Fin, the customer service software company formerly known as Intercom, has introduced a commercial AI system whose sole job is to supervise, diagnose and improve another AI: the company’s own customer-facing support agent. The product, called Operator, was announced in May 2026 alongside the company’s rebrand and the release of Fin 4.
Operator does not speak to customers. According to the company, it monitors the conversations that the Fin customer agent has with users, identifies where the agent gave wrong or outdated answers, got stuck in loops or saw its performance slip, and then works out why.
What Operator does
The company says Operator analyzes conversations automatically, traces the root cause of failures, drafts new instructions for the customer agent, updates knowledge base articles, produces reports and recommends which changes would reduce errors in the future. Work that previously took support operations teams hours or sometimes days, Fin says, can now be completed in roughly ten minutes.
Fin says its customer agent already resolves more than two million customer issues a week for clients including Anthropic, DoorDash and Mercury. As those deployments have grown, so has the operational burden of keeping the agent accurate, which is the gap Operator is meant to fill.
Humans still approve changes
The system is deliberately not fully autonomous. According to the company, no change Operator proposes takes effect until a person approves it. Fin frames that as a safeguard: companies deploying AI in customer-facing roles have learned that unsupervised changes can introduce new errors as easily as they fix old ones.
VentureBeat reported that Operator runs on Anthropic’s Claude models rather than on models built by Fin itself, with the company saying Claude was better suited to the analysis, reasoning and debugging tasks involved. Fin had earlier said Operator went through an early access period in which it generated roughly 20,000 improvements before general release.
A new layer in the stack
The launch illustrates a pattern taking shape across the AI industry. People once operated software directly; then they operated AI systems; now vendors are selling AI systems whose purpose is to operate other AI systems, with humans moving into a supervisory role. Whether that structure reduces errors or merely moves them up a level is a question support teams will be able to test as Operator rolls out.
