Where Is Your AI Adoption Stuck? Turning One Person's 10x Into Organizational Results¶

For / Key Points

For: Engineering leaders whose strongest AI users move faster while team delivery and quality remain unchanged.

Key Points:

“10x” describes an individual's output, not a guaranteed organizational result
As adoption advances, the bottleneck moves from implementation to review, judgment, and control
The next stage depends on the right guardrail, not merely on running more agents

One engineer produces a large volume of work with Claude. The review queue grows, and release frequency barely moves.

Is that successful AI adoption, or local optimization?

This article asks one question: how can an individual's “10x” become an outcome the organization can absorb?

Boris Cherny, who leads Claude Code, says he repeatedly hears that one person has “10x'd” their output with Claude while the rest of the organization has not caught up. He organized that pattern into Steps 0 through 4.¹ The 10x figure is not measured company-wide productivity. It is an observational framing, not a promised return on adoption.

Why a “four-step” model runs from Step 0 to Step 4¶

The model places organizations that cannot use AI effectively at Step 0, followed by four stages of progress. It contains five states but four transitions beyond the starting gate.

flowchart LR
    S0[Step 0<br/>Gated] --> S1[Step 1<br/>Assisted]
    S1 --> S2[Step 2<br/>Parallel]
    S2 --> S3[Step 3<br/>Supervised autonomy]
    S3 --> S4[Step 4<br/>AI-native]

The difference between stages is not simply model intelligence. It is where people wait, what they must inspect, and which failures the operating system can stop.

Stage	State of work	Main bottleneck	Next operating mechanism
Step 0: Gated	Approvals or connection paths keep AI outside the workflow	Security review, identity, budget decisions	Single sign-on, permissions, data boundaries, approved execution environments
Step 1: Assisted	One person supervises one AI closely	Continuous attention and manual review	Tests, builds, linting, plan review before edits
Step 2: Parallel	One person coordinates several AI workstreams	Review volume and steering congestion	Automated review, isolation, shared rules, usage analytics
Step 3: Supervised autonomy	AI starts recurring work with limited supervision	Trust, exceptions, and operating cost	Evaluation criteria, reusable rules, monitoring, stop conditions, cost caps
Step 4: AI-native	People set intent and supervise mainly by exception	Maintaining controls across many use cases	Use-case guardrails, outcome metrics, model routing

This is not an international maturity standard. It is a practical map from the perspective of people building and operating Claude. Agent counts and product names should not become adoption targets. The useful question is simpler: where will work wait next?

Individual speed moves work downstream¶

When AI accelerates implementation, work does not disappear. It moves into verification, approval, integration, and release.

The Claude Code team reports that once writing code, tests, and refactors became less limiting, verification, code review, and security became the new bottlenecks.² Anthropic's Code Review announcement also says code output per engineer increased 200% in one year and that review became constrained.³

That does not mean every company becomes 200% faster. The figure describes code output inside Anthropic, not another organization's lead time or business result. Its value here is structural: accelerating one stage exposes the next constraint.

Suppose a Step 1 team starts several agents at once. Implementation volume rises, but manual tests and reviews remain unchanged. Unverified changes accumulate. More prompt training will not fix that queue. The team needs agents to complete tests, type checks, and linting before humans judge acceptance and exceptions.

So where should the organization intervene first?

Diagnose the stage by waiting time, not user count¶

Do not infer maturity only from active users or token consumption. Find where work waits the longest.

Waiting for access or integration approval points to Step 0
Reading every AI edit in real time points to Step 1
A growing queue of PRs or deliverables points to Step 2
Continuous human monitoring of autonomous decisions and cost points to Step 3

Anthropic's organizational rollout guide recommends starting with 20–50 experienced users, validating workflows against the real codebase, and documenting what works and what fails before broader expansion.⁴ Its Champion Kit similarly recommends turning one person's discoveries into shared examples and recurring team habits so adoption does not depend on a single expert.⁵

This is the point where individual speed becomes an organizational asset. Distributing a power user's prompts is not enough. The team must preserve the successful procedure, required context, acceptance criteria, and return path after failure in a form that people and agents can reuse.

Step 4 should not be the goal for every workflow¶

Mature organizations do not automate everything. They advance work that is easy to verify and easy to reverse.

Dependency updates, routine tests, and first-pass log classification can support automated validation and rollback. Identity infrastructure, payments, personal data, and legal judgment have stronger reasons to retain human expertise. The Claude Code team likewise keeps human domain review for legal risk, security boundaries, and product judgment.²

Anthropic's agent design guidance recommends starting with the simplest workable approach and adding complexity only when it creates enough value.⁶ Under that principle, Step 4 is not a company badge. It is an operating choice made per workflow.

Summary: fix the next constraint, not the next stage¶

An individual's “10x” fails to become an organizational result when review, approval, integration, and operations still run at the previous pace.

The first move can stay small.

Choose one workflow with the longest waiting time
Identify whether the wait comes from people, permissions, quality, or cost
Add one guardrail required for the next operating pattern

The new implication is that an organization does not need one maturity score. Document summarization may operate at Step 3 while payment changes remain at Step 1. AI adoption matures when the organization can choose the right level of delegation for each workflow and explain why.

Boris Cherny, post introducing Steps of AI Adoption, July 17, 2026. The post frames the gap between individual output and organizational adoption as Steps 0 through 4. ↩
Claude by Anthropic, Running an AI-native engineering org, June 3, 2026. It describes how verification, review, and security became bottlenecks and where the Claude Code team retains human domain judgment. ↩↩
Claude by Anthropic, Bringing Code Review to Claude Code, March 9, 2026. It describes Anthropic's internal code-output and review metrics and its multi-agent Code Review system. ↩
Claude by Anthropic, How to scale agentic coding across your engineering organization, October 15, 2025. It recommends starting with experienced users and sharing validated workflow patterns before broader rollout. ↩
Claude Help Center, Claude Code champion kit, April 15, 2026. It explains how to turn individual success into reusable knowledge and shared adoption habits. ↩
Anthropic, Building effective agents, December 19, 2024. It recommends beginning with the simplest viable design and adding agent complexity only when necessary. ↩