Part 3 of the series "The Enterprise AI Vendor Landscape"
Agentforce 360 and the Gamble That the CRM Platform Becomes the Agent Platform
The Platform Pivot
Salesforce is not adding agents to its CRM. It is rebuilding the CRM around agents.
That is a larger bet than what any other vendor in this series is making within its core product. Microsoft is layering agent governance on top of an existing productivity suite. SAP is embedding AI into established ERP workflows. Salesforce is doing something more disruptive to its own business: repositioning the entire platform so that agents, not human users clicking through screens, are the primary interface for customer operations. The per-seat CRM model that generated $38 billion in FY2026 revenue is being supplemented, and potentially replaced, by a consumption-based model where the unit of value is not a licensed user but an action performed by an AI agent.
Agentforce 360, introduced at Dreamforce in October 2025, is the architectural expression of this bet. It reorganizes Salesforce's product portfolio into four parts: the Agentforce 360 Platform (the reasoning, trust, and agent-building layer), Data 360 (the rebranded Data Cloud for unified customer data), Customer 360 Apps (the familiar Sales, Service, Marketing, and Commerce applications), and Slack (repositioned as the collaboration surface where agents and humans work together). The naming signals what Salesforce wants the market to understand: this is not a feature addition. It is a new product architecture where agents sit at the center and everything else, data, applications, collaboration, orbits around them.
The financial trajectory suggests the market is responding. Agentforce reached $800 million in annual recurring revenue by the end of FY2026, a 169% year-over-year increase. By Q2 FY2027, that figure exceeded $1.5 billion, with 240% year-over-year growth, though Salesforce broadened the definition to include adjacent AI offerings, Slackbot, and Headless 360. The company reports 29,000 Agentforce deals and 7.0 billion total Agentic Work Units delivered, with 3.2 billion in the most recent quarter alone.
The question is not whether the numbers are growing. They are. The question is whether Salesforce can execute a platform transition of this magnitude, turning a per-seat software company into a consumption-based agent platform, without stumbling over the technical integration challenges or the revenue model uncertainty that any per-seat-to-consumption transition creates.
The Reasoning Engine
At the core of Agentforce is the Atlas Reasoning Engine, which distinguishes Salesforce's approach from vendors that simply wrap API calls to 3rd party language models.
Atlas is not a single inference call. It is an iterative loop: reason, act, observe, adapt, act again. When an agent receives a request, Atlas plans the steps, retrieves grounding data from Data Cloud, executes actions through the Salesforce platform, evaluates the results, and adjusts its approach if the initial plan does not produce the expected outcome. The loop continues until the task is complete or the agent determines it needs human guidance. This iterative architecture is what allows Agentforce agents to handle multi-step workflows rather than simple question-and-answer interactions.
The Trust Layer sits beneath the reasoning engine, inspecting every transaction between Atlas and the underlying language models. It masks sensitive data before model calls, enforces user-level permissions on every action, applies toxicity and bias detection, and maintains audit logs. The design reflects Salesforce's recognition that enterprise customers will not deploy autonomous agents against customer data without verifiable guardrails. The Trust Layer is Salesforce's answer to that requirement: every agent action passes through the same permission model that governs human users.
In September 2026, Salesforce announced Koa, its first CRM-specific reasoning model, at Dreamforce. Built on NVIDIA's Nemotron 3 Super architecture, Koa was trained using supervised fine-tuning and reinforcement learning (specifically GRPO, Group Relative Policy Optimization) on synthetic data drawn from roughly 30 years of CRM domain experience. No customer data was used in training. Salesforce controls the model weights and runs inference within its own trust boundary, a critical distinction for regulated industries.
The performance profile is telling. Koa outperforms the base Nemotron model on CRM-specific tasks but trails frontier general-purpose models on broader benchmarks. That is the point. Salesforce is not trying to build a general-purpose AI lab. It is building a domain-specialized reasoning model that is good enough for CRM workflows while keeping the entire inference pipeline within Salesforce's security perimeter. The NVIDIA partnership extends further through Missionforce, which adapts Nemotron-based models for government and regulated organizations, a segment where model provenance and data residency are not preferences but requirements.
Koa is currently in pilot with select customers, with general availability expected in (calendar) winter 2026. Its significance is strategic, not just technical: Salesforce is vertically integrating its AI stack rather than remaining dependent on third-party model providers. Whether Koa delivers enough performance advantage over calling external models through the Trust Layer is an open question that pilot results will need to answer.
Multi-Agent Orchestration
The single-agent model, one agent handling one task, was always a stepping stone. The strategic capability is multi-agent orchestration: multiple specialized agents collaborating on complex workflows, handing off context to each other, and coordinating their actions without requiring human intervention at every transition point.
