
- Data 360 MCP Server: an open door for external context sources like Atlan’s own MCP server
- Enterprise Data Graph: Atlan’s lineage, ownership, and certification layer for everything Data Cloud doesn’t ingest
| Salesforce Agentforce: Quick Facts | |
|---|---|
| What it is | AI agent platform built on the Atlas Reasoning Engine and Data Cloud |
| Key benefit | Agents that research, plan, and act inside CRM and business workflows |
| Best for | Salesforce-native enterprises automating service, sales, and ops work |
| Launched | September 2024 (Dreamforce); Agentforce 360 reached GA October 13, 2025 |
| Cost range | ~$0.10 per action (Flex Credits) to $2 per conversation; free Foundations tier available |
| Core components | Atlas Reasoning Engine, Agent Builder, Data Cloud, Trust Layer, MCP support |
| Context gap | Reaches Data Cloud, connected apps, and Snowflake/BigQuery via Zero-Copy; metric definitions, cross-system lineage, and other SaaS still need an external context layer |
What is Salesforce Agentforce?
Permalink to “What is Salesforce Agentforce?”Salesforce Agentforce is an AI agent platform built on the Atlas Reasoning Engine and Data Cloud that lets autonomous AI agents handle service, sales, and operations work inside a company’s CRM estate. Salesforce launched Agentforce at Dreamforce in September 2024, positioning it as “digital labor” that works alongside human employees rather than a chatbot that only answers questions. Agentforce 2.0, released two months later, added the Atlas Reasoning Engine, the reasoning core Salesforce’s own engineering team describes as running cooperative agent swarms that plan, act, and self-correct across multi-step tasks.
By 2026, Agentforce had scaled well past a pilot product. Salesforce’s own Agentforce Metrics page reports more than 18,000 companies across 121 countries using the platform, part of the broader “Agentic Enterprise” push Salesforce announced with Agentforce 360’s general availability in October 2025. The business reached an estimated $800 million annual revenue run rate by the end of fiscal 2026, even as investor skepticism about the broader AI bet weighed on Salesforce’s stock. That adoption curve is a useful proof point for how fast an enterprise buys into agentic tools once the underlying AI agent architecture is proven in production, the same pattern shaping how enterprises use AI agents more broadly and the one documented in how to build an AI agent tech stack.
Agentforce sits in a similar competitive lane as Gemini Enterprise from Google, AWS Bedrock for enterprise agents from Amazon, and the frameworks compared in AWS Bedrock Agents vs. LangGraph: a cloud vendor’s own agent platform, tightly coupled to that vendor’s data estate, competing to be the default place enterprises build agents. Each platform is strong within its own boundary and faces the identical governed-context question once an agent needs data the vendor’s platform was never built to hold.
| Version | Date | Key addition |
|---|---|---|
| Agentforce 1.0 | September 2024 (Dreamforce) | Initial launch; “digital labor” framing |
| Agentforce 2.0 | December 2024 | Atlas Reasoning Engine |
| Agentforce 3.0 | June 2025 | Command Center observability; MCP support |
| Agentforce 360 | October 13, 2025 (GA) | “Agentic Enterprise” platform; Intelligent Context; Data 360 MCP Server |
How does the Atlas Reasoning Engine and Agent Builder work?
Permalink to “How does the Atlas Reasoning Engine and Agent Builder work?”Agentforce’s architecture runs on four layers: the Atlas Reasoning Engine, Agent Builder and Agentforce Studio for building agents, an Action Layer for taking real steps, and a Trust Layer for guardrails. Each layer handles a distinct part of how an agent reasons and acts, and together they determine whether an agent behaves like a reliable coworker or an unpredictable script.
