Current as of October 16, 2026. Updated October 16: Human in the Loop now covers chat as well as email, with new defaults and billing wording. Beta labels change fast; each status below links to the page it came from.
Fin 4 is the release Intercom (now trading as Fin) announced at Pioneer on October 7, 2026: new Apex 2.0 answer models, Human in the Loop, Skills, a second-generation Copilot, and Routines and Connections for Operator. Only one of those is already running in your workspace. The rest are betas or add-ons you switch on, and two of them change how a "Fin resolution" gets counted.
Fin 4 moves more conversations into "Fin resolved", including the ones a teammate quietly answered. Supportman posts every Intercom rating into Slack as it lands and scores the conversations nobody rated, so you can see whether the resolutions held.
Free trial · no card · about 2 minutes to set up
Here's what each piece is, whether you can use it today, and the setting I'd decide before it goes live. The launch post and the Fin 4 page are the announcement. The availability details come from the help center, which is where the caveats live.
Fin 4 at a glance
| Release | What it does | Status on Oct 16 | Decision for you |
|---|---|---|---|
| Apex 2.0 | New default answer model for Fin chat and email | Live for every Fin customer | None to switch on. Mark Oct 7 in your reporting. |
| Apex Flash 2.0 | Faster model behind Fin Voice, now in nine languages beyond English | Rollout scope not stated | Re-listen to a sample of voice calls |
| Apex for Procedures | Model trained for multi-step Procedures | Announced; no opt-in documented | Re-run your Procedure simulations |
| Human in the Loop | Fin pauses low-confidence answers and asks a teammate | Beta, email and chat, Intercom helpdesk only | Confidence threshold, who gets asked, the time limit |
| Skills | Goal, tools and instructions instead of a scripted path | Closed beta, via your account team | Which flow to try first, given there's no Skill reporting yet |
| Operator Routines, Connections, Memory, Artifacts | Scheduled and event-driven Operator work; reads Slack and GitHub | Announced; Operator itself is GA on the Pro add-on | What the first Routine should be |
| Copilot (second generation) | Drafts, investigates, tags and writes notes before a teammate opens the conversation | Announced | Whether Copilot's tags count in your topic reports |
| Fin for Success | Fin starts onboarding, adoption and re-engagement conversations | Announced | Who owns proactive outreach, and its frequency cap |
What is Fin 4?
It's the fourth major version of Fin, and the first since the company changed hands. Intercom renamed itself Fin in May, after its AI agent. Salesforce completed its acquisition on September 10. The inbox you work in is still Intercom, but the Fin 4 announcement lives on fin.ai, not the Intercom changelog.
The pitch is that Fin stops escalating and walking away. In practice that means three changes to the work your team sees: Fin asks a teammate a question instead of handing over the whole conversation, Fin works through problems nobody wrote a Procedure for, and Operator does recurring analysis without being asked each time.
Apex 2.0 is already answering your customers
You don't turn this one on. Apex 2.0 is live for every Fin customer, built on NVIDIA Nemotron 3 Super, and Fin's own figures against Apex 1.0 are:
- 8.6% more hard resolutions, meaning the customer explicitly confirmed Fin answered them
- 16.5% fewer hallucinations
- 1.7% higher resolution rate and 1.3% higher CSAT
Read those as vendor numbers on vendor-chosen conversations. The post doesn't say whether 8.6% is relative or percentage points, and it doesn't say how hallucinations were measured. The 1.7% resolution lift is the one that will show up in your dashboard, and on most workspaces it'll be inside normal week-to-week noise.
What to do: write "Apex 2.0, Oct 7" on whatever your team uses as a reporting annotation. Then compare the two weeks either side on the split between confirmed and assumed resolutions, not the headline rate. If the model is better, confirmed should grow as a share of the total. If only assumed grows, Fin got better at ending conversations, which is a different thing. Fin resolution states covers what each state counts.
The hallucination claim is the one your QA sample should test. If you run a weekly Fin review, keep the same sample size through October so the before and after compare. The five checks are in how to QA Intercom Fin.
Human in the Loop changes what an escalation is
This is the Fin 4 feature you can switch on today, and the one most likely to move your numbers. It's in beta on email and chat, and only with the Intercom helpdesk. Intercom rewrote the article on October 16: at launch it was email only, defaulted to a Moderate threshold, and said you'd be billed for "HITL resolutions".
Fin scores its confidence on each answer. Below your threshold, it holds the answer, writes its own question for a teammate, and assigns the conversation to whoever the rule names. On chat the customer sees "Let me check with the team. I'll be back shortly." On email they see nothing until Fin replies. The teammate can approve Fin's draft, give Fin the missing fact or decision (Fin rewrites the draft for them to approve), or take the conversation over. Handover offers skip the check entirely.
