Eric Lubow

You can only push decisions as far down as context travels. Orgs want decisions made at lower levels, but keep the information needed to make them concentrated among Sr leaders. Distributing authority while centralizing context isn't delegation. It's escalation with extra steps

The SDLC wasn't just how Engineering built software. It's how security reviews, access controls, and operational experience reached what was built. AI didn't necessarily make those controls worse. It made it possible to build software without ever passing through them.

Engineering Judgment Includes Knowing When Not to Fix Things An engineer can look at a prototype & identify 10 things wrong with it & they might be right about all 10 But if the prototype exists to answer a product question, fixing those things first might be the wrong decision

AI coding agents can write a Makefile in seconds. That doesn't mean it'll recognize that repeatedly running the same commands is a reason to create one A lot of engineering experience isn't knowing how to implement something. It's recognizing when a particular practice is needed

Psychological safety isn't established by policy. Every time someone surfaces a questionable decision they've already made, your reaction teaches them whether to raise the next one or keep quiet. People learn what's safe from how you respond, not what you wrote down.

The more I work with AI, the less I think the job is replacing human thinking. The job is deciding where thinking, synthesis, and insight should happen now that humans no longer participate in every workflow. That’s a design problem, not a model problem.

Giving people AI coding agents is the easy part. At Mapp, we've been building the engineering foundation that lets people across the company create applications using our existing infrastructure & capabilities What we built, what worked, & the tradeoffs: eric.lubow.org/2026/mapp-lab…

The most useful thing about a meeting transcript might not be the summary. “What did I miss?” is useful “What does this meeting change about a decision being made elsewhere?” is more interesting The value is connecting what happened in 1 room to decisions being made in another

Some people think with writing. Some think with building. AI lets you do both. Write to develop the idea. Prototype to test it. Write again to capture what you learned. The PRD after the prototype is better because it captures knowledge, not speculation.

Automation of workflows with AI forces an interesting question: why did this workflow exist? The obvious answer is to move work. The less obvious one is that doing the work was also how the organization learned. Automating the first means figuring out what happens to the second

AI gave people across the company the ability to build software without Eng. But the context needed to do it responsibly often stayed behind. This is a recurring AI operating-model: giving someone the ability to act is easier than giving them the context to act responsibly.

Engineering judgment is scar tissue, and it's portable It's the accumulated context of incidents at 2am on a Sat. It's knowing what to monitor, how failure cascades, & where trade-offs live. The judgment was never about writing code. Judgement still applies when AI writes code.

Most delegation advice focuses on permission. Can people decide? Will their manager override them? But authority isn't the same as context. Sometimes we're asking people to make broad decisions with a narrow view.