Atlas Reasoning Engine
Permalink to “Atlas Reasoning Engine”The Atlas Reasoning Engine is Agentforce’s reasoning core. Phil Mui, Senior Vice President and Head of Products and Architecture at Salesforce AI Research, describes it as running “cooperative agent swarms” that “reason deliberatively with self-reflection toward specific tasks.” In practice, that means an Agentforce agent doesn’t just retrieve an answer and stop. It plans a sequence of steps, checks its own output against the goal, and revises the plan when a step fails, the same self-correcting loop that separates a genuine AI agent from a single-shot chatbot response.
Agent Builder and Agentforce Studio
Permalink to “Agent Builder and Agentforce Studio”Agent Builder is Agentforce’s no-code interface for defining what an agent can do: topics, instructions, and the actions it’s allowed to take. Agentforce Studio extends that into a fuller workbench for testing, versioning, and monitoring agents once they’re in production, similar in spirit to how teams treat enterprise-ready AI agents as products with a lifecycle, not one-off scripts.
The Action Layer (Flows, Apex, MuleSoft)
Permalink to “The Action Layer (Flows, Apex, MuleSoft)”Once an agent decides what to do, the Action Layer is what lets it actually do it. Agentforce agents can trigger Salesforce Flows, run Apex code, and call out through MuleSoft connectors to systems outside Salesforce, giving agents real tool use rather than the ability to only answer questions. This is also where ungoverned or duplicate CRM records do the most damage: an agent that acts through a Flow on a record with a duplicate ID or a stale owner field propagates that error downstream immediately, a failure mode covered in more detail below.
The Trust Layer
Permalink to “The Trust Layer”The Trust Layer sits underneath the other three layers, applying data masking, audit logging, and toxicity and bias detection to every agent action. It’s Salesforce’s answer to the governance question every enterprise asks before letting an agent touch production data: what did it do, and why. That question mirrors what enterprises already ask about agents generally, regardless of which platform runs them.
Build your AI context stack
Get the blueprint for implementing context graphs across your enterprise. This guide walks through the four-layer architecture, from metadata foundation to agent orchestration, with practical implementation steps for 2026.
Get the Stack GuideWhat’s the difference between Agentforce and Einstein Copilot (Agentforce Assistant)?
Permalink to “What’s the difference between Agentforce and Einstein Copilot (Agentforce Assistant)?”Einstein Copilot was renamed Agentforce Assistant as part of Salesforce’s broader Agentforce rebrand, and the two aren’t competing products. One folded into the other.
Einstein Copilot launched as a conversational assistant embedded in Salesforce workflows, answering questions and drafting content on request. When Salesforce consolidated its AI portfolio under the Agentforce name, Copilot’s functionality carried over as Agentforce Assistant, a conversational layer that still answers and drafts, while Agentforce itself covers the fuller range of autonomous agents built in Agent Builder and Agentforce Studio that plan and act without a human prompting every step.
The rename mirrors a broader industry pattern: Google folded Agentspace into Gemini Enterprise the same way, keeping old functionality intact under a new umbrella name rather than retiring it. For a Salesforce admin deciding what to build, the practical question isn’t which name is correct. It’s whether the task needs a quick conversational answer, Agentforce Assistant, or a multi-step autonomous process: an Agentforce agent built in Agent Builder and defined by an enterprise copilot’s scope versus a fully autonomous one.
What does Salesforce Agentforce cost?
Permalink to “What does Salesforce Agentforce cost?”Agentforce prices either by the action, through Flex Credits, or by the conversation, with a free Foundations allotment for Enterprise Edition-and-up customers.
Flex Credits are the default consumption unit: a standard agent action costs 20 credits, roughly $0.10, and Salesforce sells credits in blocks of $500 per 100,000. Some customers prefer a simpler unit and choose Conversation pricing instead, a flat $2 per conversation regardless of how many actions the agent takes inside it.
