Progressive Delegation

Everybody wants the same thing from AI: “I want it to just do this for me.”

And then they stare at the process they want to automate and freeze. Either they don’t know where to start, or they try to build the fully automated version on day one and hit a wall of complexity that kills the project before it produces anything useful.

The pattern is the same every time. Someone looks at the end state, the AI handling the whole workflow, and tries to figure out how to get there in one move. They skip straight to the destination without building the road.

That’s the mistake. You don’t automate a process. You earn automation by progressively removing yourself from it.

Think about how self-driving cars actually happened.

Nobody went from “human drives everything” to “the car handles it all” in one step. It was a progression. First, the car assists: lane keeping, adaptive cruise control. The human is still driving, still responsible for every decision. Then the car handles more, but the human needs to be ready to take over. Then eyes off the road in certain conditions. Then, eventually, maybe, full autonomy.

Each level was earned by proving the previous one worked reliably enough to trust the next one.

AI operations work the same way. There are stages, and each one builds on the one before it. We call this progressive delegation.

Stage one: you’re the expert, AI does the work.

This is where most useful AI work actually starts. You’re doing the thinking. AI is doing the execution. And you’re paying close attention to what comes back.

The first thing we ask teams to do is write out the process they want to automate. Every step, every decision, every input. What happens next depends on the outcome here.

Almost always, the process has way more steps than anyone assumed. People carry a simplified version in their heads, and when they write it all down, the real complexity surfaces. That moment of overwhelm is actually good. It means you’re seeing the process clearly for the first time, which means you can make smart decisions about where AI fits and where it doesn’t.

From there, you build an easy first version. Not the full thing. The simplest version of the workflow with AI handling the parts it can handle and you reviewing everything that matters.

This is the stage where you’re the expert in the loop. You’re not just checking boxes. You’re accountable to the quality of the output. And every time you correct something, you’re not just fixing that one output. You’re improving the instructions so the next run is better.

Here’s the part people miss: those corrections need to go into the playbook, not just into conversation. If you’re correcting AI and expecting it to rely on its memory, you’re not building anything. You’re just having the same conversation over and over. The feedback loop only works when your corrections become better instructions.

Stage two: you’re checking guardrails, not doing the thinking.

At some point, you notice something shift. You’re reviewing outputs and mostly approving them. The corrections get smaller and less frequent. The playbook is doing the heavy lifting.

This is when you move from expert in the loop to something lighter. You’re still reviewing, but now you’re checking against guardrails, making sure the AI stayed within the standards you already set, rather than actively shaping every output.

The difference matters. In stage one, you’re making it better. In stage two, you’re making sure it didn’t go off the rails. One is quality accountability. The other is a safety check. Knowing which one you’re in tells you whether the playbook is ready for the next step or still needs your expertise.

We went through this with our own Bootcamp grading process. We give personalized feedback to learners on the playbooks they’re developing, drawing from everything we know about them during the course plus a library of playbooks we’ve built over years. At first, we wrote every piece of feedback by hand. We noticed patterns, thought about how we’d teach a smart intern to do this, and started building a playbook for it.

The first version had multiple expert review steps where our team needed to weigh in, override AI output, and keep tuning the instructions. As it got better, we started removing those review steps one by one. Eventually, we were only evaluating the end result. Now, when new playbooks are submitted, AI produces in-depth feedback and our team reviews before sending to the learner. Each week, the team logs every issue they fixed and we triage and address them.

The next step would be setting this up to be fully automatic. We might get there. We might not.

You might not reach full automation. That’s fine.

This is the part that trips people up. They think the goal is always stage three: AI runs everything, no human involved. And so they judge every workflow by whether it got there.

But progressive delegation works even if you stop at stage two. The value isn’t only in the final stage. It’s in each stage along the way. A workflow where AI does the work and a human does a quick guardrail check is already enormously valuable. You’ve freed up expert time, you’ve documented the process, and you’ve built something that gets better every week.

Self-driving cars haven’t reached full autonomy in every condition. That doesn’t mean adaptive cruise control isn’t useful. The value compounds at every stage.

The real trap is trying to skip to the end. Teams that jump straight to full automation before they’ve earned it end up in workaround hell, fighting complexity they don’t understand because they never went through the stages where they would have understood it. Teams that earn each stage build something that actually holds up.

This is an Automation protocol. It changes how you build AI into your workflows.

The question it teaches you to ask: “What stage is this workflow in, and what does it need to earn the next one?”

Know someone dealing with this? If you’ve got a colleague who keeps trying to fully automate things and hitting walls, or someone who’s been “using AI” for months but never built anything that runs on its own, send them this. The path isn’t a single leap. It’s a progression.

Read next: Protocol: Own the Playbook, Rent the Tech

Figure out your path: Who Is This For?

Stay connected:

I lead a team and want to build an AI-capable organization. [Subscribe to Leading Momentum]

I’m operationally minded and want to build AI operations skills. [Subscribe to the Operator newsletter]