Des Traynor - Real Talk about AI and Software - [Invest Like the Best, EP.340]

Episode Summary

Executive Summary: Patrick O'Shaughnessy interviews Intercom co-founder Des Treanor on how generative AI reshaped customer support, why incumbents can still win, and how product strategy must shift from copilots to workflow removal. He also shares investor lessons on durability, speed, and what makes software defensible.

Main Topics: AI transformed customer support at Intercom (Priority: 5/5): ChatGPT/GPT-4 made support automation practical, forcing Intercom to rapidly build and launch Fin. Incumbents vs. startups in the AI wave (Priority: 5/5): Treanor argues incumbents often win when AI augments existing workflows rather than replacing them. Building context-specific agents (Priority: 5/5): The hard part is grounding models in customer data, staying on-topic, and minimizing hallucinations. Adoption frictions and human-in-the-loop (Priority: 4/5): Customers adopt cautiously, using limited rollouts before trusting AI with broader support volume. Provider choice, pricing, and model economics (Priority: 4/5): OpenAI leads today, but availability and cost may shift model selection and product scope. Investor philosophy and software defensibility (Priority: 5/5): Treanor emphasizes execution, perishability of software, and the need to remove work, not just add AI features. Product, brand, and speed as moats (Priority: 4/5): Fast iteration plus brand trust and ecosystem integration matter more than first-mover status alone.

Key Arguments: AI changes support because conversational models can answer, summarize, and extract like agents do. Fast incumbents can beat startups when 80% of the old stack still remains and only 20% changes. Workflow replacement matters more than copilots; killing steps creates real value. Data quality and workflow design are more defensible than prompts or wrappers alone. Customers adopt AI cautiously, first on weekends or narrow issue types, before broad rollout. Software is perishable and easily copied, so speed and reinvestment are required to stay ahead. A strong brand converts product momentum into a durable position beyond pure feature parity.

Data Points: Intercom businesses using products: 25,000 businesses - Intercom products operate across a large customer base including Amazon, Lyft, and Atlassian. Fin beta release: December - First beta release after ChatGPT prompted the team to shift priorities. Public announcement of beta: January - Intercom discussed the AI product publicly after the initial beta release. GPT-4 beta launch: March - After getting access to GPT-4, Intercom constrained behavior and reduced hallucinations more effectively. General availability launch: late May or early June - Fin moved from beta to GA after rapid iteration. Support volume drop: 50% - Some customers saw support volume reduced by half after turning Fin on. Support volume reduction on default use: 15, 20, 25% - Many customers saw this level of support volume disappear simply by enabling Fin. Resolution price: 99 cent resolution - Intercom charges per AI resolution and compares it to human support costs. Intercom conversations per month: 20,000 support conversations a month - Des cites Intercom’s own support volume as training data for future snippets. Historical conversation volume: a quarter million a year - Intercom's annual support conversation count was used to illustrate knowledge accumulation. Two-year conversation lookback: half a million conversations - Used to show the scale of data available to train and improve Fin. Customer support team size note: thousands of intercom seats - Used to explain why even small UI changes can require retraining many support agents. Time to start AI work after ChatGPT: 14 hours - Intercom moved very quickly once ChatGPT launched. Support rep hourly floor: $8, $9, $10 an hour - Treanor contrasted AI pricing with typical human support labor costs. Future vision threshold: 80% - Used in discussion of what percent of a workflow might be automated or displaced. Timeframe for software becoming dated: 36 months - Treanor argued software best-in-class today can be outdated in three years.

Pivotal Quotes: "“the world of customer support was going to change”" — Des Treanor: Explaining why Intercom treated ChatGPT as an existential product shift. "“I want you to stop thinking about copilots. I want you to think about what work can we remove entirely?”" — Des Treanor: His central product thesis for AI-native software. "“software has a really it's very perishable”" — Des Treanor: Part of his argument that software businesses need constant reinvestment to stay competitive.

Implications: The unresolved question is which workflows can be fully deleted next; teams should prioritize end-to-end automation, not cosmetic AI add-ons.