Salesforce's Multi-Agent Orchestration reached general availability in 2026. The architecture uses a primary agent powered by Atlas to evaluate incoming requests, plan the sequence of steps, and dispatch work to specialist agents. A service inquiry that begins as a billing question but involves a product return, a loyalty adjustment, and a shipping change can be routed across multiple specialist agents, each operating within its own domain expertise and permission boundaries, with the orchestrating agent maintaining context across the entire interaction.
At Dreamforce 2026, Salesforce extended this capability in two important directions. First, Agentforce Coworker: an autonomous agent that can build a plan and execute it by orchestrating other agents within a conversation, positioning agents not as tools that humans invoke but as teammates that operate alongside them. Second, third-party orchestration: the ability to coordinate agents from AWS, Azure, Google, and other platforms, or to let Agentforce act as a specialized subagent within any A2A-compliant architecture.
The third-party orchestration capability connects directly to the interoperability analysis in our Agent Coordination Gap series (Part 3). Salesforce's implementation of the Agent Service Bus functions, routing, state management, error handling, protocol translation, and observability, determines whether Agentforce can participate in heterogeneous multi-agent environments or whether it remains a closed ecosystem. The A2A compliance claim is significant: it signals Salesforce's intent to be a participant in open multi-agent architectures rather than a walled garden. Whether that intent translates into frictionless interoperability in practice is something enterprise architects will need to validate in their own environments.
The Enterprise AI Harness, also announced at Dreamforce 2026, adds monitoring, tracing, security, and governance capabilities across all AI agents operating within Salesforce. Agent Optimizer, an agent that builds and refines other agents, was announced at Dreamforce with an October 2026 GA target release. And the Long-Horizon Runtime, introduced alongside seven job-ready agents (with Hunter as the first to use it), extends agent capabilities to longer, more complex task sequences. These are incremental capabilities, but they signal Salesforce's direction: agents that operate over longer time horizons, across more complex workflows, with less human intervention.
The Pricing Transition
The shift from per-seat licensing to consumption-based pricing is the most consequential business model change in Salesforce's history. It is also the hardest to execute.
Salesforce offers two consumption models for Agentforce. The first is a flat $2 per conversation, where a conversation is defined as a session of up to 24 hours. The second, and newer, model uses Flex Credits: $500 per 100,000 credits, with a standard action consuming 20 credits (approximately $0.10 per action) and a voice action consuming 30 credits (approximately $0.15). For most use cases, Flex Credits are cheaper than the per-conversation rate, with the break-even at roughly 20 actions per conversation.
Salesforce has also introduced edition-based bundles. Core at $195 per user per month, Advanced at $395, and Max at $550, each including different levels of Agentforce capability alongside the traditional CRM applications. A $5 user license option provides basic access while metering actual usage through credits.
The strategic logic is sound. Per-seat pricing makes little sense when the value proposition is that agents reduce the number of human seats needed to accomplish the same work. If Agentforce genuinely automates customer service interactions, billing by the seat penalizes adoption: the more successful the agent deployment, the fewer seats an organization needs, and the less revenue Salesforce earns. Consumption-based pricing aligns Salesforce's revenue with the value its agents deliver rather than with the number of humans who log in.
The execution risk is real. Consumption revenue is inherently less predictable than seat-based recurring revenue. Wall Street values predictability, and the transition from a known recurring revenue base to a variable consumption model creates uncertainty that investors and analysts will scrutinize closely. Salesforce's broadening of the ARR definition to include adjacent AI offerings suggests the company is already working to present the consumption trajectory in the most favorable light, a move that is understandable but that analysts should adjust for.
The comparison with Microsoft's approach is instructive. Microsoft charges $15 per user per month for the Agent 365 governance layer, with agent compute billed separately through Azure consumption. Salesforce is attempting a more radical model: the governance, reasoning, and execution all live within the same consumption meter. Oracle, at the other extreme, includes AI agents at no additional cost within its Fusion Cloud applications, betting that agent value accrues through broader platform adoption rather than direct AI monetization. Each approach carries different risks, and the Part 1 pricing fault line analysis applies directly here: the vendor whose pricing model most closely aligns with how enterprises capture value from agents will have a structural advantage as the market matures.
There is also a customer-side pricing risk. Consumption models can surprise buyers with costs that scale faster than expected. Salesforce's published rate card notes that usage types, tiers, and multipliers may change, and credits do not roll over. Enterprise procurement teams accustomed to fixed per-seat contracts will need new budgeting and forecasting disciplines for a model where monthly costs vary with agent activity. The three-tier decision authority model from the Operating Model Gap series applies here: organizations need clear policies about which agent actions (Tier 1: agent acts freely, Tier 2: agent recommends and human decides, Tier 3: human only) are worth consuming credits on, and which should remain human-performed.