Salesforce Foundations, available on Enterprise Edition and above, ships a free starter allotment of Flex Credits and Data Cloud credits, enough for many teams to pilot a first agent without a separate purchase. Salesforce revises that allotment as the tier evolves, so check the current Foundations allotment before budgeting rather than relying on a fixed number. Per-user licensing for full Agentforce access starts around $125 per user per month. Because Flex Credits meter every individual action, teams scaling past a pilot should model consumption the same way they’d model the cost to run AI agents at scale on any platform, since action volume, not seat count, drives the bill.
| Pricing model | Unit | Cost | Best for |
|---|---|---|---|
| Flex Credits | Per action (20 credits) | ~$0.10/action; $500 per 100,000 credits | High-volume, variable-length interactions |
| Conversation pricing | Per conversation | $2/conversation | Predictable, bounded interactions |
| Salesforce Foundations | Free allotment | Starter Flex Credits + Data Cloud credits; allotment shifts, check current tier | Piloting a first agent |
| Per-user licensing | Per user/month | From $125/user/month | Teams standardizing on full Agentforce access |
Does Agentforce need context beyond Data Cloud and the CRM?
Permalink to “Does Agentforce need context beyond Data Cloud and the CRM?”Salesforce has already extended Agentforce’s context reach with Intelligent Context and a Data 360 MCP Server, but both still operate inside the Data Cloud boundary. Everything outside that estate still needs a separate context source.
Intelligent Context, Agentforce 360’s flagship new capability, grounds agents in unstructured data, PDFs, tables, and images, through Data 360. It’s a genuine admission from Salesforce that structured CRM records alone were never enough context for reliable agent behavior, and it closes a real gap: an agent answering a support question from a scanned contract or a product spec sheet needs exactly this kind of grounding.
Salesforce has also made its ecosystem MCP-native. In May 2026, Salesforce shipped a Data 360 MCP Server in developer preview, connecting Data 360 APIs to any MCP client, including Claude Code, Cursor, Codex, and Agentforce’s own Vibes tooling, with the code developed openly on GitHub. That confirms Agentforce agents aren’t limited to Salesforce-built tools; they can call out to context served over the Model Context Protocol from any compatible source.
Data Cloud’s own Zero-Copy access into Snowflake and BigQuery is real and meaningful reach, not a marketing claim. But warehouse metric definitions, cross-system lineage, and the other SaaS tools in a typical enterprise stack, Tableau, MuleSoft, and dozens of others, still sit outside what Data Cloud natively ingests. An agent that can query a warehouse table still needs to know what “revenue” means in that table, who certified the number, and whether the definition changed last quarter, the same problem addressed by giving AI agents access to enterprise data more broadly and by making AI agents context-aware in the first place. It’s also the exact case why AI agents need an enterprise context layer that reaches past any single vendor’s estate.
This is the kind of gap context-layer and catalog vendors, Atlan among them alongside Alation, Collibra, Informatica, and dbt Labs, exist to close. Atlan’s specific approach picks up the thread Salesforce’s own roadmap already started: Atlan’s MCP server is positioned as an external context source Agentforce agents can call for exactly the data Data Cloud doesn’t hold, the Enterprise Data Graph carries lineage, ownership, and certification across every connector in the stack, and the Active Ontology, an organized semantic layer of glossary terms and metric definitions, resolves what a business term means before an agent acts on the number. Atlan and Salesforce also both back the Open Semantic Interchange standard, a real, verifiable point of technical alignment rather than a marketing claim: the same context graph logic Data Cloud applies inside its own boundary extends outward through a shared open standard, distinct from a general-purpose knowledge graph.
| Aspect | Data Cloud only (native Agentforce) | + Governed context layer (e.g. Atlan) |
|---|---|---|
| Context scope | CRM, Data Cloud, connected apps, and Zero-Copy warehouse access | Adds metric definitions, cross-system lineage, and other SaaS outside Data Cloud’s native connectors |
| Unstructured data | Intelligent Context (PDFs, tables, images via Data 360) | Same, plus context from systems outside Data Cloud |
| Metric definitions | Not natively resolved | Certified via a semantic/ontology layer before an agent acts |
| Lineage | Within Data Cloud’s own pipelines | Cross-system, including non-Salesforce sources |
| Protocol reach | Data 360 MCP Server, MCP-native | Any MCP client, extended by an external MCP server |
Data Cloud already reaches Snowflake and BigQuery natively. What it doesn't resolve on its own, business-term definitions, cross-system lineage, and other SaaS, is where an Enterprise Data Graph and MCP server pick up.
