There’s a person in your company right now getting two or three times more done than everyone around them, and they’re probably not talking about it.
Boris Cherny, the creator of Claude Code at Anthropic, says he hears the same thing from engineers at other companies every day: “one person is 10x’ing their output with Claude but the rest of the org hasn’t caught up.” In July he published Steps of AI Adoption, a map of the five stages he keeps watching teams move through: Gated → Assisted → Parallel → Supervised autonomy → AI-native.
His version is written for software teams, but we’ve walked enough small and mid-size businesses through this exact climb to tell you the ladder is the same whether the work is code, client intake, bookkeeping, or answering the phone. And the most useful thing about it isn’t the theory; it’s that it tells you precisely where you are and what’s blocking the next step.
So before the theory: find out where you stand. Seven questions, three minutes, scored against Cherny’s ladder.
Which step is your business on?
Seven questions, about three minutes. Answer for how things actually work today, not how the plan says they should.
The ladder behind your score
Here’s what each step actually looks like in a business, and the specific bottleneck between you and the next one. Everything below is grounded in Cherny’s framework, with the engineering examples translated into operations any company will recognize.
Step 0: Gated (“We’re still evaluating AI”)
AI is technically not allowed, or allowed in a way so process-heavy nobody bothers. Requests sit in an approval queue. Legal wants a policy first. Someone forwarded an article about hallucinations in March and that settled it.
Here’s the uncomfortable part: your team is already using AI. They’re just using it on personal accounts, on personal phones, with your company’s information. Wharton professor Ethan Mollick calls these people “secret cyborgs”: employees hiding their AI use because they’re worried about how it’ll look.
The bottleneck: leadership. Not technology, not budget. Cherny points at legacy approval processes and companies obsessing over cost-per-token instead of outcomes. In an interview with Fortune, he made the point bluntly: comparing AI spend to your software budget is the wrong frame; compare it to the cost of the work it replaces.
How to climb: someone with authority has to own this. Pick approved tools, write a one-page usage policy, and give people a sanctioned way to do what they’re already doing in secret.
Step 1: Assisted (you plus one AI)
This is where most businesses actually are. One person, one AI, working like a fast pair: drafting the proposal, summarizing the contract, writing the job posting. You review everything before it goes anywhere.
The unlock is real. Cherny describes it as “a change that used to fill an afternoon becomes something you finish between meetings.”
The bottleneck: your attention. You don’t fully trust the output yet, so you read every word. The work is synchronous: you sit and watch. You’ve made one person faster; you haven’t changed how the business runs.
And be honest about measuring it. A randomized study by METR found that experienced developers using AI tools felt 20% faster while actually completing tasks 19% slower. The gap between how productive AI feels and what it actually delivers is real, and the difference comes down to whether there’s a verification loop that doesn’t depend on you personally re-checking everything.
How to climb: stop reviewing keystrokes and start reviewing outcomes. Write down what “done and correct” means for each task: a checklist the AI’s work can be checked against, ideally automatically. Trust isn’t a feeling; it’s a track record against a standard.
Step 2: Parallel (one person, many streams)
Now one person runs five to ten AI workstreams at once; in Cherny’s world, an engineer orchestrating agents that check their own work (tests, builds, security scans) before a human ever looks.
The business translation: your invoice follow-up runs itself and flags exceptions. Your AI receptionist answers every call and books appointments, and you review the log, not the calls. Your workflow automation moves data between systems and reports what it did. You’ve gone from doing the work to supervising streams of it.
Cherny’s unlock line for this step is the one to write on the whiteboard: “A backlog that used to take the team weeks becomes one engineer’s afternoon of orchestration.”
The bottleneck: review. You’re producing more than you can inspect, so the question becomes which checkpoints actually need a human. The businesses that stall here are the ones that never decided what needs sign-off and what doesn’t.
How to climb: build self-checks into every automated process: confirmation numbers, exception reports, spot-audits. Then let the routine 90% flow and put your attention on the flagged 10%.
Step 3: Supervised autonomy (the question changes)
At this step, AI does most of the work in its lane, and the question flips. Cherny frames it perfectly: “Did you read the code?” becomes “What context was the model missing, and how do we solve it for next time?”
That question is the whole game for a business. When AI-produced work is wrong at this stage, it’s almost never because the model is dumb; it’s because it didn’t know the thing everyone in your company knows. That the client hates Tuesday calls. That the refund policy changed in March. That “the Henderson job” means the second Henderson job.
The bottleneck: context. Your company’s knowledge is scattered across your CRM, inboxes, call recordings, and three people’s heads, and an agent without that context is, as we like to put it, a very confident intern. This is exactly why we built Brain, our company brain product: one permission-aware knowledge base that both your people and your AI agents draw answers from, with citations.
How to climb: stop feeding AI context by hand, task by task. Centralize it once, keep it synced automatically, and every agent you deploy after that inherits it on day one.
Step 4: AI-native (steering by intent)
The loop closes. Work kicks off automatically: a missed call triggers the follow-up, a competitor’s price change triggers the analysis, month-end triggers the close. Humans steer by intent and monitor by exception. Cherny’s version: “the quarter-long migration becomes a workflow you kick off and check on.”
Andrej Karpathy sketched where this is heading in his talk Software Is Changing (Again): natural language is becoming the way we program computers, which means describing what you want done is the new operations manual. Anthropic, by Cherny’s own account, is at step 3 and pushing toward 4. Almost nobody is fully there yet. That should be encouraging: the frontier lab building these tools is one step ahead of where a well-run mid-size business can realistically get.
The bottleneck: knowing which work to automate and which guardrails each type needs. This is a judgment problem, not a technology problem.
The three rules of the ladder
1. You can’t skip steps. Cherny’s warning to engineering leaders applies to every business: the trap is “scaling agent count before the loop has earned widespread trust.” Jumping from step 1 to step 3 because a vendor demo looked good is how companies end up with expensive automations nobody trusts and everyone quietly works around.
2. The bottleneck is never the AI. Read the ladder again: the blockers are approval processes, attention, review capacity, trust, and context. All human. All fixable. As Cherny put it, at each step “tokens aren’t enough… you need to find and break down the next set of bottlenecks, and build up the next set of guardrails.”
3. One step is a business transformation. Most businesses we audit are at step 1 while believing they’re further along, and the single most common reason is that nobody ever wrote down what “done and correct” means for the work they want to hand off. Going from 1 to 2, from one assisted person to supervised parallel streams, routinely gives back 10+ hours a week per person. You don’t need step 4 to win. You need the next step.
Got your score? Here’s the honest next move.
Whatever step the quiz put you on, the pattern is the same: the blocker is specific, human, and fixable: a policy, a checklist, a review system, or a knowledge base. Figuring out which one, on your actual stack and team, is literally what our AI audit is for. We’ll confirm your step, name what’s blocking the next one, and tell you whether we’re the right people to build it. Sometimes the answer is a checklist, not a contract.
Sources & further watching
- Boris Cherny, Steps of AI Adoption (Anthropic, July 2026): the framework behind the quiz (X thread)
- Boris Cherny on measuring AI by outcomes, not token cost: Fortune interview
- Boris Cherny, Claude Code and the Evolution of Agentic Coding (YouTube): how the orchestration workflow actually looks
- Ethan Mollick, Making AI Work: Leadership, Lab, and Crowd (One Useful Thing): why adoption is an organizational problem
- Andrej Karpathy, Software Is Changing (Again) (YouTube, YC AI Startup School): the big-picture “Software 3.0” thesis
- METR, Measuring the Impact of Early-2025 AI on Developer Productivity: the perception-vs-reality gap