Home
Apr 5, 2026
View All

Anthropic's Growth Playbook and What It Means for AI PMs

The Short Version#

A quiet Saturday — but Lenny's interview with Anthropic's Head of Growth on scaling from $1B to $19B ARR in 14 months is the kind of rare, first-person account of AI product growth strategy that's worth slowing down for. Meanwhile, Simon Willison is doing what he does: shipping security tooling fast and publicly, which is itself a useful pattern to watch.

Lenny Rachitsky / Amol Avasare — How Anthropic Grew from $1B to $19B ARR in 14 Months#

Source: https://www.lennysnewsletter.com/p/anthropics-1b-to-19b-growth-run Credibility: High (first-person account from Anthropic's Head of Growth, on a verified, high-quality editorial platform)

What happened: Lenny published an interview with Amol Avasare, Anthropic's Head of Growth, walking through how Claude went from $1B to $19B ARR in 14 months — which would make it one of the fastest-growing AI products ever. The excerpt mentions three specific levers: big bets (presumably on the right enterprise and API segments), intentional onboarding friction, and a tool called CASH — an internal AI system for autonomous growth work. That last detail is notable: Anthropic is apparently using AI agents internally to drive growth, not just selling them externally.

Key growth patterns:

  • Intentional onboarding friction — the opposite of the usual "reduce friction everywhere" playbook. Suggests Anthropic is optimizing for the right users, not just user volume
  • Big bets — concentrated resource investment in a few high-conviction channels or segments rather than spreading thin
  • CASH (internal AI system for autonomous growth) — an internal agent running some part of the growth operation. This is a direct signal that agentic workflows are crossing the threshold from experiment to core team infrastructure
  • Deep focus — mentioned as a cultural/operational value alongside the above, suggesting disciplined prioritization under rapid scaling conditions

Why it matters for PMs: The onboarding friction piece is the most PM-actionable signal here. The conventional wisdom is to remove friction, qualify users faster, get them to "aha" as quickly as possible. Anthropic deliberately added friction, which usually means they're solving for user quality or trust — not just activation rates. If you're building an AI product and your users need to understand what they're getting into before they can succeed with it, this is validation that friction can be a feature. The CASH detail is also worth sitting with: if Anthropic's growth team is already running autonomous agents for parts of growth work, the question isn't whether AI agents will change how product teams operate — it's how far along that curve your team is.

Critical questions:

  • What exactly is the friction they added, and at what point in onboarding? Friction at account creation vs. first-use vs. feature gating has very different implications.
  • Is the $1B → $19B growth primarily API (developer/enterprise) or Claude.ai consumer? The playbook looks very different depending on the answer.
  • What does CASH actually do autonomously — experiment setup, channel optimization, copy generation? The line between "AI-assisted" and "autonomous" matters a lot for how you staff growth.
  • How much of this growth is price-sensitive enterprise (where switching to a cheaper model is a real risk) vs. sticky use-case-specific adoption?

Action you could take today: If you have an onboarding funnel, pull the data on which activation steps correlate with long-term retention — not just conversion. Identify whether your friction points are filtering for users who will succeed, or just creating drop-off. That's the question Anthropic apparently answered before deliberately adding friction back.

Simon Willison — scan-for-secrets: Security Tooling Shipped Fast and in Public#

Source: https://simonwillison.net/2026/Apr/5/scan-for-secrets/#atom-everything Credibility: Medium-High (first-party post from a recognized practitioner, versioned releases suggest real tool)

What happened: Simon shipped three versions of a tool called scan-for-secrets on a single day (0.1, 0.1.1, and 0.2), plus posted a research-llm-apis update. The exact details of what scan-for-secrets does aren't fully available from the excerpt, but the name and the rapid version cadence tell a clear story: he built something to detect secrets (API keys, tokens, credentials) in code or content, likely with an LLM component given his recent work on AI-assisted tooling, and iterated on it rapidly in public. He also posted updates to research-llm-apis, a separate tool for researching LLM APIs.

Key patterns:

  • Rapid public iteration — three versions in one day signals that Simon treats his public repositories as working-in-progress tools, not polished releases
  • Security + LLM tooling overlap — scan-for-secrets in the context of LLM-heavy workflows is directly relevant to the risks of agentic systems that touch credentials, environment variables, and API keys
  • Tooling to research LLM APIs — the research-llm-apis project suggests ongoing tooling for comparing/evaluating LLM APIs, which is useful for PMs evaluating build vs. buy

Why it matters for PMs: If your team is building agentic workflows, the surface area for secret exposure has expanded dramatically. Agents that read files, execute code, or browse repos can accidentally expose credentials in their context windows, logs, or outputs. A tool specifically designed to scan for secrets in LLM-adjacent workflows is the kind of infrastructure that mature AI product teams will need. This is also a signal about what PM-adjacent builders are caring about right now: not just "will the agent work" but "will the agent accidentally expose something it shouldn't."

Critical questions:

  • Does scan-for-secrets operate on LLM inputs/outputs specifically, or is it a general-purpose secrets scanner? The distinction matters for whether it's relevant to AI workflow security or just standard DevSecOps.
  • How does it handle false positives? Secrets scanning tools are notoriously noisy — if every run flags 200 things, teams ignore them.
  • Is there an MCP integration or similar that would let AI coding tools use this as a guardrail in real-time?

Action you could take today: Ask your engineering team if there's a secrets scanning step in your CI pipeline that covers AI-generated code and LLM prompt/response logs — not just committed code. If not, that's a gap worth flagging.

Quick Hits#

The Thread#

Autonomous AI internal tooling is becoming the real signal. This week: Anthropic revealed an internal AI system (CASH) running parts of their growth operation autonomously. LangChain's Harrison Chase keeps pushing on agent observability via LangSmith. Simon Willison is shipping security tooling designed for the risks that agentic workflows introduce. The pattern across all of it: the teams closest to AI product development are already operating with AI agents running core workflows — growth, deployment, code review — and the infrastructure questions (observability, security, access control) are racing to catch up. The gap between "we use AI to help us write" and "AI is running part of our operation" is closing faster than most product teams are accounting for.

Sit With This#

Anthropic's Head of Growth says they deliberately added onboarding friction as part of a strategy that took them from $1B to $19B ARR in 14 months. Most PM playbooks treat friction as the enemy of activation.

For your AI product: Where in your onboarding does friction currently exist — and is it filtering for users who will succeed long-term, or just creating drop-off? How would you tell the difference with the data you have today?