Head of Growth (Anthropic): “Claude is growing itself at this point” | Amol Avasare

What It Covers

Amol Avasare, Head of Growth at Anthropic, details how the company scaled from $1B to $19B ARR in 14 months. He covers growth team structure, activation strategy, the CACHE automation initiative using Claude to run growth experiments, how PM and engineering roles are shifting, and why Anthropic deliberately leaves money on the table to protect brand and safety.

Key Questions Answered

• •Cold Email Acquisition: Avasare landed his role at Anthropic by cold emailing CPO Mike Krieger directly — not through any formal hiring channel. His framework: find personal emails rather than LinkedIn or work addresses, craft a subject line optimized for open rate, keep the body short with three elements (who you are, why you fit, call to action), and follow up repeatedly until explicitly told to stop. Being a growth practitioner who excels at cold email functions as a live demonstration of the skill itself.
• •Activation Friction Framework: Adding friction to onboarding consistently outperforms removing it, provided the friction helps users understand why the product is relevant to them. Anthropic, MasterClass, Mercury, and Calm all use multi-step onboarding quizzes that appear counterintuitive but drive higher completion and downstream retention. The rule: remove friction that adds no value, but keep friction that personalizes the experience. Data collected during onboarding also enables lookalike targeting and lifecycle marketing, compounding its value well beyond the signup moment.
• •Growth Bet Sizing by AI Dependency: At AI-first companies, shift the experiment portfolio from 70% small bets and 30% large bets to roughly 50/50 or even 70% large bets. The rationale: if product value will be 100x to 1,000x higher in two years due to model improvements, small optimizations capture a negligible share of future upside. This logic applies specifically when AI is the core value driver — not when AI features are peripheral additions to a non-AI product.
• •CACHE Growth Automation: Anthropic's growth platform team runs an initiative called CACHE (Claude Accelerates Sustainable Hypergrowth) that uses Claude to automate the four-stage experiment loop: identifying opportunities, building features, testing against quality and brand standards, and analyzing results. Launched only months ago and only viable after Opus 4.5, it currently performs at a junior PM level for copy and minor UI changes. The one stage not yet automatable is cross-functional stakeholder alignment, which still requires human judgment.
• •Engineer-as-PM Delegation Rule: When engineering capacity outpaces PM bandwidth — which is increasingly common as AI multiplies engineer output — use a two-week threshold. Projects under two engineering weeks are owned end-to-end by the engineer, including legal, security, and stakeholder coordination. Projects over two weeks revert to PM ownership. This requires hiring product-minded engineers who can handle coordination, and it dramatically increases the value of engineers with strong product instincts relative to purely technical peers.
• •Cowork Scheduling for Metrics and Alignment: Avasare runs scheduled Cowork tasks that automatically review 20 to 25 metric charts each morning and surface anomalies in Slack before he starts work. A separate weekly task scans Slack via MCP to identify potential misalignment across teams and projects. A third task reviews direct report activity against team OKRs and generates suggested feedback. Setup requires the Cowork desktop app, Slack MCP connection, and admin permissions — all currently available without custom engineering.
• •Leaving Money on the Table as Growth Strategy: Anthropic's growth team explicitly foregoes metric gains when tests conflict with brand standards, user experience quality, or safety principles. Avasare frames this as a structural advantage: the best long-term growth compounds from trust, not from squeezing short-term conversion. Practically, controversial tests are sorted into two categories — those that cross non-negotiable lines and are never run, and those that are merely uncomfortable but testable if the hypothesis is strong and the expected return justifies the brand cost.