Our read
Settle four things before Fin answers a real customer: the content it reads, the customer data it may look up, the actions it may take, and the topics it must hand to a person. That order is our recommendation, not Kalshi's. Kalshi's published rollout documents only an import with batch tests, then rules and escalation topics stress-tested by legal and compliance, then go-live. Treat vendor resolution rates as claims, not your target.
Start with the content Fin will answer from
Intercom says Fin starts from the knowledge you already have and is then trained on your processes and policies. Its Knowledge Hub can gather articles from a help center, Zendesk, Confluence, Notion and Guru, and a Zendesk help center import keeps categories and links. The same content serves Fin, Copilot and the public help center.
The practical task is therefore editorial, not technical: decide which articles are authoritative, remove contradictions and outdated policies, and write down the rules that exist only in agents' heads. Intercom states that Fin trains only on knowledge and Procedures you have approved. Kalshi began by importing its help center and running batch tests, before it shaped any rules.
Decide which customer data Fin may look up
An answer about a specific account needs that account's data, so this is a decision to make deliberately. Intercom lists connectors through which Fin reads Salesforce and HubSpot records, Shopify store data, Stripe billing and subscription details, Statuspage incidents and Jira or Linear issue status. For anything else it offers a REST API connector, webhooks and MCP.
Start from your ten most frequent question types and name, for each, the system that holds the answer and the narrowest read access that suffices. Kalshi's example, described after its move to Intercom, is a question like where is my money: Fin looks up the account, checks whether a market is open, pending or settled, separates deposit holds from withdrawal processing, and answers with dates and amounts. That only works because the account data is reachable.
Limit the actions and put approvals where they matter
Intercom describes Procedures as the way Fin handles multi-step workflows, applies business rules and triggers approvals inside boundaries you define. It also says that when Fin needs an approval or a missing piece of information, it asks a teammate and then finishes the conversation. The pricing page lists taking action on external systems among Fin's features.
Our recommendation, not a vendor claim: sort candidate actions by reversibility. Read-only lookups go first, reversible changes second, and anything involving money or account status stays behind a human approval until the transcripts justify loosening it. Write the list before you build the first Procedure.
Write the escalation rules and the excluded topics
Intercom says conversations are monitored against your criteria and that Fin escalates to a human when confidence thresholds are not met. It also offers Simulations, which run full conversations against edge cases, including vulnerable-customer and compliance scenarios, before launch. Fin can hand off to agents in the inbox you already use, according to the pricing page.
Three decisions are yours: the topics Fin never handles at first, who receives the handoff, and what context travels with it. Kalshi kept fraud investigations, compliance cases and account restrictions out of Fin entirely at launch, and set rules for permitted terms, topics to escalate on sight, and financial questions that must not drift into advice. Its legal and compliance teams stress-tested those guardrails before go-live.
What Kalshi's rollout shows, and where it stops
Kalshi, per its own published story, put Fin live on all production traffic roughly ten days after the first call, on top of its existing helpdesk. After that helpdesk's API rate limits blocked agents and Fin during Super Bowl traffic, Kalshi moved the whole helpdesk to Intercom in under three weeks. The team reports Fin resolving about 80% of live-chat conversations end to end, and median resolution time falling from about 23 hours to minutes.
Work changed as well: one person became a full-time Fin engineer, another focused on metrics and knowledge gaps, and five people stayed on the front line. Fraud and compliance flows were later added through an internal Kalshi system that Fin calls. These are vendor-published figures from one regulated company, so they show how one team proceeded, not a rate you should expect.
Budget on outcomes and define what you measure
Intercom's published price is $19 per seat per month plus $0.99 per outcome, with spending limits you can set. Fin alone, on an existing helpdesk such as Zendesk or Salesforce, starts from $0.99 per outcome with no seats. A free trial is available. Confirm the exact definition of a billable outcome before forecasting, because your bill follows it.
Intercom advertises a 76% average resolution rate across customers, and its customers quote figures from 50% to 80%. Those populations and definitions differ, so none is a benchmark for you. Measure your own resolution rate by topic, with the escalated share shown separately.
Method and limits
2026-11-01
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.
