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

    Intercom Fin

    Fin is paid per outcome, and Fin writes the definition of an outcome.

    Fin is sold per outcome instead of per seat. Here’s how it works, the decisions behind it, and what to check before you trust the number it bills you on.

    Fin AI Agent in an Intercom procedure simulation, replying “I can help you move your physio appointment. First, I need to verify your identity” after starting identity verification, above an Evaluation panel with green checks for Switched to procedure and Reschedule booking was not triggered.
    Image: Intercom simulation recording · 4:27

    At a glance#

    What it does

    Answers customers and finishes tasks like account updates, payments and troubleshooting, across chat, email, phone, WhatsApp, SMS and Slack.

    Price

    $0.99 per outcome, charged at most once per conversation. No seats required.

    Works with

    Intercom’s own helpdesk, or Salesforce, HubSpot, Freshdesk and others.

    Model

    Fin Apex 1.0, trained in-house. Since March 2026 it runs almost all English chat and email conversations.

    Claimed results

    76% average resolution rate across 12,000+ customers, and about 2 million resolutions a week.

    Owner

    Intercom renamed itself Fin in May 2026. Salesforce bought the company for about $3.6 billion; the deal closed September 10, 2026.

    Why it matters#

    Most software is sold per seat. Fin is sold per outcome: the customer pays $0.99 each time Fin finishes the job it was set up to do. Intercom built the product, renamed the whole company after it, and sold that company to Salesforce for about $3.6 billion.

    Fin charges per outcome, and Fin writes the definition of an outcome. That one definition sets both your bill and the resolution rate you report, so it’s the first thing to read. This page also covers Fin’s own model, the approval step inside Procedures, and its testing tools.

    How it works#

    What happens when a request reaches Intercom Fin.

    1. Learn the business

      The support team loads help content, policies and tone guidance. Answering questions needs no engineering work.

    2. Answer

      Apex, Fin’s own model, reads the conversation and answers from that knowledge.

    3. Act

      For a task, Fin runs a Procedure: plain-language steps mixed with branches and code, calling other systems through data connectors, an API or MCP. An account number can come from the customer’s stored record, so the customer can’t type in someone else’s.

    4. Ask a person

      If the Procedure includes a human-review step, it pauses. A teammate gets the request (in Slack it shows the procedure, the customer and recent messages) and decides. If nobody answers in time, the request escalates.

    5. Finish

      The conversation ends resolved, handed to a person, or escalated. Each ending is counted, and some are billed.

    6. Improve

      Teams test changes in simulations before release, and read monitors, CX Score and insights to find conversations to fix. Intercom calls the loop train, test, deploy, analyze.

    The decisions worth studying#

    What they chose, why, and what it costs.

    1It charges per outcome, and widened what counts as one

    Until March 2026, Fin billed per resolution: the customer said the answer helped, or left without asking for more. On March 12, Intercom switched to billing per outcome. An outcome now also includes a Procedure that gathers details, takes an action and hands the customer to a person, when the team set it up to do that.

    Intercom’s reasoning is fair. Full automation isn’t always the right ending, and a well-prepared handoff saves a support rep real time.

    The catch for buyers: the number you judge Fin by and the number on your invoice are now the same number, and the vendor defines it. An “assumed” resolution only means the customer stopped replying. They may have been helped, given up, or picked up the phone. Intercom’s docs also calculate the rate two ways: over all conversations, or over conversations Fin took part in. With 1,000 conversations, Fin in 600 of them and 300 outcomes, that’s 30% or 50% from the same work.

    2It trained its own model

    In March 2026, Intercom moved almost all English chat and email onto Apex, a model its 60-person AI group trained on billions of support conversations. Intercom says Apex beats GPT-5.4 and Claude Opus 4.5 on its support tests while being faster and cheaper. One gaming customer’s resolution rate went from 68% to 75% overnight.

    My read: when you’re paid per outcome, every point of resolution rate is revenue, and every model call is your cost instead of the customer’s. Owning the model gives Intercom more control over both.

    3Approval is a step in the program, outside the prompt

    When a Procedure needs a person, the builder adds a human-review step. Fin stops, a teammate decides, and Fin continues only down the branch the builder wrote for “approved.” No wording from the customer can talk the model past a step that lives in the program instead of the instructions.

