Our read
Decide by the helpdesk you already run. On Zendesk, test Zendesk's own AI agents first; on Intercom, Fin is the native choice. Intercom also states that Fin can run on top of Zendesk, Freshdesk, Salesforce or HubSpot, so a migration is not a precondition. Do not compare $0.99 per outcome with Zendesk's per-automated-resolution charge until you have mapped both definitions onto your own tickets.
Intercom Fin and Zendesk AI across the five criteria of one ticket journey, from vendor pages
| Criterion | Intercom | Zendesk AI |
|---|---|---|
| Existing helpdesk | Intercom Native on Intercom; connectors for Zendesk, Freshdesk and others Vendor-stated; no customer case reviewed | Zendesk AI Native on Zendesk; Forethought agents stated to run on any platform Which product serves a non-Zendesk stack: to confirm |
IntercomFin is built into the Intercom helpdesk. Intercom's connector directory also lists Fin on Zendesk, Freshdesk, Salesforce Service Cloud, HubSpot Service Hub and any helpdesk over API.[1] [2] Zendesk AIZendesk AI agents sit inside the Zendesk platform. Zendesk says its Forethought agents also run on other platforms.[3] [4] | ||
| Knowledge sources | Intercom Existing content plus Knowledge Hub imports | Zendesk AI Knowledge graph unifying third-party content |
IntercomFin trains on existing knowledge. The Knowledge Hub imports Zendesk help center articles and consolidates Confluence, Notion and Guru content.[1] [5] [2] Zendesk AIZendesk describes a knowledge graph that builds a knowledge base in minutes, unifies third-party sources and optimises content from recent conversations.[3] [4] | ||
| Actions | Intercom APIs or MCP, with teammate approval shown | Zendesk AI Actions across business systems; action list not published Action Builder from Suite Team, usage-billed |
IntercomFin pulls customer data and acts through APIs or MCP; Procedures can trigger approvals, and Fin can ask a teammate before finishing.[1] [5] [2] Zendesk AIAgents execute actions across business systems on multi-step requests. Babbel and Vimeo connected them to their own APIs.[4] [3] | ||
| Escalation | Intercom Escalates when confidence thresholds are not met | Zendesk AI No thresholds described; Babbel uses topics Rules to confirm in a trial |
IntercomIntercom states that conversations are monitored against your criteria and Fin hands over to a human instead of guessing.[1] [5] Zendesk AIZendesk's AI page gives no handoff rules. In the Babbel story, defined topics decide what the AI agent resolves and what goes to people.[3] [4] [6] | ||
| Billing unit | Intercom $0.99 per outcome plus seats, with spending limits Outcome definition set by Intercom | Zendesk AI Per automated resolution, without human escalation Per-resolution amount not published |
IntercomAt most one outcome per conversation; nothing if Fin cannot answer, fails an action or the customer asks for a human.[5] [1] Zendesk AIAI agents are billed per automated resolution on top of seats; a resolution means no escalation to a human.[4] [3] | ||
Sources
Start from the helpdesk you already run
If your team is on Zendesk, the shortest path is Zendesk's own AI agents: its pricing page says they come with Suite and Support plans and are charged per automated resolution. If you are on Intercom, Fin is the native option and shares the inbox and customer record with the helpdesk.
Neither choice locks you in on paper. Intercom's connector directory lists Fin running on Zendesk, Freshdesk, Salesforce Service Cloud and HubSpot Service Hub, with the existing tool kept as system of record, plus an API route for other helpdesks. Standalone Fin is priced at $0.99 per outcome and sold through sales. Zendesk says its Forethought agents work on any platform. These are vendor statements, and none of the customer examples below runs a cross-helpdesk setup.
Vanta shows the opposite move: it moved its helpdesk to Intercom and made Fin its front-line agent. Intercom reports a 71% Fin resolution rate for that account; treat it as Intercom's own figure, with the definition unconfirmed here. At Neo Financial, the product lead describes the draw as one connected system for AI and human support.
Knowledge: what each agent answers from
Fin starts from the knowledge you already have, then you train it on your processes and policies, and Simulations let you run test conversations before going live. The Knowledge Hub imports Zendesk help center articles with categories and links intact and consolidates Confluence, Notion and Guru content, so a Zendesk team can move its articles without rewriting them.
Zendesk describes a knowledge graph that unifies service content, builds a knowledge base in minutes, pulls in third-party sources and tunes content from recent conversations. Vimeo, a Zendesk customer with a six-person Digital Support unit, draws the practical lesson: its AI agent is only as good as its content, so writers and developers joined the work, and the conversation logs now feed product feedback. Whichever engine you pick, someone has to own content upkeep.
Actions and approvals before a refund
Fin pulls customer data and acts through APIs or MCP, with connectors such as Shopify, Stripe and Jira. Its Procedures handle multi-step workflows, apply business rules and trigger approvals, and Intercom shows Fin asking a teammate for approval, then finishing the conversation itself.
Zendesk says its agents execute actions across business systems on multi-step requests; Action Builder appears from the Suite Team plan and is billed on consumption beyond the plan allowance. Babbel opened API endpoints so subscribers manage their subscription alone, and Vimeo used SSO and APIs to tailor answers by account type, with scripted flows for billing and generative answers for the long tail of FAQs. Whether a Zendesk agent can pause for human approval before a refund is a question for the demo (our suggestion, not a documented capability).
Escalation: what the human receives
Intercom states that every conversation is monitored against your criteria and that Fin escalates to a human when confidence thresholds are not met, rather than guessing. The handoff lands in the same inbox where the team already works.
Zendesk's AI page describes no handoff thresholds, so its customer stories carry the detail. Babbel uses a set of topics that decide what the AI agent may resolve and what goes to people, for example data-protection requests. Vimeo built the agent to recognise user identity and plan and to escalate when needed. BritBox pairs agents with Copilot so humans work faster on what remains.
A test we would run in a trial (editorial recommendation): take a handful of your real tickets, including a refund above your limit, a data-protection request and a frustrated repeat contact, and watch which agent hands over, when, and with what context.
Cost: an outcome is not an automated resolution
Intercom lists $19 per seat per month plus $0.99 per outcome. It charges at most once per conversation, nothing if Fin cannot answer, fails an action or the customer asks for a human, refunds the outcome if the conversation reopens, and offers spending limits. Zendesk bills seats ($55 for Suite Team, $115 for Suite Professional, Copilot add-on $50) and charges AI agents per automated resolution, meaning a request resolved without escalation to a human. The per-resolution amount is not published there.
Zendesk's own pages disagree slightly: the FAQ says AI agents are included in every Suite and Support plan, while the plan cards list them from Suite Team. Confirm which tier you need. The seat tiers are not like for like either. Vendor-reported rates, such as 47% automation at BritBox, 30 to 40% at Vimeo or 43% of chats at Docebo, come from different populations and metric definitions and do not rank the tools.
Method and limits
2026-11-02
This dated snapshot counts distinct companies in public customer evidence. CONFIRMED means a case study, named testimonial or press source; DECLARED means the company itself told usedby; OBSERVED means a vendor listing; SIGNAL means a job post. All four count, but their strength differs. Removed and unlevelled lines are excluded. Examples are a capped selection verified within 90 days. Vendor facts are sourced records checked within 60 days. Coverage is uneven and missing evidence is unknown. These figures measure neither market share nor product performance; capture dates are not adoption dates.