Salesforce Agentforce: Complete Guide to Features, Pricing & Limitations (2026)

If someone handed you "go look at Agentforce" this quarter, I'd bet you've already hit the problem: nearly every page that ranks for it is written by Salesforce or by a Salesforce implementation partner. There's a popular Reddit thread titled "Can someone succinctly explain what AgentForce is without all the BS", and it pulls more search traffic than most of Salesforce's own pages. That tells you how many people bounce off the marketing and go looking for a plain answer.

Here's where most readers are sitting:

1. You're already on Salesforce Service Cloud and a director told you to "add AI." Agentforce is your default path, so read the setup and cost sections before you commit (this is the most common situation we see).
2. You're on Zendesk, Intercom, Freshdesk or HubSpot and wondering whether to move to Salesforce to get Agentforce. The integration answer changes the whole decision.
3. You're trying to model the cost, and the $2-per-conversation-versus-Flex-Credits-versus-per-user pricing is a mess. The pricing section is for you.

I'll cover what Agentforce is, how it works, where it fits, what it really costs, what resolution rate to expect, and how to roll it out without regrets. Plus the one question that decides whether it's even an option for you.

What is Salesforce Agentforce?

Agentforce is Salesforce's autonomous AI agent layer, built on top of the Salesforce Platform. Salesforce defines it as "a proactive, autonomous AI application that answers questions, takes actions, and improves productivity", and positions it firmly against chatbots and copilots. Marc Benioff calls it the "Third Wave of AI."

In support terms, it's two products that are easy to confuse (we get asked the difference on a lot of calls). Service Agent is the autonomous, customer-facing agent that answers and resolves tickets on its own.

Service Assistant (formerly Einstein Copilot) is the agent-facing copilot that helps your human reps inside the Service Console. Both run on the same building blocks, the Atlas Reasoning Engine, Topics, Actions, Instructions, and the Einstein Trust Layer, but they're sold separately.

It launched fast and rebranded even faster. Agentforce hit general availability on October 29, 2024, then shipped 2.0, 2dx, Agentforce 3, and Agentforce 360 in barely a year.

At Dreamforce 2025, Salesforce renamed much of its own stack around the brand: Service Cloud became "Agentforce Service," Data Cloud became "Data 360," and the Einstein 1 Platform became the "Agentforce 360 Platform" (fun fact: Benioff has even floated renaming Salesforce itself to Agentforce, after a text from Elon Musk).

The scale numbers are genuinely large. Salesforce reported 29,000 Agentforce deals at $800M ARR in Q4 FY26.

The caveat is just as large: third-party analysis from Stifel suggests only about 5.3% of Salesforce customers have adopted it, and SaaStr notes fewer than 10% of the 150,000+ install base has even signed an Agentforce deal. It's the loudest product in the category and one of the least-scaled relative to its own base.

G2: Agentforce scores 4.3/5 from 1,109 reviews on G2. "Not just a chatbot, it actually helps with performing actions like creating records, generating emails… very secure as it uses Einstein Trust Layer and provides a zero-retention policy." via a G2 reviewer (Software Developer, Mid-Market).

How do you set up Salesforce Agentforce?

Salesforce markets three-to-six-week deployments. In practice, real production rollouts run five to eleven months, and they almost always need a systems integrator.

The gap between the two numbers is well documented. Salesforce's own Forward Deployed Engineering team claims it reduced deployments "from six months to three weeks" and cites reMarkable (3 weeks) and OpenTable (weeks) as proof.

A Salesforce SI, Clientell, gives a more grounded range: single-use-case deployments go live in 4 to 6 weeks "when data quality is reasonable," and complex multi-agent deployments take 8 to 16 weeks. A third-party analysis pegs the full realistic timeline at 22 to 44 weeks, or 5.5 to 11 months, for production at enterprise scale.

Word-on-the-street among Salesforce admins is that the agent itself is the easy part and getting the org ready is the slog. As Salesforce Ben's ecosystem summary puts it, "creating an agent and some actions [is] not hard… but all the blocking and tackling of getting an org ready for Agentforce… I have found fiddly and not well thought through." Implementation estimates from GetGenerative land at $50,000 to $150,000 upfront plus $10,000 to $25,000 a month in ongoing consulting. And Agentforce requires Service Cloud Enterprise Edition or higher before you can even build the agent.