    One gap: data Fin fetched before the pause stays in use unless the Procedure fetches it again. The customer’s account can change while the request waits, and the approved change still runs on the old picture.

    4Its tests grade what Fin did, apart from whether it finished

    Simulations run a Procedure against a scripted customer and sample data, then score it against criteria the team writes. In Intercom’s demo, a customer asks to move a physio appointment that already happened. The run completes, so “Procedure finished” turns green. The two checks that matter turn red: leave the booking alone, and tell the customer the appointment already took place.

    That’s the test doing its job. Finishing and doing the right thing are separate checks, and Fin shows them side by side.

    Intercom · Procedure simulationProduct screenshot
    Evaluation: a green check for “Procedure finished”, and red crosses for “Reschedule booking was not triggered” and “Fin replied: That the customer's appointment has already taken pl…”.
    The run finished. The checks for what Fin should and shouldn’t do both failed, and the panel shows all three results. Intercom simulation recording · 3:40

    What to watch#

    • The vendor defines your success metric

      Billing and reporting share one definition of an outcome, set by Fin. Write down exactly what counts, and over which conversations, before you compare quarters or vendors.

    • A passing simulation isn’t a live test

      Simulations skip the step that picks which Procedure to run, and they use sample data in place of your real systems. A pass doesn’t prove the real integration works.

    • Undo covers Fin, not your systems

      Fin can roll back its own configuration in one step. Reversing what a Procedure already changed in your billing or booking system isn’t documented.

    • New owner

      Fin is now part of Salesforce, which already sells its own service agent, Agentforce. Salesforce says Fin keeps working with the helpdesks companies already use. If yours isn’t Salesforce, watch the roadmap.

    What it means for your team#

    What to decide before you ship an agent like Intercom Fin.

    • Finance

      Before you sign, write down what an outcome is, which conversations it’s counted over, and which handoffs are billed. Your invoice and your success metric are now the same number.

    • Support

      Staff the approval queue: who answers, how fast, and where a request goes when nobody does. Read a sample of “assumed” resolutions every week, because a customer who left may have given up.

    • Product

      Write tests for what the agent must never do, then rerun them against a real test account. Simulations use sample data, so a pass there proves the logic, not the integration.

    • Platform

      Fetch the customer’s data again right before an approved change runs, and build your own undo for changes Fin makes in other systems. Fin’s rollback only restores Fin.

    Steal this

    1. Make approval a step in the program that no prompt can skip, with a timeout that escalates.

    2. Grade what the agent did, separately from whether it finished.

    3. If you charge per outcome, publish exactly what counts, and report confirmed and assumed results apart.

    Reference#

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

    Walkthrough · 5 frames

    From a simulated customer to a pass/fail evaluation#

    Frames from Intercom’s simulation recording in the help article “Run Simulations for Fin Procedures” (article dated August 10, 2026). A vendor recording on sample data, not a live test.

    1. Step 1 of 5: The admin sets up a simulated customer

      A Fin simulation named “Lapsed appointment booking” with Save and Run: step 1, Simulate customer path; Simulate as Test customer and All brands; and the customer’s opening message “Hi, I need to move my physio appointment to next week.”, partly under the cursor.
      Image: Intercom simulation recording · 1:55

      1. The simulation, “Lapsed appointment booking”, tests one case: a customer whose appointment has already passed.

      2. Fin talks to a simulated test customer, not a real one.

      3. The test customer opens with “Hi, I need to move my physio appointment to next week.”

    2. Step 2 of 5: A sample response stands in for bookings

      Get customer bookings · Default sample response, then “Data response provided to Fin” and JSON with customer_name “Sarah Murphy”, an email, and booking BK-10432: type “Physiotherapy - Initial”, a practitioner, a date in April 2026 partly under the cursor, time “10:30” and status “confirmed”.
      Image: Intercom simulation recording · 2:30

      1. The Get customer bookings connector returns a default sample response, so the simulation doesn’t call a real booking system.

