Block's journey of transforming a 3,500-person engineering organization into an AI autonomous organization demonstrates an incredibly unique approach. Instead of training all employees, we selected 50 key employees, or 'AI champions,' and focused on innovating the 'code repository' rather than changing people. In this presentation from Angie Jones, who designed this transition from start to finish, we'll explore the actual sequence in which companies adopt AI and a six-stage maturity model to help diagnose where we stand in the AI ​​era. Ultimately, it emphasizes the disruptive changes that AI transition can bring and the importance of preparing for them.

1. Block's shocking AI transformation headline

In February 2026, CNN Business posted a shocking headline that Block had laid off nearly half of its employees because of AI, and CEO Jack Dorsey said most companies would follow the same path. A few months before this news came out, Angie Jones, who personally designed the AI ​​transformation for the 3,500-person Block engineering organization for two years, gave a detailed presentation on the entire process. The last minute of this presentation ends with this headline:

From this video we can take away two important things:

• The actual sequence of your company's AI transformation: Get a glimpse into the actual stages of your company's AI transformation in the future.
• 6-Stage AI Maturity Diagnostic Chart: A useful tool to determine where you and your team are in the AI ​​transformation process.

Block has been one of the fastest moving companies in the AI ​​transition. Even before AI supported the 'tool calling' function, which directly uses tools, we were creating our own coding agent, Goose, and when Anthropic's Claude model appeared, we participated in the initial design as a design partner.

2. 90% usage rate and CEO's skepticism: Introduction ≠ Impact

Angie Jones (VP of Agentic AI Foundation), who led Block's AI transition, oversaw the introduction of AI for 12,000 employees across the company, including marketing, legal, and finance, in the first half of 2025, and later received a mission from the CTO to change the entire engineering organization to agent-based. But as soon as we started, we faced a strange situation.

Within months of introduction, 90% of engineers were using Goose and Claude code on a daily basis. On paper it looked like a complete success, but the CEO was convinced just the opposite.

"It is clear that our engineers are not using AI."

As Engie Jones explains, the real problem was that, surprisingly, they were both correct.

"The indicator that 90% of people are using it was correct, and the CEO's sense that the speed of features being delivered to customers was also correct."

This shows that adoption is not leading to impact. Block's engineers were using AI, but mostly at the level of asking questions or creating boilerplate code within the IDE. In other words, it did not reach the impact stage that leads to performance.

3. Six-level AI maturity model and 1/9/90 rule

Based on this recognition of the problem, Block created a 6-stage AI maturity model to define the goal. The standard of this model measures the 'relationship between engineers and agents' rather than AI skills.

• Level 0: No use of AI at all.
• Step 1: Use autocomplete only.
• Step 2: Chat, but does not lead to actual code reflection (PR).
• Step 3: Entrust the entire work and inspect the results.
• Step 4: Operating multiple agents simultaneously.
• Step 5: Distributable results are produced without direct human involvement.

Most teams will remain somewhere between Stage 1 and Stage 2. The majority of Block's 3,500 engineers were also between Level 1 and Level 2, and the goal was all Level 5.

But Block's approach was unexpected. They didn't try to train all 3,500 people. The situation was very bad at that time. There was no standard method (playbook) to refer to, and the organization was already suffering from AI fatigue due to rapidly changing technology. Moreover, employees were already tired of the top-down pressure of "AI or die."

At this time, what Angie Jones suggested is the 1/9/90 rule. Just as in an online community, 1% of people create content, 9% comment and react, and the remaining 90% only watch, engineers' AI adaptation also follows this distribution.

"Some people will completely master using AI and discover useful patterns and techniques for working with agents. Some people will make small modifications to the agent.md file, but most people won't put in the extra effort to figure it all out."

In conclusion, I realized that a strategy that relies on individual leveling up is bound to fail.

"If AI strategies rely on individuals leveling up, we'll never see the impact."