What channels does Agentforce work in?

Agentforce covers a broad set of messaging channels (voice is the newest and most gated addition).

Generally available channels include web and in-app messaging (MIAW), SMS, WhatsApp (Enhanced), Facebook Messenger, Apple Messages for Business, and LINE, plus email-to-case, which was added after the initial launch. Agentforce Voice went GA on October 21, 2025 with native partners Amazon Connect, Five9, Genesys, NiCE, and Vonage, though voice is US and Canada only as of early 2026. Agentforce Contact Center reached GA in February 2026.

Two gaps stand out for social-heavy teams: X/Twitter and Instagram DMs aren't listed as native channels in current Digital Engagement materials, and the legacy Chat and Standard Messaging channels are end-of-life.

What are the limitations of Salesforce Agentforce?

Salesforce Agentforce's biggest limitation decides the whole thing, so I'll put it first among the trade-offs: Agentforce only works where Salesforce is already the system of record.

There are no native integrations with Zendesk, Intercom, Freshdesk, or HubSpot Service Hub. Salesforce runs active comparison pages pitching Agentforce as a Zendesk replacement, not as something that plugs into it. The only Zendesk-to-Salesforce link that exists is a CRM-data sync for accounts and contacts, which does nothing to let Agentforce resolve tickets inside those helpdesks (we've had this exact conversation with teams switching over).

To put Agentforce into a non-Salesforce helpdesk, you'd have to expose it through MuleSoft APIs, a separately licensed product that adds real cost. If your team lives in one of those four helpdesks, Agentforce is effectively out of reach unless you migrate. We hear from those teams a lot.

The pricing model has its own trap: you pay per conversation or per action whether or not the issue is resolved. As Salesforce Ben summarized the customer complaint, "if every query gets escalated, now you're paying twice: once for Agentforce and again for a human… you've paid $2 just to be told someone else needs to help." Knowledge ingestion is credit-hungry too, since every chunk indexed into Data Cloud consumes credits.

There's also a public trust wobble I'd want to know about before buying. In September and October 2025, Salesforce removed the search bar from its own Help site and forced customers to use Agentforce instead. The community revolted, calling it "slow, inaccurate, and producing repetitive answers or even incorrect information", and Salesforce brought search back.

That last point is the one I'd underline hardest. Even big, well-funded vendors are surprisingly primitive on the fundamentals: we've watched a large vendor's agent answer a support question with advice on caring for someone's dog, and another well-known tool that still can't respond past the first message in an email thread.

Good AI support is a skill a company has to build, and the spread across vendors is huge, so run your own hardest tickets through any agent (Agentforce included) before you trust a resolution number.

What knowledge sources can you train Agentforce on?

Agentforce ingests a reasonable spread of sources (all routed through Data Cloud, now Data 360).

Native sources include Salesforce Knowledge articles, uploaded files (PDF, HTML, and text), web-search grounding with inline citations, and a Beta Web Content Crawler and Sitemap connector that crawls external sites with configurable depth. The retrieval layer runs RAG over that content through a vector database inside Data Cloud. To give a sense of scale, Salesforce's own internal "Customer Zero" deployment ingested roughly 740,000 content pieces to power its own Help site.

The practical catch is that every ingested chunk consumes Data Cloud credits, so a large knowledge base keeps costing you every month on top of the initial setup.

What features does Salesforce Agentforce have?

Agentforce's feature set breaks into three layers (we cover the same ground a different way, more on that in the pricing and improvement notes below).

The autonomous Service Agent handles ticket deflection, resolution, knowledge retrieval, and case classification and routing through Omni Flow. The Service Assistant copilot (GA Spring '25) does case summarization, multi-step action plans, and suggested next steps for human reps. The action framework offers five ways to actually do things: Apex code, Flow automations, Prompt Templates, MuleSoft API actions, and native Model Context Protocol (MCP) actions, which shipped with Agentforce 3 in June 2025.

Underneath sits the Command Center, a real-time dashboard for session tracing, latency, error and escalation analytics, and AI-tagged conversation analysis. Agentforce 360 added hybrid reasoning that blends deterministic workflows with LLM output, controlled through a new "Agent Script" language.

How do you improve Salesforce Agentforce's responses?

Improving Agentforce is a monitor-test-refine loop, and it needs an owner.