You set it up as an escalation rule (Fin AI Agent › Train › Escalation, action Ask for input), and every new rule starts with the same defaults:
| Setting | New-rule default | What I'd set |
|---|---|---|
| Channels | Chat and email | Two rules, one per channel. They need different time limits. |
| Confidence threshold | Low (slider to High; High checks every answer) | Keep Low for two weeks, then raise it if teammates approve most drafts without changes |
| Who to ask | All teammates | One team inbox. "Everyone" means nobody, the same as an unassigned queue. |
| Time limit | 24 hours, then send the original draft | Chat: just over your chat first-response p90. Email: just over your email p90. Never "wait indefinitely". |
The time limit deserves the most thought. If nobody answers in time, Fin sends its original draft, the one it scored as low-confidence. Intercom's advice is to set the limit above your team's typical response time, so it rarely fires. The 24-hour default is fine for email and wrong for chat, where the customer was just told Fin would be back shortly. Waiting indefinitely avoids the risky draft but leaves a customer stuck if the question lands in a queue nobody watches.
Check where your team will see the questions. They appear in the Inbox under Needs teammate input, and not in Slack. If your team works the queue from Slack, someone has to own that Inbox folder or the time limit fires every time.
Then the billing and the reporting. "There's no additional charge for Human in the Loop." Billing stays per resolution, and takeovers aren't charged. So a conversation Fin finishes after a teammate supplied the answer is billed as an ordinary Fin resolution. A worked example with numbers you should swap for your own:
| Line | Value |
|---|---|
| Fin conversations a month that match the rule | 2,000 |
| Paused for a teammate (assume 15%) | 300 |
| Of those, finished by Fin after teammate input (assume 80%) | 240 |
| Billed at $0.99 per resolution | $237.60 |
| Added to Fin's resolution count | 240 conversations a teammate shaped |
$237 a month is cheap if each one saved a teammate a full handover. The reporting is the bigger issue. Those 240 sit inside Fin's resolution rate, so Fin's rate rises while your team still did part of the work. There's now a Human in the loop report listing every such conversation and each teammate's average time to action. Use it from day one to keep these conversations separate. Otherwise, in December, someone will compare Fin's resolution rate with September's and credit the model for work your team did.
For Procedures there's a separate step that pauses for a teammate's approval before Fin acts, covered in Human in the Loop approvals for Procedures. That's the one for refund exceptions and account closures. If you already run approvals through Slack, this guide shows the pattern.
Skills are Procedures without the script
A Procedure is a path you write step by step. A Skill is a goal, a set of tools (data connectors, content, attributes, tags, handover targets, other Procedures) and written instructions, and Fin decides which tools to call and when it's done. Skills are in closed beta; you ask your account team for access.
Intercom's own comparison, from the getting-started article:
| Skills | Procedures | |
|---|---|---|
| Suited to | Open-ended goals: diagnose a failed payment, work out why email isn't delivering | Specific, well-defined cases that follow a set path |
| Behaviour | More flexible, less predictable | More consistent, closer to a script |
| Setup effort | Lower | Higher |
The line that matters in the beta article: Skill reporting is "coming soon", and "you won't be able to track Skill performance until this is released". Human in the Loop for Skills is also listed as coming soon.
So pick the first Skill for low volume and low cost of a wrong answer. Email deliverability diagnosis is a good candidate: the customer is already stuck, the tools are read-only lookups, and a bad answer means a handover rather than a refund. Keep anything that moves money in a Procedure until you can see a Skill's numbers. Fin's launch example is PrizePicks resolving 38,000 conversations with Skills in under 30 days, which is a scale you'd want reporting for.
Write the "Do not use when" examples as carefully as the "Use when" ones. The article recommends 10 to 20 trigger examples, and a Skill that fires on the wrong conversations is harder to notice than one that never fires.
Operator gets Routines, Connections and Memory
Operator is the agent for your support ops team: it analyses conversations, drafts content and Procedure changes, and proposes them for a human to approve. It became generally available on August 6 on the paid Pro add-on, with a monthly credit allowance. Fin says more than 1,600 customers use it.
Fin 4 adds four things:
- Routines: Operator runs on a schedule or on an event, such as a Monday performance report or an incident alert with a diagnosis.
- Connections: Operator reads Slack and GitHub, so it can draft Fin changes from a product announcement in Slack, or file a bug in GitHub.
- Memory: it keeps your preferences, like report structure, across sessions.
- Artifacts: shareable outputs, such as a drafted Procedure, for teammates to review.
As of today the Operator FAQ still says it can't send Slack messages, can't change workflows or Fin settings, and caps real-time analysis at 250 conversations per session. Expect the FAQ to change as Connections rolls out; until it does, plan around those limits. The proposal model hasn't changed: Operator drafts, a person approves, nothing goes live on its own.
My first Routine wouldn't be the performance report. You already have one, and a second report that disagrees with the first costs a meeting. I'd ask it for a weekly list of the conversations where Fin's answer contradicted a help center article changed in the last 7 days. That's the drift no dashboard catches, and it's the job Memory and the Slack connection were built for.
Copilot and Fin for Success
The second-generation Copilot does its work before a teammate opens the conversation: it pulls context from GitHub, Linear and Notion, drafts a reply, and recommends a next step. It also applies tags, writes internal notes, and prepares handovers, following guidance you define.
