Once a week, I take an electrician's class. It's a fallback in case AI obliterates white-collar work. An 18-year-old whiz kid sits next to me. Last Friday, attempting to stay current with the youth, I asked him what he watches for entertainment. "YouTube," he said. No surprise. I followed up, "Do you know about NBC or ABC?" He replied, "I know they're news stations, but I don't watch them."
He relegated the 20th century's most powerful media networks to "news stations." ABC and NBC are where you go for local traffic and weather.
When I was his age, I joined tens of millions every Thursday to watch Rachel and Ross endlessly flirt on Friends and half-naked strangers compete for a million dollars in Borneo's jungle on Survivor. We watched Ellen come out and Will and Grace normalize gay romance. Networks that entertained the entire country and assisted meaningful change in social causes are now "news stations."
To be fair, I cannot name a single show on network television. I haven't had cable since 2012. I consume plenty of media (too much). Like the whiz kid, it's primarily on YouTube. I watch Doug Demuro's insufferable car reviews, Shohei Ohtani highlights, old Top Gear episodes, and Marco Pierre White.
YouTube's most popular creator, MrBeast, recently uploaded a video entitled "Hi me in 10 years", which he had recorded and scheduled for release 10 years ago. In it he's an awkward high school kid with a shaky camera and bad lighting hoping to have a million subscribers by 2025. Today, he has 435 million subscribers, his most popular video has 889 million views, and his channel's production quality rivals professional studios. For comparison, the most watched television show in U.S. history was the final episode of M*A*S*H, which drew an estimated 105 million viewers.
YouTube made content distribution easy and free. The iPhone made it good. Inside everyone's pocket is a camera capable of studio-quality film. A tripod, an iPhone, and a Rode mic are enough equipment to produce professional-grade content.
AI will similarly disrupt the software industry. Complex enterprise applications typically developed by large firms for enormous amounts of money over long development cycles will be developed by small teams and even individuals for a fraction of the cost and time.
We're now learning the skills required to make this work — and it does require real work, as with content distribution: Not everyone gets to be a YouTube star. You must find an open niche to appeal to an unserved audience; develop and edit compelling content; and release new content at a reliable cadence.
The same goes for software. Small teams and individuals can either develop custom software to solve problems for which no solution exists, or they can compete against larger firms by selling similar software at lower cost.
Two emerging skills will be necessary. The first is a sort of AI system architect. This person is capable of developing complex end-to-end workflows — from ideation to deployment — by developing a series of integrated instructions for AI to follow. In other words, it's a sort of value stream for a suite of AI agents with standardized work embedded in each stage of the process.
Currently, AI fails when it lacks clear instruction and when its "context" becomes diluted by too much information or polluted by conflicting information. The system architect will be responsible for delineating where one stage of work ends and another begins, and what the precise output of each stage of work should be. This enables clean handoffs between agents with high-quality context so work can continue with quality.
Ironically, AI may help us to get better at developing standardized work. Clear instructions built for feedback loops are the ticket to success with AI. While your people are wading through half-baked job instruction buried in nested folders on SharePoint, your AI agents will have exceptional standardized work. Humans are capable of dealing with uncertainty, while machines are still figuring it out. It reminds me of an ATM factory I observed where robots worked alongside people in assembly. Because of the robot's vision system's limitations, engineers had to precisely design part placement. It had the most perfect 5S I had ever seen. Meanwhile, the human adjacent to the robot had a sloppy workstation and could have used the same attention to workplace organization.
The system architect will also be adept at creating specialized agents with domain-specific skills. Rather than asking Claude or ChatGPT to build an application, the architect will design a host of agents to serve distinct parts of development. For instance, in the ideation stage, a market research agent trained to learn about competitors and customer frustrations will help shape the early product, while a quality agent will develop testing systems to ensure the developer agent's code works as intended.
The system architect's work will be used by another new role: AI agent manager. It's odd to think, but people will manage AI agents. In fact, in my spare time, I manage a team of 10. At times, it can be as rewarding and frustrating as managing people.
The AI manager's purpose will be twofold. First, they must follow the process developed by the system architect. This is critical because when the process fails, you need a reliable PDCA-loop to improve the system. As you correct system failures, the overall process becomes more reliable. And, as explained earlier, AI tends to go haywire without crystal clear instruction, which the system provides. By following the process, you have a much greater chance at success. Second, the AI manager will be the "human in the loop" to review the agents' work before committing it and pressing forward.
What does this mean for companies?
More are encouraging employees to vibe code. Anyone can pop into Claude, ChatGPT, or Gemini and build a nifty front-end application, even ones that solve annoying work headaches. But I'd encourage companies to push much further. Find a project that would normally take a multi-skilled team of five or 10 and see if you can accomplish it with two people by developing an AI-first development architecture and managing agents through it. Pick a challenging project — something that would normally take weeks or months.
It begs the question: if the team of two is successful, what happens to the rest?
The conversation around AI has been far too focused on cost savings through headcount reduction. Companies that layoff people by replacing them with AI will face consequences down the road. The big opportunity is not bottom-line cost reduction but top-line revenue growth.
Companies that invest in developing the system architecture and train employees on managing AI agents through it will be able to create enormous amounts of new value much faster. There are people in your companies with compelling ideas without the skills to realize them. These tools will allow them to contribute and add new, additional value.
Just like YouTube degenerated NBC and ABC to "local news" for everyone under 30, AI will render five and six-figure SaaS obsolete. Don't get me wrong. Companies like SAP and Salesforce aren't going anywhere (yet). But software companies offering niche solutions at exorbitant prices will face headwinds.
I am so compelled by these ideas that I have decided to pursue them full-time. After 10 years of rewarding work alongside amazing people at the Lean Enterprise Institute, I will depart to see if any of this has merit. Over the past year, I have stared at my screen slack-jawed too many times. Something incredible is happening. And I would like to see what's possible.
If it doesn't work out, I can always be an electrician.
Matthew Savas is a lean practitioner exploring the intersection of AI and lean thinking. He spent 10 years at the Lean Enterprise Institute before departing to pursue AI-enabled software development.
Get articles like this delivered to your inbox. Join technology leaders learning to apply lean thinking.