What does it really take to build AI that can resolve customer support at scale reliably, safely, and with measurable business impact?
We explore how Intercom has evolved from a traditional customer support platform into an AI-first company, with its AI assistant, Fin, now resolving 65% of customer queries without human intervention. Intercom’s Chief AI Officer, Fergal Reid, discusses the company’s journey from natural language understanding (NLU) systems to their current retrieval augmented generation (RAG) approach, explaining how they’ve optimised every component of their AI pipeline with custom-built models.
The conversation covers Intercom’s unique approach to AI product development, emphasising standardisation and continuous improvement rather than customisation for individual clients. Fergal explains their outcome-based pricing model, where clients pay for successful resolutions rather than conversations, and how this aligns incentives across the business.
We also discuss Intercom’s approach to agentic AI, which enables their systems to perform complex, multi-step tasks, such as processing refunds, by integrating with various APIs. Fergal shares insights on testing methodologies, the balance between customisation and standardisation, and the challenges of building AI products in a rapidly evolving technological landscape.
Finally, Fergal shares what excites and honestly freaks him out a bit about where AI is heading next.
Timestamps
00:00 – Intro
02:31 – Welcome to Fergal Reid
05:26 – How to train an NLU solution effectively?
08:56 – What gen AI changed for Intercom
10:57 – How would you describe Fin?
14:30 – Fin’s performance increase
17:18 – Intercom’s custom models
22:14 – Large Language Models vs Small Language Models
30:40 – RAG and the ‘full stop problem’
40:08 – Agentic AI capabilities at Intercom
50:40 – Intercom’s approach to testing
1:04:46 – The most exciting things in the AI space
Show notes
Learn more about Intercom
Connect with Fergal Reid on LinkedIn
Follow Kane Simms on LinkedIn
Article – The full stop problem: RAG’s biggest limitation
Take our updated AI Maturity Assessment
Subscribe to VUX World
Subscribe to The AI Ultimatum Substack