The Platform Integration Question
The most important question about Salesforce’s agentic strategy is not whether the individual pieces are strong. They are. The question is whether the pieces work as one platform.
Salesforce built its current portfolio partly through acquisition: MuleSoft (2018), Tableau (2019), Slack (2021), Spiff (2024). Each brought genuine capability, and each brought integration debt. The four-layer Agentforce 360 architecture looks coherent on a diagram, with Data 360, Customer 360 Apps, the Agentforce Platform, and Slack working as a unified system. Production deployments expose the seams. The Tableau rollout illustrates the pattern. Tableau Next’s full agentic analytics capabilities, conversational insights via Slackbot, proactive alerts, and headless MCP access, are generally available on Slack but still in pilot on Microsoft Teams, with conversational insights and headless flexibility listed as “coming soon” for Teams on Cloud and Server editions. When the owned collaboration layer gets first-class integration and the one used by most enterprise customers lags behind, it is a visible signal of how acquisition-assembled platforms prioritize their engineering resources. Data Cloud’s real-time unification depends on connectors to legacy systems that vary in maturity. Slack’s role as the agent interface layer requires deeper integration with Customer 360 workflows than the current implementation consistently delivers. Tableau’s analytics capabilities need tighter coupling with the Atlas reasoning engine to provide the observability that production agent deployments demand.
There is a case that MuleSoft, not Agentforce itself, is Salesforce’s most strategically significant AI asset. MuleSoft’s integration layer, now branded as Agent Fabric, is what connects Agentforce to the enterprise systems where customer data resides: ERP, supply chain, legacy databases, and third-party applications. Without that connectivity, Agentforce agents are limited to the data and actions available within Salesforce’s own applications. With it, they can orchestrate across the enterprise. If this reading is correct, the strategic center of gravity in Salesforce’s agent architecture is the integration layer, not the reasoning engine.
AgentExchange, Salesforce’s marketplace for pre-built agents and agent components, raises a related concern. The marketplace model creates value for customers when it is genuinely open. It creates lock-in when the platform owner controls which agents can participate, how they access data, and what they can charge. Enterprise buyers should evaluate whether AgentExchange functions as an open ecosystem or a curated storefront, and whether the agents available through it can interoperate with agents built on competing platforms.
The pricing bundles deserve scrutiny as well. The three edition tiers (Core at $195, Advanced at $395, Max at $550 per user per month) layer Flex Credits on top of per-seat fees, but the bundled credit allocations may not align with an organization’s current agent maturity. An enterprise in early pilot stages may find itself paying for capacity it will not use for quarters, while an organization with high agent volume may exhaust bundled credits and face overage rates that change the unit economics.
A November 2026 briefing on Tableau’s collaboration integrations offered evidence on both sides of the integration question. Tableau now exposes its analytics through MCP servers that connect not only to Slack but to Microsoft Teams and Copilot, allowing customers to surface governed Tableau insights inside competing agent platforms. Salesforce’s internal sales organization built a Tableau Next and Slack integration in roughly 40 days that thousands of sales leaders now use for daily intelligence briefings. The cross-platform openness is a genuine signal that Salesforce is building integration rather than just announcing it. The gaps, your questions will be answered in a different layer, the knowledge and context layer briefing expected in December, which suggests the full picture is still incomplete.
These are not reasons to dismiss the strategy. They are execution risks that are specific enough to be testable. Enterprise evaluators should ask: Does the Data Cloud integration with our legacy systems work in our environment? Does multi-agent orchestration perform reliably across Salesforce and non-Salesforce agents in our architecture? Are the pricing bundles aligned with our current deployment scale, or are we being sized for a future state we may not reach? The fair read is that Salesforce has strong products and a bold vision, and the gap between vision and integrated execution is the risk to manage.
The Dreamforce 2026 Signal
Dreamforce 2026, held September 15 to 17 in San Francisco, made the "agentic enterprise" the organizing frame for Salesforce's entire product strategy. (See our research report for Dreamforce)
The messaging was deliberate. Salesforce positioned the agentic enterprise not as a future vision but as a present-tense capability: organizations where people and AI agents work together across company data and processes, not only inside the traditional Salesforce interface. Multi-agent orchestration, the Koa reasoning model, seven job-ready agents, Agentforce Coworker, the Enterprise AI Harness, and Agent Optimizer were all announced or updated during the event.