4. 50 AI Champions and Storage Innovation

So Block chose 50 'AI Champions' instead of 3,500. Rather than receiving volunteers, we selected them by meeting with technical leaders and managers for a week and nominating them. There were three conditions.

• Be able to devote more than 30% of work time to AI champion activities.
• When the AI ​​doesn't go your way (which it says it often does), you won't give up.
• Represent **the most important **repositories (repositories)** in the company (from Square, Cash App, Afterpay to front-end, back-end, and mobile infrastructure).

What was the first thing these 50 champions did? It was not an on-campus lecture or distribution of prompt guides. We fixed the code repository. Instead of replacing people, we built an environment in the repository where agents can do their jobs well.

Here's why this is key: The mid-2025 models had one function well written, but the results could not be trusted because they did not know the team's conventions, code style, and what not to do. I couldn't trust him, so I couldn't entrust him with the job. In other words, it wasn't model performance that was blocking it, but trust.

So a standard called AI Friendly Storage was created. It includes four elements:

1. Context file: Serves as a guide to the repository, such as agent.md or claude.md.
2. Rules file: Acts as a guardrail that the agent must not cross.
3. Slash Commands and Skills: Contains repetitive tasks and allows them to be executed quickly.
4. AI Code Reviewer: Turn on the AI ​​code reviewer and even leave a mark in the PR (Pull Request) that it was written by AI.

The reason these four things are so powerful is because the repository is where the entire team passes by every day. If 1% of champions implant knowledge here, the remaining 90% automatically benefit without having to learn anything. This does not increase individual skills, but has the effect of raising the basic level (floor) of the entire team.

What's interesting is that, rather than forcing the standards down, we let the champions find solutions that fit their own repositories, and similar teams naturally converged on the same pattern. Monorepo had a common context at the root and inherited detailed rules for each service, while web and mobile took a completely different approach.

5. 'Delegation' planted at the beginning of work and poor sprints

Once the repository was ready, the next problem was that delegation was annoying. Because the champions were constantly watching the agent from the sidelines, and the rest of them didn't even start.

So what this team focused on were the three ways work comes in for engineers.

• Issue tracker such as Jira
• GitHub Issue
• Slack

At all three entrances, we've made it possible for you to hand off work to your agents right on the spot, without having to learn new tools. For example, it is said that the process of finding a bug in Slack and creating a PR was done in 5 minutes.

"This is a story that really happened. One day, an engineer found a bug in a product in Slack and asked, 'Hey, has anyone seen this bug?' Engineer 2 responded, 'No, I haven't seen it.' At that time, Engineer 3 said right away in Slack, 'Hey, Goose, have you seen this bug before? Can you confirm if this is a bug?' Then Goose downloaded the file from the repository and said, 'Yes, this is a bug. It's here.' But there were three options as solutions. I will present it.' Engineer 1 said, 'Option 1 is good,' and Engineer 2 agreed, 'Goose, please implement option 1.' Goose completed the implementation and returned the PR link. The entire cycle of discussion, diagnosis, issue creation, agreement, and correction was completed in 5 minutes."

As this method becomes the standard, agents join Sprint. And in the first sprint, the team ran out of work, so tickets were pulled twice more. At first glance, it may seem like a story about increased productivity, but it refers to the rate at which the number of people needed to do the same job is reduced.

6. Move bottlenecks and build company-wide maps

As a result, amazing numbers were achieved after only 3 months of the AI Champions Program.

• AI-written code 69% increase
• Time savings experienced by engineers 37% increase
• Automatically created PR (Pull Request) 21x increase

Now we come to level 4 parallelization. It is said that in organizations where delegation is possible, increasing the number of agents is almost free. Instead, the bottleneck moved somewhere else. When one person started pumping out three or four times as many PRs, this time code reviews were pushed out.

The solution was interesting. We turned on code reviewers for the entire repository and created an auto-fix loop where when a reviewer points out an issue, another agent automatically fixes it and commits it. By the time a person opens the PR, most of the points have already been sorted out.

