Skip to content
Product Agents

Search Product Agents

Start with a guide, or search a topic

    New here? Read the five guides in order

    Subscribe

    Agentforce Service

    Agentforce has changed its billing unit four times in two years.

    Salesforce’s service agent has charged per conversation, per action, per seat and now per resolution. The last price arrived with a $3.6 billion purchase of Fin. Here’s what each change says about building an agent into a product.

    Agentforce Builder for the Coral Cloud Service Agent, showing a booking action whose required contactId input is assigned the Verified Customer Id variable with Collect data from user unchecked, beside a Conversation Preview with the Agentforce robot and “Let’s chat!”.
    Image: Salesforce Developers recording · August 21, 2025 · 4:15

    At a glance#

    What it does

    Salesforce’s customer service agent. It answers customers on chat, email, voice, WhatsApp and SMS, runs actions on their records (verify identity, create a booking, update an account) and transfers them to a person.

    Price

    Four ways. $2 per conversation (since October 2024), 10 cents per action in Flex Credits (since May 2025), $125 to $550 per user per month, and $2 per resolution for the prepackaged Help Agent (since July 2026).

    Model

    New agents default to GPT-4.1. Claude Haiku 4.5, hosted inside Salesforce’s own Amazon Bedrock, and Gemini 3.5 Flash are the other options. Since August 26, 2026, Claude is a reasoning model for Atlas, the engine that plans each step.

    Takes actions?

    Yes. Actions are Flows, Apex code, prompt templates or API calls the team already owns. On messaging channels the agent runs them as its own Salesforce user.

    Claimed results

    Agentforce annual recurring revenue passed $1.5 billion in the quarter ended July 31, 2026, up over 240%. On Salesforce’s own help site it has handled 4.3 million inquiries and resolved 70%. Marc Benioff says Salesforce cut its own support staff from 9,000 to about 5,000.

    Owner / Status

    Salesforce. Since September 10, 2026 it also owns Fin (formerly Intercom), bought for about $3.6 billion and sold alongside Agentforce as the fast-to-deploy option.

    Launched

    Generally available October 29, 2024. Agentforce 360 followed on October 13, 2025. Seven named prebuilt agents, including Casey the Help Agent and Fin, arrived September 11, 2026.

    Why it matters#

    Agentforce is the agent Salesforce sells to run customer service inside its CRM. In two years it has carried four prices: $2 a conversation at launch, 10 cents an action from May 2025, a flat $125 to $550 a seat, and since July 2026, $2 per resolved issue. In June 2026 Salesforce also agreed to pay about $3.6 billion for Fin, a finished service agent that charges 99 cents per outcome.

    My read: each billing unit matched what Salesforce was selling at the time. A kit of building blocks sells by the action; a finished agent sells by the result. This page covers the four prices, the Fin purchase, the agent user and the written rule for a billable resolution.

    How it works#

    What happens when a request reaches Agentforce Service.

    1. Ask

      A customer writes “Book the zip line activity for two guests for tomorrow.” in web chat. Booking is a protected action, so the agent asks for an email address before anything else.

    2. Verify the customer

      The agent emails a one-time code. Three wrong codes restart the process. This lookup runs with system-level permissions, because nobody is verified yet. Success stores a Verified Customer Id.

    3. Confirm the exact request

      The agent restates the booking, two guests at 9:00 AM on July 9th, and the customer says “Yes”. A builder can also make any action require a confirmation step in the platform’s own settings, where the prompt can’t skip it.

    4. Act

      The booking action is a Flow, Salesforce’s no-code automation. Its customer field is filled from the stored Verified Customer Id, never from the chat. On messaging channels the Flow runs as the agent user, a Salesforce user record holding the permissions the agent needs.

    5. Hand off

      If the agent can’t help, its Escalation subagent runs an Omni-Channel flow, Salesforce’s routing tool, to a rep or queue, passing the conversation history. It tries once per session; if no rep is free, the agent carries on with the same context.

    6. Count and bill

      For the Help Agent, a session with two turns, no negative feedback and no escalation bills $2. For agents you build yourself, a model reads the session and scores deflection from 1 to 5.

    The decisions worth studying#

    What they chose, why, and what it costs.

    1It changed its billing unit four times in two years

    Agentforce Service Agent went on sale October 29, 2024 at $2 per conversation. On May 15, 2025 Salesforce added Flex Credits: $500 buys 100,000 credits and a standard action costs 20, so an action costs 10 cents. The pricing page now also sells seats, $125 per user per month for the add-on and from $550 for Agentforce 1 Editions. The fourth unit is $2 per resolution for the Help Agent, the prepackaged service agent on general sale since July 2026.

    Salesforce’s release note gives the reason for the fourth change: metering the Help Agent by action “made costs difficult to predict for high-volume deployments.” To investors it reports a fifth unit, Agentic Work Units, 7 billion so far. One unit can be a prompt processed or a tool invoked.