      2. The sample hands Fin a confirmed physiotherapy booking for Sarah Murphy, dated April 2026.

    3. Step 3 of 5: The admin writes the pass/fail criteria

      Step 2, Evaluate Fin’s behaviour, with the rule “The simulation must meet all criteria to pass.”, Add criteria, and two criteria: Data connector “Reschedule booking was not triggered” and Fin reply “that Fin told the customer their appointment has a...”.
      Image: Intercom simulation recording · 2:55

      1. Intercom’s rule is strict: “The simulation must meet all criteria to pass.”

      2. One criterion checks an action: the Reschedule booking connector must not run.

      3. A second criterion checks Fin’s reply: the customer must hear that the appointment has already taken place.

    4. Step 4 of 5: The Procedure finishes, but two checks fail

      Evaluation: a green check for “Procedure finished”, and red crosses for “Reschedule booking was not triggered” and “Fin replied: That the customer's appointment has already taken pl…”.
      Image: Intercom simulation recording · 3:40

      1. “Procedure finished” is green: by completion alone, the run looks like a success.

      2. The no-reschedule check fails, a cue to inspect the simulated events; no real booking was involved.

      3. The reply check fails too: the evaluation doesn’t credit Fin with saying the appointment already took place.

    5. Step 5 of 5: A later run passes changed criteria

      Evaluation: green checks for “Switched to procedure” and “Reschedule booking was not triggered”, with the cursor over the second.
      Image: Intercom simulation recording · 4:27

      1. The first check now reads “Switched to procedure”, not “Procedure finished”: the criteria changed between the two runs.

      2. “Reschedule booking was not triggered” now passes, but the recording can’t credit the pass to one fix, since the test changed too.

    Hands-on session · 5 frames

    Fin’s citation worked. The written link didn’t.#

    A live conversation in Intercom’s public Help Center. Fin explained how to pause a channel, stated the chat’s limits, and corrected a broken help link.

    1. Step 1 of 5: Fin returns two links to the same article

      Fin’s answer has a numbered citation after the pause instructions and a written help URL containing article ID 8279188.
      Image: Original browser capture · September 29, 2026

      1. Fin’s numbered citation opened article 8286630, the working deployment guide.

      2. Fin’s written URL used article 8279188, a different destination.

    2. Step 2 of 5: Fin explains the chat’s action boundary

      Asked whether the Help Center chat can turn Fin off, Fin says the chat can explain steps but cannot change a customer’s workspace or take account actions.
      Image: Original browser capture · September 29, 2026

      1. Fin said the public chat could explain the steps but could not change a customer’s workspace.

    3. Step 3 of 5: The written URL opens a 404

      Intercom Help displays a 404 notice for article 8279188, with a search field below the notice.
      Image: Original browser capture · September 29, 2026

      1. Intercom’s Help Center displayed a 404 for the written link; the numbered citation loaded the article.

    4. Step 4 of 5: Fin returns the working URL after feedback

      The researcher identifies both article IDs and asks for the canonical URL only. Fin replies with the working 8286630 help link.
      Image: Original browser capture · September 29, 2026

      1. Fin returned the working article URL after the correction named the failed and successful destinations.

    5. Step 5 of 5: Messages keeps the earlier conversation after reload

      Intercom Messenger’s Messages screen lists ‘Pause agent on channel’ with the corrected help URL in the preview.
      Image: Original browser capture · September 29, 2026

      1. Intercom’s Messages list kept the earlier conversation; selecting the entry restored all three prompts and responses.

    What a builder can take from the session

    A support team should check the destination link separately from the answer text. A useful implementation renders links from structured source records and tests the resolved destination. The working citation and the broken written URL were separate outputs in the same answer.

    The return path

    Closing and reopening the messenger kept the conversation. After a page reload, the launcher opened a fresh chat. Messages still listed the earlier session, and selecting the entry restored the original conversation.

    Exact prompts from the session

    1. Ask for instructions

      In a test Intercom workspace, how can an admin pause Fin on one channel without deleting the agent? Please link the relevant help article. This is product research; please keep the conversation with the AI and do not contact a human teammate.

    2. Check the action boundary

      Are you able to turn off Fin for a customer's workspace from this Help Center chat, or can you only explain the steps? Please do not make any changes or start a human handoff.

    3. Correct the destination

      The written /8279188-deploy-fin-ai-agent-over-chat link opens a 404 page. The numbered source opens /8286630-deploy-fin-ai-agent-over-chat and works. Please return the working canonical help link only. Please do not start a human handoff.

    Sources#

    1. Fin pricingintercom.com

    Sources checked September 28, 2026.

    Next teardownAgentforce Service

    Agentforce Service is Salesforce’s customer-facing agent.