Short answer
AI connects to Intercom through Fin, Intercom’s AI agent, which is billed per outcome, through the REST API and webhooks, or through Intercom’s MCP server for US and EU hosted workspaces. For New Zealand support teams the most valuable integration around Fin is an independent quality check that grades its answers daily, alongside escalation rules that hand the right conversations to people.
- Fin costs $0.99 per outcome, and an assumed resolution counts as an outcome.
- Escalation Rules and Escalation Guidance override Fin’s default handoff behaviour.
- Intercom’s MCP server does not yet support Australian-hosted workspaces.
- For AU-hosted workspaces, Intercom says Fin’s processing is still US-based.
What Intercom’s built-in AI does today
Intercom’s AI agent is Fin, billed per outcome rather than per seat. On 19 October 2026 Intercom listed $0.99 per resolution, procedure handoff, disqualification or self-serve routing, and $9.99 per sales qualification, with at most one outcome charged per conversation (Fin outcomes). A resolution includes an “assumed resolution”, where the customer leaves without asking for more help; the charge is reversed if they reopen the conversation. Fin can also run on other helpdesks, including Zendesk and Salesforce, at the same rate, and Fin Voice is $1.99 per voice outcome through sales (Fin pricing). Intercom shows “$” without a currency code.
For teammates, Copilot is an inbox assistant add-on at $29 per teammate per month, free for 10 conversations per agent each month.
Integration routes and when to use each
| Route | What it gives you | Use it when |
|---|---|---|
| Fin configuration | Content sources, procedures, escalation rules and guidance | Changing what Fin says and when it hands over |
| REST API and webhooks | Conversations, contacts, notes and events; 10,000 calls a minute per app | Quality checks, reporting and handoff enrichment in any region |
| MCP server | 14 tools to search and fetch conversations and contacts, manage articles and add internal notes | US or EU workspaces using Claude or other MCP clients |
| Fin Operator tools over MCP | Ask Operator to do work in the workspace and review its proposals | Workspaces with Fin Operator, connected by OAuth |
| n8n, Make or Zapier | Intercom triggers plus AI steps | Light routing and alerts |
The MCP server note tool “cannot send customer-visible replies”, which makes it a safe write path for quality grades. It is “available for US and EU hosted workspaces. AU hosted workspaces are not yet supported.” If your New Zealand business chose Australian hosting, check your workspace region before planning on MCP. See MCP for business systems for the wider design.
Automation 1: a daily answer quality check
- Inputs
- Yesterday’s conversations Fin answered, with outcome, the content Fin cited and any rating.
- Sampling
- Every assumed resolution, every conversation with negative sentiment, and a random tenth of the rest.
- Model step
- A different model from Fin grades each conversation against a written rubric: supported by your content, correct under policy, missed an escalation trigger, acceptable tone. Each judgement quotes the line it relied on.
- Human review
- The support lead reviews every fail and 20 random passes a week, to measure whether the grader agrees with the team.
- Write-back
- An internal note with the grade and a tag such as qa_fail_grounding. Fixes go to the content source or escalation guidance.
- Measures
- Fail rate by topic, wrong resolutions per week, grader agreement with people.
| Grader output | Intercom field | Rule |
|---|---|---|
| Verdict | Conversation tag (qa_pass, qa_fail_grounding, qa_fail_handoff) | One verdict tag per conversation |
| Reasoning and quotes | Internal note | Notes never reach the customer |
| Topic | Conversation attribute | Lets you report fail rates by topic |
| Fix needed | Task for the content owner, outside Intercom | The source gets fixed, not just the reply |
The methods are those in our guides to AI agent evaluations and monitoring AI quality.
Automation 2: handoffs a teammate can act on
- Rules for data. Add Escalation Rules for facts you hold: order value over a threshold, a VIP attribute, an overdue account.
- Guidance for judgement. Add Escalation Guidance for cases such as bereavement, safety concerns or legal threats, under Train > Escalation (Intercom Help).
- Review before sending. For high-risk rules, use the beta Ask for input option so a teammate approves Fin’s draft first.
- Enrich the handoff. On escalation, your integration adds an internal note with the customer’s recent orders and account status from your own systems, so the teammate does not re-ask.
Permissions and approval design
Grading needs read access only. Writing grade notes needs the Write conversations permission, and Intercom attributes every MCP note to the admin who authorised the connection, so authorise with a dedicated QA teammate rather than the head of support. Fin Operator tools appear only over OAuth, not bearer tokens. Keep the grader unable to reply to customers by design, not by instruction.
Costs and limits
Illustrative example: 3,000 conversations a month with Fin resolving half costs about $1,485 in outcomes. Grading 600 of them at 6,000 input and 400 output tokens on Claude Sonnet 5.5 (US$2 and US$10 per million tokens, 19 October 2026) costs about US$10. The check costs under 1% of the Fin bill, and each wrong assumed resolution it catches was billed at $0.99 and left a customer unhelped. Private apps get 10,000 API calls a minute per app and 25,000 per workspace (rate limiting).
Where it goes wrong
- Assumed resolutions. A customer who gives up looks like a success in reports and on the invoice. Grade them all.
- One offer only. Intercom says Fin “will never offer escalation twice in a row”. Make hard triggers rules that hand over, not offers.
- Grader drift. A grader that agrees with Fin is not evidence. Recalibrate against people monthly.
- Data location. For AU-hosted workspaces, Intercom’s Fin FAQ says processing is “currently still US-based”. Under the Privacy Act 2020 a provider processing only on your behalf is not a disclosure, but you remain responsible, so say so in your privacy statement.
- Manual recovery. When the grader finds a wrong answer, a teammate contacts the customer, the content is fixed, and similar conversations from that day are re-graded.
Compare AI and Zendesk. We set up evaluation for support agents through our LLM evaluation services and AI customer support systems.
Sources
- Fin AI Agent outcomes, Intercom Help, accessed 19 October 2026
- Fin pricing, Intercom, accessed 19 October 2026
- Manage Fin AI Agent’s escalation guidance and rules, Intercom Help, accessed 20 October 2026
- Fin AI Agent FAQs, Intercom Help, updated 11 September 2026
- Model Context Protocol (MCP), Intercom Developers, accessed 19 October 2026
- Rate limiting, Intercom Developers, accessed 19 October 2026
- Claude API pricing, Anthropic, accessed 19 October 2026
Written by the Levy Street team in Wellington. Last updated 19 October 2026.