    My read: the unit tracks what Salesforce is selling. A kit of Flows and prompts sells by the piece. A finished agent sells by the result. Ask which unit a vendor leads with before you trust the demo.

    2It paid $3.6 billion for a finished agent

    On June 15, 2026 Salesforce agreed to buy Fin, formerly Intercom, for about $3.6 billion. The deal closed September 10. Fin is a finished agent: 30,000 companies use it, it claims a 76% resolution rate, and it charges 99 cents per outcome with no seats. Salesforce said Fin’s “packaged offerings and proprietary models will complement Agentforce’s deeply customizable platform.” At closing it framed the choice as “rapid deployment on existing systems” against “deeply tailored, enterprise-scale transformation.”

    Ten days after the Fin deal, Salesforce announced its own Help Agent, set up from one screen and billed per resolution. On September 11 it named seven prebuilt agents. Two do customer service: Casey, the Help Agent, and Fin, which keeps its own Apex models.

    My read: buying Fin concedes that most support teams wanted a finished agent. Casey is $2 per resolution. Fin is 99 cents per outcome. One company now sells two per-result service agents at two prices, plus the kit. Ask which one your account team is paid to sell.

    3The customer comes from a verified ID. The agent runs as its own user.

    In Salesforce’s Coral Cloud sample, booking is a protected action. The agent emails a one-time code, and three wrong codes restart the process. Success stores a Verified Customer Id. The booking action, a Flow, fills its contactId from that stored value. “Collect data from user” is unchecked, so a request naming someone else can’t redirect the booking.

    That answers who the customer is. It doesn’t answer whose permissions run the action. On channels where customers aren’t logged in, such as messaging, the agent runs as an agent user: a Salesforce user record with every permission the agent needs. The verification flow itself runs with system-level permissions, because the lookup happens before anyone is verified. An action can also require a confirmation step set in its own definition, outside the prompt.

    The risk: a polished preview can stand in for identity. Salesforce’s newer sample ships a verification stub that accepts any code. Write down, per action, the customer, the running user and the record that changes.

    Salesforce · Booking action in Agentforce BuilderProduct screenshot
    Agentforce Builder, the booking action: Loading Text “Booking experience...”; input 1, contactId, “Contact record ID that will be used to create the related booking.”, Data Type lightning__textType, Require input checked, Collect data from user unchecked, Assign a Variable “Verified Customer Id”.
    “Collect data from user” is unchecked. The booking’s customer comes from the Verified Customer Id the verification step stored, so the chat can’t change it. Salesforce Developers recording · August 21, 2025 · 4:15

    4A resolution is defined by three rules. A model grades everything else.

    For the Help Agent, Salesforce wrote the billing rule down. A session is a billable resolution when it has at least two turns, the customer’s last feedback is not explicitly negative, and it ends without escalation to a person (or the last feedback is explicitly positive). Anything else is free, and actions inside a resolved session are included. A chat resolution covers up to two hours. A voice resolution covers ten minutes; a resolved 12-minute call bills as two.

    For agents you build yourself, the scoreboard changed on July 16, 2026. Deflection is now a 1 to 5 score from a model reading the session, and a 3 lands in an “ambiguous” bucket. Before that, the metric came from runtime events like end-conversation clicks and timeouts, and Salesforce once counted timed-out sessions as deflected.

    My read: Salesforce built the resolution meter before it started charging per resolution. A written rule beats a model’s opinion, with one gap: a customer who asks twice and leaves quietly is a $2 resolution.

    What to watch#

    • Quiet exits are billable

      Two turns, no complaint, no escalation: that is a $2 resolution, whether or not the customer got help. Read a weekly sample of resolutions that ended in silence.

    • Salesforce’s own numbers disagree

      Its June 2026 announcement says the help site resolved 70% of 4.3 million inquiries. Its service page says Agentforce resolves 85% of Salesforce’s customer service requests. Ask which denominator any rate uses.

    • Three service agents from one vendor

      Build-your-own Agentforce, the prepackaged Casey, and Fin with its own models now sit in one lineup. Salesforce frames Fin as rapid deployment on existing systems and Agentforce as deeply tailored work. Ask which one gets the roadmap before you commit.

    • The default model and the headline partner differ

      New agents default to OpenAI’s GPT-4.1 while the Claudeforce deal makes Claude the default for Agentforce Vibes and Coworker. Salesforce says to retest custom actions and prompts after any model switch, so budget for that.

    What it means for your team#

    What to decide before you ship an agent like Agentforce Service.

    • Finance

      Model your volume under both units before signing. Per action is 10 cents in Flex Credits with every call metered; per resolution is a flat $2 that includes the actions, and Salesforce won’t run Flex Credits and conversation pricing in the same org.

