Boris Cherny’s Steps of AI Adoption: A Roadmap

Every executive I talk to asks some version of the same question: how far along are we with AI, really? Boris Cherny created Claude Code, runs the product at Anthropic, and spends his days directing thousands of agents, which makes him one of the few people who answers that question from experience instead of theory.

He just published “ Steps of AI Adoption,” a one-page table that locates your organization on a path from employees locked out of capable models to executives directing a thousand agents at once.

Boris’s table tells you where you are. I’ve added some advice about how to overcome the challenges of getting from step to step.

Stage 0: Gated. (Number of Agents: 0)

Status: Employees at Stage 0 have little access to current models, and approved tools are often too limited for serious work. SecOps controls access and data policy with a dotted line to CDO oversight. Legal or risk & compliance interprets liability and confidentiality. Procurement manages vendors. The CFO governs spending and the CTO owns the operating environment.

The Bottleneck: Each function can delay deployment, while no single function can approve the whole system. Employees respond by waiting, using consumer tools without permission, or performing sanctioned demonstrations that never touch real work. The company calls this caution. Employees experience it as paralysis.

The Roadmap: Name one executive sponsor with authority to reconcile security, legal, procurement, finance, and technology requirements. Give that sponsor 30 days to publish an approved-tool list, data-classification rules, financial limits, prohibited uses, incident procedures, and a named owner for every policy decision. Then select a small group of employees and give them access to current models inside those boundaries.

The first milestone is operational evidence: approved employees using approved tools on real assignments with real company information. Count completed assignments, policy exceptions, incidents, time saved, and work rejected during review. Training begins after access because employees cannot learn to use tools they are forbidden to touch.

Stage 1: Assisted. (Number of Agents: ~1)

Status: You + an agent (a pair). At Stage 1, an employee assigns work to one agent, supplies context, corrects mistakes, and approves every material decision. These sessions teach employees where the technology performs well, which instructions improve results, what information the agent requires, and which errors recur.

The Bottleneck: Human attention sets the limit. Employees spend less time producing first drafts and more time checking confident answers whose sources, assumptions, and actions are difficult to inspect. A polished output without evidence can increase supervisory work.

The Roadmap: Choose repeatable assignments with clear inputs and observable outcomes. Require every assignment to produce evidence with the answer: source links for research, reconciliations for financial work, citations and defined-term checks for legal work, test results for software, and rights and brand checks for marketing.

Create a shared library of successful task briefs, failure examples, verification procedures, and approved information sources. Measure the employee’s total time from assignment through approval, including corrections. A fast first draft followed by an hour of forensic review has weak economics.

The team is ready for Stage 2 when employees can predict recurring errors, verify routine outputs quickly, and describe completion in terms an agent can test.

Stage 2: Parallel. (Number of Agents: ~10)

Status: Orchestrator. At Stage 2, one person runs five or 10 agents in parallel, with each agent receiving a bounded assignment and an isolated workspace. A 2026 longitudinal study of professional software engineers found that 82 percent of participants reported spending less time writing code as they devoted more attention to directing, evaluating, and correcting AI output, a category the researchers called “supervisory engineering work.”

The Bottleneck: Ten agents can finish assignments faster than one person can understand and approve their combined results. Weak task boundaries produce duplicate work, incompatible recommendations, conflicting changes, and a review queue that consumes the productivity gain.

The Roadmap: Break work into assignments with one objective, one owner, the minimum required context, a defined output, and an explicit quality test. Isolate workspaces so agents cannot overwrite one another. Route routine failures back to the agent automatically, and reserve human review for exceptions, consequential decisions, and outputs that pass their mechanical checks.

Limit concurrent work according to review capacity. Five agents generating 50 deliverables for one overloaded employee helps no one. Track cycle time, rejection rate, reviewer minutes, duplicate effort, and the age of the review queue. These measures show whether parallel production improves the operation or just relocates the bottlenecks.

The team is ready for Stage 3 when routine work can start, check itself, report its status, and escalate uncertainty without an employee managing every step.

Stage 3: Supervised Autonomy. (Number of Agents: ~100)

Status: Manager of managers (an org tree). Stage 3 introduces supervised autonomy. Agents delegate limited assignments to other agents, while schedules and defined events initiate recurring work. People supervise outcomes, costs, priorities, and exceptions.

The Bottleneck: No person can inspect every action taken by 100 agents and their subagents. Informal instructions, shared credentials, invisible costs, and unclear escalation rules turn small errors into big problems.

The Roadmap: Convert policies into controls the system can enforce. Give each workflow a named owner, approved information sources, permitted actions, spending limits, quality tests, escalation thresholds, an audit record, and a tested stop mechanism. Version the instructions so every output can be traced to the rules that governed it.

Run failure drills before increasing volume or velocity. See what happens when you revoke a permission, introduce bad source data, force a quality test to fail, exceed a budget, or simulate an unavailable vendor. The workflow should stop safely, preserve evidence, notify the correct person, and resume only after an accountable human approves the recovery.

The organization is ready for Stage 4 when executives can compare automated workflows using consistent measures of cost, quality, risk, and business value.

Stage 4: AI-native. (Number of Agents: ~1,000+)

Status: VP steering by intent. At Stage 4, hundreds or thousands of agents perform bounded work across the organization. Executives set objectives, constraints, budgets, and measures of success while people supervise the automation portfolio and intervene when results cross defined boundaries.

The Bottleneck: Technical success can hide poor allocation. An agentic system may automate low-value activity, accelerate a process that should have been eliminated, or consume more computing and supervisory resources than its output is worth. Teams will always defend the systems they built. Busy work may look like work, but it isn’t work.

The Roadmap: Manage automated workflows as a portfolio. Give every system an accountable executive, a business objective, a full cost model, a quality measure, a risk classification, and a review date. Expand systems that produce measurable value, redesign those with correctable constraints, and terminate those whose economics fail.

Make models and vendors replaceable. The company should own the objective, process, evidence, interfaces, and accountability structure. Test important workflows against alternative models and preserve the ability to move when capabilities, prices, contract terms, or risk profiles change.

Stage 4 has no finish line. The operating discipline is continuous because models improve, prices fall, regulations change, and the company’s own processes evolve. In our AI advisory practice, we call this creating a “culture of continuous improvement.” For most of our clients, this massive organizational shift is “the” leadership challenge of our time.

Cultural Debt Determines The Pace

The gap between technological capability and organizational capacity is cultural debt.

In the past, resistance to change was easy to understand and, while hard to fix, it was generally easy to see.

This is different. This is the first time humanity has shared the planet with an alien intelligence. And no one, myself included, has any idea what this will ultimately mean. We’ve put our workforce in an untenable position. Everyone has to applaud the advent of AI, put it on their résumés, learn it, act like they own it, etc. all while dealing with the gnawing feeling that they are training their replacement or maybe even training their future overlords.

When you combine cultural debt with the silent fear of AI, the size and scope of the problem become obvious. Most of my clients have people and teams in every stage of Boris’s table. Hopefully, these defined stages will help you name your various constraints, assign owners, define the evidence required to remove bottlenecks, and measure the results.

I’m grateful to Boris for sharing his “Steps of AI Adoption.” It is a welcome validation of our approach to deploying AI at enterprise scale.

Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.