The Agentforce Testing Center (GA December 2024) generates AI test cases, runs batch CSV-based test suites, and executes sandbox actions at 16 Flex Credits versus 20 in production. The Command Center gives you the analytics to see where the agent is failing, across latency, errors, and escalations, and from there you refine Topics, Instructions, and the underlying knowledge. Like any agent, it needs someone watching the dashboards and closing the gaps.

What resolution rate can you expect from Salesforce Agentforce?

Salesforce markets "up to 72% of routine inquiries" deflected and "32% faster resolution times." The real answer, in my experience, depends entirely on how each customer defines "resolved," and the definitions are all over the place.

The customer numbers span an enormous range. Salesforce's own Customer Zero cites 77% resolution across 50,000+ conversations.

Published results include 1-800Accountant at 90% case deflection during tax week, Heathrow's "Hallie" at roughly 95% response accuracy, Reddit at 46% case deflection, OpenTable at 70% of inquiries resolved, Pandora at 60% deflection, and GE Appliances at 25%. Benioff himself has repeatedly pegged accuracy at around 93%, adding "you need the human in the loop."

Two things stand out to me here. Salesforce's own CRMArena-Pro benchmark found an out-of-the-box Agentforce agent hit "accuracy of nearly 35%" before customization, so the strong customer numbers come from heavy configuration. And there's no consistent definition of "resolution": some of those percentages are case deflection (the ticket didn't escalate), some are customer-confirmed resolution, and some are knowledge-deflected sessions where no ticket ever existed.

For a sense of where the field actually sits, our own aggregate across 195 rated AI-support deployments puts the median AI-handling rate at about 70%, with most vendors clustered between 56% and 80%. Treat that as directional. Every vendor counts a different event, so the field number is a rough gravity point, and the metric label moves it more than the underlying capability does.

What AI model does Agentforce use?

Agentforce is a multi-model platform. The default reasoning model is OpenAI's GPT-4o, with Anthropic's Claude Sonnet 4 hosted on Amazon Bedrock for regulated industries, Google Gemini support added at Dreamforce 2025, and GPT-5 selectability announced alongside it. You can also bring your own model (Mistral, Llama or Cohere) through the Models API and LLM Open Connector, which consumes 30% fewer Einstein Requests as a pricing nudge.

The much-hyped "Atlas Reasoning Engine" is the orchestration layer that plans and sequences actions; the LLMs sit underneath it. Salesforce describes it as "the brain behind Agentforce", using a Reason-Act-Observe loop rather than plain chain-of-thought.

Here I'd push back on the question everyone asks first, which is "what model does it use?" It's a fair question if you don't build these systems, but the wrong place to start.

Picking the model is the vendor's job. Models change constantly, and only the vendor can see across all their customers' data how each one moves resolution, accuracy and escalation.

On top of that, a real agent is never one model: it fires upwards of ten orchestrated LLM calls between question and answer, each doing a different job and tuned differently, so one rewrites the conversation, one detects the language, one applies your guidance and tone, one decides whether a task should run. We A/B test each of those steps continuously.

Agentforce's multi-model design is a sign of that maturity; a vendor who can answer "what model?" in a single word is telling you the setup is primitive.

What languages does Salesforce Agentforce support?

Agentforce supports a broad set of languages out of the box, with more on the way (Salesforce has 30+ additional languages on the Agentforce 360 roadmap). As with most platforms, some features lag the full language set, so if a specific channel-and-language combination is critical, I'd confirm it against the current documentation rather than the headline count.

How secure is Salesforce Agentforce?

Security is genuinely Agentforce's strongest dimension. It's the real reason a regulated enterprise picks it, and I'd give Salesforce full marks here.

The Einstein Trust Layer provides PII masking, prompt-injection defense, toxicity scoring, zero-retention contracts with the LLM providers, geo-aware routing for data residency, and a full audit trail stored in Data Cloud.

Agentforce also landed FedRAMP High authorization in 2025, which unlocks federal and public-sector procurement. If your buying decision is driven by compliance depth, this is where Salesforce earns its price.

Who is using Salesforce Agentforce?

Agentforce's customer base is overwhelmingly large enterprise, which is the clearest signal of who it's built for.