What's interesting is that we didn't force this AI reviewer to be turned on in the beginning. In the beginning, the quality of reviews was so poor that engineers were dissatisfied. I only turned it on once the repository was ready and the model was good enough. This shows that the order was important.

The second bottleneck was hardware. As I was running four or five agents at a time, my laptop's memory and CPU could not withstand it. So we invested in providing each agent with an isolated cloud workspace, which can now run in parallel from anywhere.

In the final step 5, as the number of agents increased, we created an orchestrator to direct them. The name is BuilderBot. And the final piece we needed to move toward true autonomy was company-wide guidance. We analyzed all 25,000 repositories and created a machine-readable list of which services are where and what they depend on.

Now, multiple agents can explore different parts of the system in parallel, and an orchestrator can piece them together to create an execution plan that spans multiple code bases. What has become possible at this point is that anyone in the company, even if they are not an engineer or do not have a GitHub account, can call Builderbot on Slack to fix bugs and create features. In the speaker's words, it was 'like a dream.'

"This felt like a dream... until it became a nightmare."

7. Shocking ending and a sign of the AI era

The presentation ends here. Leaving only three questions without answers.

"I had a lot of questions. Was this my fault? Did I enable my employees to do the most amazing things of their careers that ultimately led to their dismissal? Just the other day I was so proud. I was completely amazed at the way we were working and felt quite accomplished that we had successfully built an autonomous engineering organization. But for what purpose? I want to end with a few questions: What are we doing? Where are we headed? And is that where we're really going to get? Is this the right place for you?"

This journey can be condensed as follows. An organization that had no impact even though 90% of people used AI changed the repository to 50 AI champions instead of training everyone, planted delegation at the beginning of work, and created a map for the entire company to reach the 5th level of autonomy. And a few months later, that very headline came out.

This announcement is not simply a success story, but clearly shows the destructive consequences of technological advancement. This is a story we all know from textbooks. Since the Industrial Revolution, technology has increased productivity, and increased productivity means producing more output with fewer people, and the purpose of business is profit, not employment. The 'sprint where we had to pull tickets twice because we ran out of work' was a scene where this proposition became reality.

Three signs are now building up that this isn't just Block's story.

1. Anthropic CEO Dario Amodei's prediction: As of 2026 In 2025, Dario Amodei said in a Financial Times interview that half of entry-level office jobs could be automated within the next five years. At the time, it seemed like an exaggeration, but looking at Block's case, it reads differently. To the eyes of insiders using AI models indefinitely, it may have been just an observation rather than a prophecy.
2. Falling price of AI models: The objection that high-performance AI models are expensive so not everyone can use them is unlikely to last long. The token price of Opus 4.1, released in 2025, was $75 for input/output total, but as of 2026, the top model, Claude 3.5 Sonnet, is $10/$50. Even though it is the top model, it is cheaper than the previous generation's launch price, and the price of Opus 4.5 has been reduced to $5/$25. Performance is increasing, but unit prices are continuing to fall.
3. OpenAI's Seoul Recruitment Notice: It's not far across the ocean. Currently, job openings in Seoul are posted on the OpenAI recruitment page. The job is called 'Forward Deployed Engineer'. If the first sentence of the announcement is translated into Korean, this person is "a person who leads the entire process of deploying the latest model to an actual operating system by working with the most strategic customers." It's not just a demo, we're responsible for everything from identifying production requirements to designing, building, and deploying. The criteria for success are also specified in the announcement. This is Measurable Work Impact. Their job is to push AI into the bottlenecks of Korean companies, and another name for that impact is the 'efficiency' we saw at Block.

8. Conclusion: Prepare, not fear

My reason for telling this story is not to scare you. What Block's track record shows is that AI-driven transformation is not just a vague idea, but **a process that already works.

Then, the only option left to us is to face this transition not from the 'side' but from the 'knowing side' and the 'preparing side'. The beginning is not grand. We need to start by knowing our current position and the position of our team through the **6-step AI maturity diagnosis table** presented above. What stage is your team at now and how do you see the trend of this AI era? 🚀