The tagging is the part to plan for. If you report on conversation topics by tag, the day Copilot starts tagging is the day your topic mix shifts, because Copilot tags more consistently than people do. Note the date, and keep a way to tell Copilot's tags from a teammate's if your setup allows it. Intercom conversation topics has the reporting side.
Fin for Success is Fin starting conversations: onboarding plans built from the customer's goals, nudges toward features they haven't used, re-engagement when usage drops, using data from Snowflake, Shopify and Salesforce. This belongs to whoever owns lifecycle messaging, not support. Before it goes live, agree a cap on how many conversations Fin can open per customer per month, and where replies to those conversations route. Otherwise the replies land in the support queue with no context on why Fin got in touch.
What else shipped since the July edition
Fin 4 got the stage, but the Intercom changelog kept moving. The releases from August 13 to October 16 that matter to a support ops team, dated as published on the Intercom changelog:
| Date | Release | Why it matters |
|---|---|---|
| Aug 13 | Fin remembers past conversations | Returning customers pick up where they left off, across channels |
| Aug 17 | Real Time Incident Detection | Alerts as a customer problem starts forming |
| Aug 18 | Take and drop permissions | Limit who can self-assign or unassign themselves |
| Aug 20 | Phone tab on the Inbox Dashboard | Live queue, agent status and SLA data for phone |
| Sep 1 | Configurable unsnooze controls | Internal notes and reassignments no longer have to wake a snoozed conversation |
| Sep 8 | Snowflake MCP Connector | Fin queries your own warehouse |
| Sep 8 | Ticket assignments dataset and lifecycle report | Time a ticket spends with every team that touches it |
| Sep 11 | Data connector templates | Stripe, Shopify and Statuspage connectors, generally available |
| Sep 23 | Custom metrics | Filter or divide built-in metrics, across your full history |
| Sep 28 | Merge duplicate users | Preview what moves, behind a new permission |
| Oct 6 | Macro fill-in fields | Macros prompt for details before inserting the reply |
| Oct 9 | Macro Admin Mode | Admins can view, share or delete teammates' personal macros |
| Oct 13 | SMS numbers by brand | Texts go out from the number assigned to the customer's brand |
| Oct 14 | Dictation in the Inbox | Teammates speak replies and notes, then review the text before sending |
| Oct 15 | Search inside a Messenger conversation | Customers find text in a long thread |
| Oct 15 | Workforce Management | Forecasting and scheduling, with the Pro add-on |
Custom metrics is the one I'd open this week. It's how you build "Fin resolutions excluding Human in the Loop" without exporting to a spreadsheet, which is the number the section above says you'll need. Workforce Management is the other one worth a look if you're already on Pro for Operator: forecasting is only as good as the volume history behind it, and Fin 4 is about to change how much of that volume reaches people.
What Fin 4 doesn't tell you
Every part of Fin 4 pushes the same direction: more conversations end as Fin resolutions. Human in the Loop keeps the resolution when a teammate supplies the answer. Skills resolve conversations nobody scripted. Apex 2.0 resolves a few more on its own.
None of them tells you whether those resolutions held. A customer who got a confident wrong answer through a Skill doesn't reopen the thread; they email again next week, or leave. The check is the same one that applied before Fin 4, run more often: CSAT on the conversations Fin handled, split by how they were resolved, plus the customers who came back after an assumed resolution. Measuring whether Fin is actually resolving conversations sets up the first, and the Fin resolutions that came back covers the second.
Give each change two weeks before you add the next one. Switch on Human in the Loop and a Skill in the same week and you won't know which one moved the number.
The running list of Intercom releases and the settings worth changing is in What's New in Intercom, and every Fin page on this site is indexed on the Fin hub.
Fin 4 FAQ
When was Fin 4 released?
Fin announced Fin 4 at its Pioneer event on October 7, 2026. Apex 2.0 went live for every Fin customer at launch. Human in the Loop is in beta on email and chat, Skills are in closed beta, and Operator is generally available on the Pro add-on.
Do I need to turn anything on to get Fin 4?
Not for Apex 2.0, which already answers your customers. Human in the Loop is an escalation rule you create, Skills need your account team to grant beta access, and Operator needs the Pro add-on.
Does Intercom charge for conversations a teammate helped Fin answer?
Yes, at the normal rate. Intercom says there's no additional charge for Human in the Loop and billing stays per resolution, so a conversation Fin finishes after a teammate's input is billed like any other Fin resolution. Takeovers aren't charged. Use the Human in the loop report to keep those conversations separate when you report Fin's resolution rate.
Should I replace my Procedures with Skills?
No. Intercom describes them as complementary. Keep Procedures for set paths and anything that moves money, and use Skills for open-ended diagnosis. Skill reporting isn't available in the beta yet, so start with a low-volume flow.
Is Intercom still called Intercom?
The company renamed itself Fin in May 2026, and Salesforce completed its acquisition on September 10. The helpdesk is still Intercom and its changelog is still at intercom.com, while Fin announcements, including Fin 4, are published on fin.ai.