Intercom bet 90% of the company on a product earning 1% of revenue, renamed itself Fin, and sold to Salesforce for $3.6B. Now 94% of its code is written by Claude

AI will make your product irrelevant (unless it was born after 2023), your real competitors are startups founded today rather than the incumbents you benchmark against, and slow adapters die. Some product categories stop needing to exist at all; his (deliberately fictitious) example is status-update and project management tools, which lose their point once AI connects Slack, GitHub, and the rest directly.

When the UI becomes "tell us what you want and we'll do it," queries stop respecting product boundaries ("show me which of my best performing sales reps take the most holidays" spans the HR tool and the CRM). Everyone building an agent is suddenly crossing seven or eight product lines. His scoping advice: stop where you hit a different buyer, a different level of data permission, a brand you'd need to rebuild, or a really popular product you don't want to fight.

The old world split people who know what they want from the few who know which menu does it (his example: conditional formatting in Excel versus just asking for negative cells in red). Their product Operator works this way: upload a PDF of a product launch and say "update all the documentation," and it works out which docs change where. He's blunt about the implication for his own former craft, design: it's not obvious what role classic design systems have left.

A single agent step at 99% reliability sounds fine; multiply a realistic workflow of such steps and the agent lands near 85%, which nobody buys. Getting from the 80% an enthusiastic weekend with Claude gives you to the 99.99% a sellable product needs is the hard part, and he sees startups announce the homepage before discovering this. Big companies too: he points to a product announced in 2024 that still hasn't shipped.

Fin resolved about 25% of support conversations at launch; now it's around 75%. Nothing in the UI changed, which leaves marketing with nothing to screenshot: the product simply does what it always claimed, more often. Intercom's famous product principles had to be abandoned wholesale; deterministic build-ship-measure became a messy loop of "what is possible, how reliable can we make it, can we productize it, how do we improve it after we see real-world performance."

Empirical evaluation, scientific rigor, causality, prompting, and post-training their own models (plus about 15 small language models running alongside them). Design flipped from iterating on one expensive direction to exploring many directions in parallel, then refining the winner.

A hard rule since last August: everyone learns to deploy. Designers no longer file UI tickets, they fix them; they vibe code the roadmap as interactive prototypes shared with customers; they own the front end entirely. In his words, the decade-long "should designers code" debate is over, "because coders don't even code. Claude codes."

Their CTO targeted doubled productivity before Claude Code's breakout; they tripled it, measured as features shipped to production and used by real people. They also measure individual engineers' AI adoption against output, which he admits nobody likes.