Claim
Run growth experimentation through a four-stage substrate, identify opportunity, build the change, test against a quality + brand bar, ship and analyze, driven by Claude. The fifth stage (cross-functional alignment) stays human, and that is the lasting bottleneck.
Mechanism
Most growth experimentation is loosely coupled steps that already have rich playbooks: ideation, implementation, QA, analysis. A capable model can drive each step end-to-end against a written brand and quality bar, with current win rates around "junior PM 2–3 years in." The expensive human input is no longer building the experiment, it is the political and aesthetic work of getting six people in a room to agree on what to ship.
Conditions
Holds when:
- The team has a written quality + brand bar codified as skills with explicit dos / don'ts.
- A frontier model (Opus 4.5+ in Anthropic's case) is wired to the relevant tools.
- Growth output is a high-volume stream of similar experiments, not one-shot strategic bets.
Fails when:
- Brand and quality guardrails are tacit, not written. The model has nothing to align to.
- The work depends on novel research or category creation, not iteration on a known surface.
- Stakeholder alignment was already the bottleneck, automation does not solve org-design problems.
Evidence
"Identify opportunities → build the feature → test against quality + brand bar → ship + analyze."
"We will have AGI and it will still be impossible to get six people in a room to align."
The team is led by Alexey Komissarouk inside Anthropic. Win rate today is named at "junior PM 2–3 years in." The substrate wasn't viable before Opus 4.5; it is now. Human-in-loop review need is decreasing weekly.
· Amol Avasare on Lenny's Podcast, 2026-04-05
Signals
- Number of shipped experiments per growth-PM-month rises 3–5x without quality regression.
- Brand-bar violations caught in pre-ship review trend down to a stable low.
- PM time shifts from "building the experiment" to "deciding which experiment matters" and "negotiating cross-team alignment."
Counter-evidence
Operators outside frontier labs may not have the brand-bar maturity, the model access, or the org buy-in to run this. Without the codified guardrails, automating stages 1–4 produces fast slop. The win is not in the automation; it is in the prerequisite of having explicit quality definitions.
Cross-references
- Claude Code multiplies engineers 2–3x; PM and design become the bottleneck, the staffing implication
- Give the model tools and a goal; do not hard-code the workflow, why the substrate exposes Claude with light scaffolding