Named service customers include Wiley, OpenTable, FedEx, Saks Fifth Avenue, Heathrow Airport, Pandora, 1-800Accountant, Reddit, Air India and GE Appliances, spanning retail, travel, financial services, publishing and manufacturing. Salesforce disclosed that more than 60% of Agentforce and Data 360 Q4 bookings came from existing-customer expansion, and that every top-10 Q4 win included the full Agentforce 360 stack. The SMB proof points it parades, Safari365 (a 35-employee company) or Engine's 12-day deployment, are the exceptions.

How much does Salesforce Agentforce cost?

Agentforce has had three pricing models in eighteen months, and untangling them is the single most-searched thing about the product. I'll take them in the order they launched.

Conversation and Flex Credit pricing

Agentforce's conversation pricing launched in October 2024 at $2 per conversation, where a "conversation" ran from the agent's first response until the issue closed or 24 hours of inactivity passed. That model backfired publicly. Salesforce closed about 5,000 deals by early 2025 but only 3,000 were paid, and SaaStr wrote that the $2 rate "immediately priced out nonprofits, SMBs, and any organization without a large AI budget."

The replacement, Flex Credits, launched in May 2025: $500 per 100,000 credits, where one standard action costs 20 credits ($0.10) and one voice action costs 30 credits ($0.15).

The gotchas matter, and we field questions about these constantly. Unused credits don't roll over, you can't mix Flex Credits and conversation pricing in the same org, and overages bill monthly.

Per-user editions

Later in 2025 Salesforce added per-user add-ons on top of the consumption models (so these seats stack on top of the credit cost). The Agentforce add-on starts at $125 per user per month on Enterprise or Unlimited Edition, and the Agentforce 1 Editions run $550 per user per month with 2.5 million Flex Credits included per year. On August 1, 2025 Salesforce also raised list prices 6%, taking Service Cloud Enterprise to $165 per user per month, and it lists at $175 today.

The full Year-1 TCO

None of the AI pricing exists in isolation. It sits on top of the Service Cloud and Data Cloud licenses it requires. For a 30-seat mid-market team running 10,000 conversations a month, here's the first-year total I'd model:

Add it up and a realistic Year-1 total lands between $200,000 and $450,000+, depending on how much MuleSoft, extra Data 360 and systems-integrator time you need. There's no published Agentforce starter SKU for smaller teams. The Starter Suite ($25/user) and Pro Suite ($100/user) don't include it, so by design Agentforce is Enterprise Edition and up.

Play video

Does Agentforce have a free trial?

There's no true free trial of the paid product, but I count a few gated entry points.

Salesforce Developer Edition is free forever for building and testing. Salesforce Foundations gives Enterprise Edition and up customers $0 access with 100,000 to 200,000 Flex Credits and 250,000 Data 360 credits.

There's also a standard 30-day Service or Sales Cloud trial, and the Testing Center for sandbox runs. What you won't find is a "sign up and try Agentforce in production this afternoon" path, since the free tiers are aimed at developers and existing enterprise customers.

Is Salesforce Agentforce worth it?

Agentforce is worth it for a specific buyer and a poor fit for everyone else. Here's how I'd draw the line.

One piece of forward-looking context: Salesforce has also agreed to acquire Fin (the company formerly known as Intercom) for around $3.6 billion, a deal expected to close in Salesforce's FY2027, so the enterprise AI-support landscape is consolidating around it. That doesn't change the buying decision today, but it's worth watching if you're choosing a platform for the long term.

What are the Pros and Cons of Salesforce Agentforce?

Pros

• Deepest enterprise compliance grid: the Einstein Trust Layer plus FedRAMP High authorization is the strongest security-and-governance story in the category.
• A genuine CRM data moat: if Salesforce is already your system of record, Agentforce sits directly on your data with no integration gap.
• Broad channels and a real action framework: wide messaging coverage plus voice, and native MCP support for agentic actions.

Cons

• Salesforce-only: no native Zendesk, Intercom, Freshdesk or HubSpot support, so teams on those helpdesks can't reach it without migrating.
• Enterprise-only cost: a realistic Year-1 TCO of $200,000 to $450,000+ and no SMB starter SKU shut out most mid-market and smaller teams.
• Long, SI-dependent deployments: five-to-eleven-month production rollouts that usually need a systems integrator.
• You pay even when it escalates: per-conversation and per-action billing charges you whether or not the ticket is resolved, and the pricing model has changed three times in eighteen months.

FAQs