Agentforce

Salesforce announced Agentforce in September 2024 and made it generally available in October 2024. Since Dreamforce in October 2025, the entire stack has been marketed under the umbrella Agentforce 360: the Agentforce 360 platform, the data layer Data 360 (formerly Data Cloud), the Customer 360 applications and Slack. When you read "Agentforce" today, it can therefore mean three things depending on the context: the product for customer and employee agents, the construction-kit platform underneath, or the brand name for Salesforce's entire AI strategy. For clarifying the term, the middle meaning is the relevant one: Agentforce is an AI agent framework that is firmly coupled to Salesforce data and Salesforce automation.

Chatbot, copilot, autonomous agent: where Agentforce stands

The term "agent" is used inflationarily in marketing. For a purchasing decision, however, the distinction is the most important of all, because it determines which processes you can automate at all and what risk you take in doing so.

How to recognise a real agent

An autonomous agent has three things that a chatbot does not: a goal instead of a script, a toolbox of executable actions and a loop in which it checks after each step whether it has come closer to the goal. Salesforce calls this loop the Atlas Reasoning Engine and publicly describes it as a ReAct pattern (reason, act, observe), in contrast to the rigid chain-of-thought plan of the previous generation. This is not specific to Salesforce but the state of the art in agent frameworks; what is interesting is the concrete implementation.

In practice this means: you describe the agent's job in natural language ("answer questions about delivery status, offer an alternative in case of delay, escalate complaints"), store the actions it may call for this and define the limits. A large language model handles language understanding and planning, while execution is handled by deterministic building blocks: flows, Apex code, API calls. This separation is the reason why agentic systems become production-ready at all: the language model chooses, but it does not calculate the discount itself.

Why the distinction matters when buying

Anyone who buys a chatbot and expects an agent pays for disappointment. Anyone who buys an agent and treats it like a chatbot causes damage. The question you should put to every provider is therefore very specific: which systems may the thing touch with write access, and what happens when it gets something wrong? Everything else, such as answer quality, tone and language coverage, is secondary in comparison. Salesforce answers this question via the rights of the executing user context, via instructions in the subagent and via a native "hand over to a person" action.

How Agentforce is built technically

Agentforce is not a model but an orchestration layer. It consists of four building blocks you should know when you read offers or project plans.

Agent Builder, subagents and actions

The Agent Builder is the low-code interface in which an agent is created. The central structural element is subagents: each subagent maps an intention ("order status", "return", "quote request") and carries its own instructions, business rules and actions. An incoming request is first assigned to a subagent, which keeps the solution space for the language model small. Salesforce describes this mechanism as the reason why an agent can scale to thousands of actions without latency and accuracy collapsing; you can read about it in the technical description of the Atlas Reasoning Engine. In addition, since 2025 there has been Agent Script for cases in which the flow has to be fixed deterministically instead of being planned by the model.

Data 360 as the data foundation

An agent is only as good as what it is allowed to see. Data 360 (Data Cloud until October 2025) is the layer that brings CRM data together with external sources and prepares it for retrieval methods. Through it, the agent draws context, such as customer history, contract conditions and knowledge articles, instead of relying on the language model's general knowledge. Important for the calculation: Data 360 consumption is measured separately and is not included in the Agentforce prices listed below. This is exactly where most business cases in mid-sized companies get out of hand, because data integration appears as a side issue in the offer and becomes the main issue in the project.

Guardrails, trust layer and handover to people

Salesforce bundles its protective mechanisms under the Einstein Trust Layer: toxicity checks, contracts without data retention at the model provider, defence against prompt injection. In addition there are the guardrails you set yourself, as a natural-language instruction in the subagent and, more effectively, as a rights and role model on the objects. For critical processes, Agentforce treats the handover to a person as a regular action; human in the loop is therefore not a stopgap but a plannable part of the process. Via MCP support, external tools can also be connected via the Model Context Protocol, which loosens the tie to Salesforce's own actions somewhat.

What Agentforce costs

Salesforce lists its prices publicly, including euro values. As of August 2026, there are three consumption and licence models side by side; they can be looked up on the official Agentforce pricing page.

A detail that can overturn calculations: Flex Credits and Conversations cannot be mixed in the same organisation, and unused Flex Credits expire at the end of the contract term. Overconsumption does not incur penalty fees but is billed monthly at the contractually agreed rate.