Why do so many Agentforce implementations stall or underdeliver?
Permalink to “Why do so many Agentforce implementations stall or underdeliver?”Independent practitioner sources agree on one thing Salesforce’s own materials don’t emphasize: Agentforce rollouts stall on CRM data quality and governance readiness, not on the reasoning engine.
Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Anushree Verma, Senior Director Analyst at Gartner, put it directly: “Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied.” That statistic isn’t specific to Agentforce, but it describes the exact failure mode practitioners report when an agent is deployed against real CRM data rather than a clean demo environment.
A 2026 Forrester and Anaconda survey, compiled by DigitalApplied’s analysis of enterprise AI agent adoption, found that roughly 88% of agentic AI pilots never reach production, with the leading causes traced to unclear success criteria, insufficient tool or data access, and evaluation drift, not model quality, a secondary compilation of the primary survey data cited here as such.
The Salesforce-specific version of this story shows up directly in practitioner commentary. Connie Hazendonk, a Salesforce Trainer, told Salesforce Ben plainly: “Most are not ready for Agentforce. Not even close.” The recurring, specific complaint isn’t the Atlas Reasoning Engine. It’s duplicate leads, stale record ownership, and undocumented custom fields, the unglamorous CRM hygiene work that has to happen before an agent can act on a record with any confidence. Not every practitioner report is negative: the same Salesforce Ben reporting cites admins who built a working Agentforce agent in a single day, a useful counterpoint showing Agentforce performs well once the underlying data is actually ready. The gating factor practitioners report, over and over, is context and data readiness, exactly the problem a context layer is built to solve, and the same accuracy gap covered in AI agent accuracy: the fix is context and governance, not a better model. Left unaddressed, that gap is also what keeps agent pilots from ever scaling past a proof of concept, the same readiness work covered in preparing enterprise data for AI agents.
Find your context gap
Most Agentforce rollouts stall on data readiness, not the reasoning engine. See where your CRM and Data Cloud context has gaps before you scale past a pilot.
Calculate Your GapHow to choose the right context strategy for Agentforce
Permalink to “How to choose the right context strategy for Agentforce”Enterprises rolling out Agentforce should evaluate their context strategy against five criteria before scaling past a pilot.
Data quality and deduplication come first: an agent that acts on a duplicate lead or a stale ownership record fails at the source, before the reasoning engine even gets involved. Cross-system lineage matters next, since an agent that surfaces a number needs to trace it back to where it came from and whether it’s still trusted. Business-term definitions, resolved through a semantic layer, settle what “revenue,” “active account,” or “churn” mean before an agent uses them in a decision, the same semantic layer for AI agents logic that applies across any agent platform, not just Agentforce. Governance and access policy determine who and what an agent is allowed to touch, and protocol compatibility, whether the context source speaks MCP or another open standard, determines how easily new tools plug into the stack later.
Practical questions worth asking before scaling: Does your context source reach beyond Data Cloud into warehouses and BI tools? Can you trace a number an agent surfaces back to its lineage? Is your CRM data deduplicated and certified before an agent acts on it? Does your semantic layer resolve conflicting definitions of the same metric across teams? Would adding a new agent platform next year require rebuilding this context from scratch? Enterprises that can answer all five with confidence are ready to scale; the ones that can’t have found the real blocker, and it’s rarely the model, the same conclusion reached in how to implement an enterprise context layer for AI.