    • Security

      For every action, write down who the customer is, which Salesforce user runs it and which record changes. On messaging channels the agent’s own user runs the action, and verification runs with system-level permissions.

    • Platform

      Fill each action’s customer field from the Verified Customer Id in configuration, and require the confirm step in the action’s definition. Salesforce doesn’t document how to cancel a booking the agent made, so build that undo yourself.

    • Support

      Staff the queue the Omni-Channel flow routes to, and decide what the agent says when nobody is free. It tries the transfer once per session, and the handoff can only be tested on a live channel, never in the builder.

    Steal this

    1. Bill for what you sell: a kit by the action, a finished agent by the result.

    2. Take the customer for every action from the verified ID, never from what the customer typed.

    3. Write the billing rule for a resolution before you charge, and report quiet exits separately.

    Reference#

    The evidence behind this teardown: recordings, sessions, what Agentforce Service gets right and wrong, and every source.

    Walkthrough · 5 frames

    From “Book the zip line activity” to the customer ID the booking uses#

    Frames from Salesforce Developers’ customer-verification recording, August 21, 2025, built on Salesforce’s fictional Coral Cloud sample. A vendor recording, not a live test: no frame tries to book for a different customer.

    1. Step 1 of 5: The agent asks who the customer is

      Coral Cloud Service Agent chat: the customer writes “Book the zip line activity for two guests for tomorrow.”, and the agent replies “To proceed with booking the zip line activity, please provide your email address or username for verification. Once verified, I'll be able to assist you with the booking.”
      Image: Salesforce Developers recording · August 21, 2025 · 0:55

      1. The customer asks the agent to book the zip line activity for two guests “for tomorrow”.

      2. The agent asks for an email address or username for verification before any booking.

    2. Step 2 of 5: A code verifies the customer

      The agent writes “If you have provided a valid email or username, you should receive a verification code to verify your identity. Please enter the code.” The customer sends 926279, and the agent replies “Your verification was successful! You can now proceed with booking the zip line activity for two guests”. The message box holds an unsent “Go ahead and book the activity”.
      Image: Salesforce Developers recording · August 21, 2025 · 1:25

      1. The customer sends the code from the verification email.

      2. Verification answers who the customer is; the reply names no session, date or time yet.

      3. The customer’s next message, still in the message box, is a broad “Go ahead and book the activity”.

    3. Step 3 of 5: The agent confirms the exact booking

      After the customer’s “Go ahead and book the activity”, the agent asks “Would you like me to proceed with booking the zip line activity for two guests at 9:00 AM on July 9th?” The customer answers “Yes”, and a spinner shows “Booking experience...”.
      Image: Salesforce Developers recording · August 21, 2025 · 1:35

      1. The agent restates the booking before acting: the zip line, two guests, 9:00 AM on July 9th.

      2. The customer’s “Yes” agrees to this exact booking, a separate step from proving identity.

      3. The spinner’s “Booking experience...” is the booking action’s loading text, set in Agentforce Builder (step 5).

    4. Step 4 of 5: The Contact shows a matching booking

      Salesforce Contact record for Sofia Rodriguez on the Related tab: “We found no potential duplicates of this Contact.” Bookings (4), “4 items • Sorted by Date • Updated 19 minutes ago”; the first row is Canyon Zip Line Excursion, 09/07/2025, 09:00 to 12:30, 2, 300,00 €.
      Image: Salesforce Developers recording · August 21, 2025 · 1:45

      1. The recording then shows Sofia Rodriguez’s Contact record in Salesforce, on the Related tab.

      2. The first booking matches the confirmation: Canyon Zip Line Excursion, 09/07/2025 at 09:00, two guests, 300,00 €.

      3. The list reads “Updated 19 minutes ago”, and the recording doesn’t show which run created the row.

    5. Step 5 of 5: The booking takes the verified customer’s ID

      Agentforce Builder, the booking action: Loading Text “Booking experience...”; input 1, contactId, “Contact record ID that will be used to create the related booking.”, Data Type lightning__textType, Require input checked, Collect data from user unchecked, Assign a Variable “Verified Customer Id”.
      Image: Salesforce Developers recording · August 21, 2025 · 4:15

      1. The action’s Loading Text, “Booking experience...”, is the line the chat showed after the customer’s “Yes”.

      2. Collect data from user is unchecked, so the configuration doesn’t take contactId from the chat.

      3. Agentforce Builder assigns contactId the Verified Customer Id that verification stored; whose permissions run the action is a separate question.

    Sources#

    1. Agentforce pricingsalesforce.com
    14. Fin pricingintercom.com
    31. Verify Customerhelp.salesforce.com

    Sources checked September 28, 2026.

    Next teardownHubSpot Customer Agent

    HubSpot’s Customer Agent answers customers from the team’s knowledge base, can read and update their CRM record, can call the team’s own systems, and hands off to the support team.