Key Takeaways

  • Cloudflare CEO Matthew Prince is building for an exponential AI future by betting on raw, young talent; they hired 1,111 interns last summer who are naturally fluent in AI tools.
  • To make AI practical across the company, Cloudflare launched an internal initiative called "Cloudflare OS," designed to give finance, legal, and sales teams access to custom AI tools.
  • They seeded this system with a clever trick: a "magic AI model" email address that employees thought was a cutting-edge AI, but was initially staffed by a hidden human team.
  • This human team covertly collected critical "jobs to be done" data, like how a sales rep does an account plan, by observing employee requests and then using AI to refine responses.
  • The Cloudflare's 'Magic AI Model' for Internal Process Capture framework is how they harvested real-world workflow data to build practical AI tools that drastically boost productivity.

The Cloudflare's 'Magic AI Model' for Internal Process Capture

Prince saw the AI wave coming and understood that generic tools wouldn't cut it. To truly embed AI into Cloudflare's daily operations, they needed to understand exactly how their own people worked. Their solution was a covert data collection strategy masked as an AI assistant.

1. Establish 'Magic AI' Email Address: Set up an internal email address and initially tell the team that it's a magic AI model that can help with job tasks.

2. Collect 'Jobs to Be Done': Team members write to the email address asking how to get specific parts of their job done. The 'AI' sometimes asks more questions and provides a response.

3. Human-Powered Backend for Data Collection: A hidden team of humans initially receives these emails, using AI systems to flesh out responses, while critically recording all key 'jobs to be done' within the organization.

4. Seed Internal AI Systems: Use the collected 'jobs to be done' data to seed internal AI platforms (like Cloudflare OS) with the ability to automate or assist these specific tasks (e.g., a '/account plan' command for sales).

Matthew Prince explains the initial deception: “We actually set up this email address. Uh and initially, we told the team that it was this magic AI model. ... What what we didn't tell people was actually it was a team of team of humans behind the scenes.” This allowed them to capture genuine, real-world problems without the bias of formal surveys or theoretical discussions. For Prince, the goal was to build systems that deliver real results: “If you're on the sales team and you have to do like an account plan, you know, you can literally just type {slash} account plan and then describe what it is that you want to do and it will output that. And that's made that's made our team so much more productive.”

When This Works (and When It Doesn't)

This method shines when an organization wants to quickly and organically capture real internal workflows and 'jobs to be done.' By simulating an AI interface, it provides the raw material and user context needed to train and deploy practical AI tools across the business efficiently. It works best in environments where employees are curious and willing to experiment with new tools, and where there's a baseline of trust in leadership's intentions. The approach is less effective, and potentially risky, if employees feel tricked or manipulated once the human element is revealed. It also requires a committed team ready to analyze the collected data and then actually build the AI solutions, rather than just collecting requests.

What to Do With This

As a 27-year-old founder, you can adapt Cloudflare's trick to embed AI into your own small team. This week, create a dedicated email, say ai.internal@yourcompany.com, and announce it as an experimental AI assistant for drafting specific documents or messages. When your marketing team asks it to "write a social media post for our new product launch," you (or a trusted ops person) secretly generate the response using a tool like ChatGPT, then critically log both the request and the successful output. Once you have 30-50 such examples, you'll have a clear dataset of internal "jobs to be done" – a goldmine for building custom AI prompts or even simple internal commands that boost your team's specific productivity, just like Cloudflare's /account plan.