| Criterion | Why it matters | What to look for |
|---|---|---|
| Data quality & deduplication | Agents act on records at face value; duplicates and stale fields propagate errors | Automated dedup and certification before an agent can query the record |
| Cross-system lineage | An agent’s answer needs a traceable source, not just a plausible one | Lineage that spans CRM, Data Cloud, and non-Salesforce systems |
| Business-term definitions | Different teams calculate the same metric differently | A certified semantic layer resolving one definition per term |
| Governance & access policy | Agents need the same access controls as a human user, checked at query time | Policy enforcement at the point the agent acts, not just at ingestion |
| MCP/protocol compatibility | New tools and agent platforms should plug in without a rebuild | An open, MCP-native context source, not a closed connector |
How Atlan extends Agentforce’s context beyond Data Cloud
Permalink to “How Atlan extends Agentforce’s context beyond Data Cloud”The challenge. Agentforce agents are only as good as the CRM and Data Cloud context they can see. Practitioners cited throughout this page report the same recurring blockers: duplicate leads, stale record ownership, and undocumented custom fields, the exact context gap that stalls implementations covered above.
Atlan’s approach. Atlan’s MCP server acts as an external context source Agentforce agents can call for data Data Cloud doesn’t hold, extending the same pattern Salesforce already uses for connecting enterprise data sources to LLMs securely. The Enterprise Data Graph adds lineage, ownership, and certification across every connector in the stack, not just the Salesforce-native ones, and the Active Ontology resolves what a business term means before an agent acts on a number, work that matters most in exactly the domain Salesforce owns: AI agents for sales run on numbers that need one certified definition, not fourteen. Atlan and Salesforce both back the Open Semantic Interchange standard, a real point of technical alignment rather than a marketing claim, and the same context discipline extends to how a team builds an AI agent harness around any model, not just Agentforce’s own.
The outcome. Enterprises bridging Data Cloud with warehouses like Databricks and Snowflake report the same pattern directionally: agents that could already reason well start acting reliably once the context feeding them is governed and certified, not because the model changed, but because what it was allowed to see did.
See context agents live
Watch how a governed context layer connects CRM, Data Cloud, and everything outside it to AI agents in real time, across systems and vendors.
Watch a Live DemoWhat Agentforce’s own roadmap shows about enterprise AI context
Permalink to “What Agentforce’s own roadmap shows about enterprise AI context”Agentforce’s own product roadmap keeps landing on the same insight. Intelligent Context and the Data 360 MCP Server are both, in different ways, Salesforce building toward the same conclusion this page opened with: raw CRM records alone were never enough context for an agent to act on reliably. The Atlas Reasoning Engine is a genuinely capable planning layer on top of that grounding, and it performs only as well as the context it’s given, which is exactly why Salesforce kept extending that grounding release after release rather than treating the reasoning engine alone as the finish line.
Atlan extends that same context engineering discipline across everything Data Cloud doesn’t natively resolve: business-term definitions, cross-system lineage, and the other SaaS systems every enterprise Salesforce customer also runs. The question worth asking isn’t whether Agentforce’s reasoning engine is capable enough. It’s whether the context feeding it reaches as far as the business actually operates.
FAQs about Salesforce Agentforce
Permalink to “FAQs about Salesforce Agentforce”1. What is an AI agent, and how does that relate to Agentforce?
Permalink to “1. What is an AI agent, and how does that relate to Agentforce?”An AI agent is software that can perceive a goal, plan steps, and take action across systems with limited human intervention, rather than only answering a question when asked. Agentforce is Salesforce’s own AI agent platform, built specifically to run this kind of autonomous, multi-step work inside CRM and business workflows.
2. How does Agentforce differ from other Salesforce AI solutions like Data Cloud or Tableau AI?
Permalink to “2. How does Agentforce differ from other Salesforce AI solutions like Data Cloud or Tableau AI?”Data Cloud is Salesforce’s data unification and grounding layer, and Tableau AI adds AI-assisted analytics inside Tableau’s visualization product. Agentforce sits on top of both, using Data Cloud for grounding and taking autonomous action, while Data Cloud and Tableau AI stay focused on unifying and analyzing data rather than acting on it.
