{"id":10879,"date":"2025-08-14T14:39:22","date_gmt":"2025-08-14T06:39:22","guid":{"rendered":"https:\/\/www.dkmeco.com\/en\/?p=10879"},"modified":"2025-12-11T22:00:05","modified_gmt":"2025-12-11T14:00:05","slug":"how-did-intercom-use-100-million-and-gpt-4-to-build-the-hit-ai-customer-service-fin-in-4-months","status":"publish","type":"post","link":"https:\/\/www.dkmeco.com\/en\/how-did-intercom-use-100-million-and-gpt-4-to-build-the-hit-ai-customer-service-fin-in-4-months\/","title":{"rendered":"How did Intercom use $100 million and GPT-4 to build the hit AI customer service Fin in 4 months?"},"content":{"rendered":"

In 2022, when GPT-4 had just been released and most companies were still busy discussing the news, Intercom had already quietly moved into action. Within just a few hours, this customer service software company began hands-on testing; in only four months, it launched the AI Agent Fin\u2014now capable of handling millions of complex customer inquiries every month.<\/p>\n

This first-mover advantage was no accident. Faced with the rapid evolution of large language models (LLMs), Intercom\u2019s leadership made a decisive bet on AI. They quickly assembled a cross-functional team, shut down all non-AI projects, invested $100 million to rebuild the business architecture, and fully migrated to an AI platform.<\/p>\n

This decision triggered a company-wide transformation from top to bottom: reshaping product teams, establishing an \u201cAI-first\u201d customer service strategy, and building a technical platform capable of powering Fin\u2019s high-speed operations.<\/p>\n

Next came the three key lessons they summarized from this AI transformation journey\u2014lessons that any team, no matter their current stage, can immediately apply.<\/p>\n

\u201cAI must be embedded into product design from the start, not crammed in as an afterthought.\u201d \u2014Paul Adams, Chief Product Officer, Intercom<\/p>\n

Lesson 1: Start early and experiment continuously to improve model fluency<\/strong><\/p>\n

Intercom began experimenting with generative models early and often, gaining valuable real-world experience\u2014identifying the limitations of models and finding opportunities for optimization. When GPT-4 launched in early 2023, they were fully prepared, releasing the AI customer service agent Fin in just four months and rapidly expanding its use.<\/p>\n

\u201cWith GPT-3.5, we achieved smooth conversational experiences, even some \u2018magic,\u2019 but its reliability wasn\u2019t high enough for customer service. Because we had laid the groundwork early, when GPT-4 arrived, we knew the time was right and moved quickly to launch Fin.\u201d \u2014Jordan Neill, VP of Engineering, Intercom<\/p>\n

This grasp of model fluency enabled Intercom to design Fin Tasks\u2014a system that can automatically handle complex processes like refunds and technical support. While the team initially planned to use a retrieval-based architecture, evaluation showed that GPT-4.1 could complete tasks independently and efficiently, with higher reliability and lower latency.<\/p>\n

Today, GPT-4.1 remains the core engine of Intercom\u2019s AI systems, including the key logic of Fin Tasks. The team also found that adding \u201cchain-of-thought prompting\u201d to non-reasoning queries improved performance without building a full RAG pipeline.<\/p>\n

The conclusion is clear: the earlier and deeper you understand the model, the faster you can seize opportunities as the technology evolves.<\/strong><\/p>\n

Evaluations showed GPT-4.1 delivered the highest reliability in task execution while reducing costs by 20% compared to GPT-4o.<\/p>\n

<\/p>\n

Lesson 2: Use rigorous evaluation to drive rapid iteration and complete upgrades in days<\/strong><\/p>\n

To speed up technical upgrades, you must measure precisely what works and understand why.<\/p>\n

Intercom\u2019s ability to quickly switch to new models, modes, and architectures hinges on a structured and rigorous evaluation process. Whether it\u2019s Fin Voice (based on the Realtime API) or Fin Tasks (based on GPT-4.1), every deployment undergoes offline testing and live A\/B experiments, focusing on three key capabilities:<\/p>\n