The multi-model strategy is worth noting. Salesforce connected its architecture to Anthropic’s Claude models through the ClaudeForce partnership announced in August 2026, making Claude the default reasoning model for Slack AI, Agentforce Coworker, and the Atlas Reasoning Engine. It also integrated Google’s Gemini models for broader analytical tasks. Salesforce is not locking itself to a single model provider, even as it develops its own Koa model. The flexibility to route different agent tasks to different models, whether Koa for CRM-specific reasoning, Claude for general-purpose agent work, or Gemini for analytical tasks, gives Salesforce architectural options that a single-model strategy would foreclose. The two-way nature of the Anthropic deal, where Salesforce adopted Claude as a core reasoning layer while Anthropic adopted Salesforce as its CRM and runs its operations through Slack, suggests this is a deeper partnership than a simple model-licensing arrangement.
The introduction of Hunter, the first agent to use the new Long-Horizon Runtime, illustrates where Salesforce is heading. Hunter is not a chatbot that answers questions. It is a sales development agent that researches prospects, builds personalized outreach plans, and executes multi-step engagement sequences autonomously. The Long-Horizon Runtime gives agents like Hunter the ability to maintain context and execute plans over extended periods, hours or days rather than a single conversation session. This is the capability gap that separates demonstration agents from production agents: the ability to persist state, adapt to changing conditions, and complete complex tasks without losing track of the original objective. Hunter is in pilot with general availability expected in November 2026.
The competitive framing was implicit but clear. Where Microsoft positions itself as the governance layer above all agents, Salesforce positions itself as the platform where agents are built, trained, deployed, and orchestrated around the richest source of customer data in the enterprise. The bet is that customer data gravity, the pull of having the most comprehensive customer record, makes Salesforce the natural home for customer-facing agents. Governance is important, but the agents that drive revenue, resolve service issues, and manage customer relationships need to live where the customer data lives.
What Enterprises Should Watch
Salesforce is making the most aggressive agent-first bet in enterprise software. The architecture is coherent. The revenue trajectory is strong. The product is shipping.
The risks are equally clear. The platform integration question is not theoretical: it determines whether Agentforce 360 is a unified agent platform or a marketing umbrella over products that still require significant integration work to function as one. The pricing transition from seats to consumption creates revenue uncertainty for Salesforce and budgeting complexity for customers. And the Koa reasoning model, while strategically important as a move toward vertical integration, needs to demonstrate that domain-specialized reasoning justifies the trade-off against frontier model capabilities.
For enterprise leaders evaluating Salesforce's agentic strategy, the decision framework from the AI Strategy series applies: assess the platform against your specific requirements for customer data integration, multi-agent orchestration maturity, interoperability with non-Salesforce systems, and pricing model alignment with your deployment trajectory and cost management plan. The vision is compelling. The execution is what you are buying.
Strategy Playbook
1. Test Multi-Agent Orchestration With Your Actual Systems. Salesforce's multi-agent orchestration claims are architecturally sound, but production performance depends on integration quality with your specific systems. Before committing, run a proof of concept that orchestrates Agentforce agents alongside agents from at least one non-Salesforce platform (AWS, Azure, or Google) using A2A protocols. Evaluate handoff reliability, context preservation, latency, and error handling across the boundary. The Agent Coordination Gap series (Part 3) provides the evaluation framework: test routing, state management, error handling, protocol translation, and observability against your requirements.
2. Model Your Consumption Economics Before You Sign. The shift from per-seat to consumption pricing changes the budgeting discipline entirely. Map your current customer interaction volumes to the Flex Credits model ($0.10 per standard action, $0.15 per voice action) and the per-conversation model ($2 per session). Model scenarios at 50 percent, 100 percent, and 200 percent of projected volumes. Factor in that credits do not roll over and that Salesforce reserves the right to adjust multipliers. Build internal guardrails using the three-tier decision authority model to control which interactions consume credits autonomously (Tier 1) and which require human approval (Tier 2).
3. Evaluate the Data Cloud Integration Depth. Agentforce's value proposition depends on Data Cloud providing agents with a unified, real-time customer record that spans Salesforce and non-Salesforce systems. Test the Zero-Copy access to your data warehouse (Snowflake, BigQuery, or equivalent) under production-representative loads. Evaluate whether the RAG-based retrieval delivers accurate, current results when your data changes frequently. The gap between Data Cloud as an architectural concept and Data Cloud as a production data fabric in your environment is the gap between the Agentforce vision and the Agentforce reality.
4. Watch Koa, But Do Not Wait For It. The Koa reasoning model is strategically significant as a signal of Salesforce's vertical integration ambitions, but it is currently in pilot and trails frontier models on general benchmarks. Build your Agentforce deployment on the Atlas Reasoning Engine with current model integrations. Treat Koa as an optimization opportunity for CRM-specific workflows once it reaches general availability and independent benchmarks confirm its performance advantage over external models for your use cases. Do not delay your agent strategy waiting for a model that may or may not deliver a meaningful uplift over what is available today.
Arion Research advises enterprise leaders on AI strategy and the shift to a digital workforce.