3. What’s the difference between Agentforce 2, Agentforce 3, and Agentforce 360?
Permalink to “3. What’s the difference between Agentforce 2, Agentforce 3, and Agentforce 360?”Agentforce 2.0, released in December 2024, added the Atlas Reasoning Engine. Agentforce 3.0, released in June 2025, added the Command Center for observability and MCP support. Agentforce 360, which reached general availability in October 2025, packaged all of it into a broader “Agentic Enterprise” platform and added Intelligent Context.
4. Is Agentforce included in my Salesforce license, or is it a separate purchase?
Permalink to “4. Is Agentforce included in my Salesforce license, or is it a separate purchase?”Agentforce is a separate, consumption-based purchase priced through Flex Credits or Conversation pricing, though Salesforce Foundations includes a free allotment for Enterprise Edition and above. Most organizations need to budget for it separately from their existing Salesforce license.
5. Should I use Flex Credits or Conversation pricing for my business?
Permalink to “5. Should I use Flex Credits or Conversation pricing for my business?”Flex Credits work better for variable-length or high-volume interactions, since you pay per action rather than per conversation. Conversation pricing is easier to forecast when interactions are predictable and bounded, because the cost per conversation stays flat regardless of how many actions an agent takes inside it.
6. What data does Agentforce need to work well?
Permalink to “6. What data does Agentforce need to work well?”Agentforce needs clean, deduplicated CRM records with clear ownership, plus the Data Cloud grounding data an agent draws context from at runtime. Practitioners consistently report that duplicate leads, stale ownership fields, and undocumented custom fields, not the reasoning engine, are what causes agents to underperform.
7. Why do so many Agentforce implementations stall or underdeliver?
Permalink to “7. Why do so many Agentforce implementations stall or underdeliver?”Independent practitioner sources point to CRM data quality and governance readiness, not the Atlas Reasoning Engine, as the actual blocker. Gartner projects that over 40% of agentic AI projects will be canceled by 2027 for related reasons: unclear business value and inadequate risk controls, not weak models.
8. Does Agentforce work with data outside Salesforce, like data warehouses or BI tools?
Permalink to “8. Does Agentforce work with data outside Salesforce, like data warehouses or BI tools?”Agentforce reaches some of that data natively through Data Cloud’s Zero-Copy access into Snowflake and BigQuery, and through Intelligent Context for unstructured files. Anything outside Data Cloud’s native connectors, including most BI tools and other SaaS systems, still needs an external context source.
9. How does Atlan connect to Salesforce Agentforce?
Permalink to “9. How does Atlan connect to Salesforce Agentforce?”Atlan connects through its own MCP server, which Agentforce agents can call as an external context source for data Data Cloud doesn’t hold. Atlan and Salesforce also both back the Open Semantic Interchange standard, giving the two platforms a shared, open way to align on what business terms mean.
Sources
Permalink to “Sources”- Inside Agentforce: Revealing the Atlas Reasoning Engine, Salesforce Engineering (Phil Mui, 2024)
- Agentforce Metrics, Salesforce (2026)
- Welcome to the Agentic Enterprise: Agentforce 360 General Availability, Salesforce (2025)
- New Flexible Agentforce Pricing, Salesforce (2025)
- Salesforce Agentforce Pricing, Salesforce (2026)
- Intelligent Context in Data 360, Salesforce Help (2026)
- Introducing the Data 360 MCP Server (Developer Preview), Salesforce Developers Blog (2026)
- forcedotcom/d360-mcp-server, GitHub (2026)
- Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, Gartner (2025)
- Has Agentforce Moved from Hype to Reality?, Salesforce Ben (Sasha Semjonova, 2025)
- Salesforce’s Agentforce Bet Made It the Dow’s Worst Performer, Yahoo Finance (2026)
- AI Agent Adoption 2026: 120+ Enterprise Data Points (secondary compilation of Forrester and Anaconda survey data), DigitalApplied (2026)
- Salesforce Foundations (official product page; verify current free-tier credit allotments here before citing a specific number)