A worked example from practice

Salesforce itself works through the classic "Where is my order?": the agent authenticates the customer via her email address and retrieves the order details, two actions, i.e. 40 Flex Credits. With 20 requests a day over 30 days, that is 24,000 credits a month, roughly €120 a month. For a service team with 100 Salesforce users who each handle three cases a day on 20 days a month, on the other hand, Salesforce quotes 360,000 credits and around €1,800 a month. The figures come from vendor material and are explicitly labelled as sample calculations, but they show the mechanics: the price scales with the number of actions, not with the number of conversations. An agent that makes many small system calls is more expensive in the Flex model than one that answers once.

Agentforce in commerce

For retailers, Agentforce has become interesting above all since the Commerce release of 24 June 2026. In it, Salesforce made three ready-to-use agents generally available and opened the platform further towards agentic commerce.

Shopper Agent, Buyer Agent, Merchant Agent

The Shopper Agent takes over the conversation in your own storefront: checking availability, confirming the shipping cut-off, offering collection from a nearby store, completing the purchase. The Buyer Agent targets B2B procurement and runs in WhatsApp and SMS; the example from the announcement is a reorder by text message, including image confirmation of the SKU, contract price and a finished order, without logging into a portal. The Merchant Agent is aimed at the back office: sorting the catalogue and maintaining boost and bury rules in natural language. For early Merchant Agent customers, Salesforce cites an 88 per cent reduction in processing time, a vendor claim without a publicly available methodology that you should treat with corresponding caution. All the details are in the official release announcement.

Connection to ChatGPT, Google and Gemini

The strategically more important part is the opening to external AI interfaces. Via Agentic Commerce Channels, the product catalogue can be connected from Business Manager directly to ChatGPT; integration into Google Search, including AI Mode, and into the Gemini app was announced for summer 2026. Salesforce stresses that the retailer remains the merchant of record and that the order lands on the same platform as service and loyalty. Salesforce has also announced support for Google's Universal Commerce Protocol and for the Agentic Commerce Protocol from OpenAI and Stripe. Anyone thinking today about how their products appear in AI answers should know this protocol level; it is relevant regardless of platform, including for Shopware or headless set-ups.

What operation in a mid-sized company actually requires

Technology is rarely the problem. Four prerequisites decide whether an agent project works:

• Data quality. An agent that reads out a wrong delivery time passes it on with complete conviction. Duplicates, outdated price lists and inconsistent item master data are not concealed by AI but made more visible.
• Rights and role model. The agent acts in the context of a technical user. If that user has too many permissions, the agent can do more than you want. The permission concept is the hardest guardrail you have, harder than any instruction in natural language.
• Defined escalation. Specify in advance for which processes a person decides: credit notes above an amount limit, complaints, contract changes. This limit belongs in the configuration, not in an operating instruction.
• Measurability. Without an analysis of what share of cases was actually resolved completely, no business case can be calculated. Agentforce brings observability tools for this; use them from day one, not only in the second year.

Regulation: the transparency obligation is mandatory

The EU AI Act applies. Since 2 August 2026, the transparency obligations of Article 50 have applied: anyone operating an AI system that interacts directly with people must inform them that they are dealing with an AI system, unless this is obvious from the context to a reasonably well-informed person. In practice this means a clear notice at the start of the chat and in the voice greeting, as well as a documented escalation option. This is not a big hurdle, but it is one you will not forget once you have planned for it. Further context can be found in our entry on the EU AI Act.

When Agentforce pays off, and when something leaner does

Agentforce plays to its strength when the data is in Salesforce anyway. Then the path from conversation to action is short, the permissions are already modelled and the handover to a human service agent happens in the same system. If, on the other hand, Salesforce is only one of several systems, or not in use at all, you buy a platform lock-in that you do not need functionally.

For this case there are vendor-neutral alternatives: a workflow automation tool such as n8n with a connected language model, your own agent service based on open frameworks, or a specialised solution for the one process that really has volume. The price difference is rarely the decisive point; the decisive point is where the business logic lies and who can change it. As a rule of thumb: with an existing Salesforce landscape and several processes at once, Agentforce is a serious option. For a single, clearly defined use case in a heterogeneous system landscape, the lean in-house solution is usually live faster and cheaper to operate.

nextlevels is not a Salesforce partner. This assessment is therefore neither a recommendation nor a warning, but an attempt to explain the term so that you ask the right questions in your next conversation with a vendor. The prices come from the public Salesforce pricing page as of August 2026 and may change; list prices are never binding, and they are negotiated anyway. And the most important question remains the same regardless of the vendor: which process should the agent take over, how do you measure whether it does so, and what happens when it gets something wrong?