What should be a human decision, and what can be an AI decision?

As machines gain cognitive abilities we’ve never seen before, we’re asking this question all the time. We’ve started keeping a running list, and I fully expect the answers to change as AI gets better.

We believe organizations need to move quickly toward being AI native. To us, that means starting with AI as the default way of doing work rather than layering it onto how we’ve always worked. This makes the question of where humans need to stay involved even more important.

So far, I see three main buckets:

1. HUMAN GUIDES

AI can do a lot, but the human still needs to guide it.

This is probably most knowledge work today: research, writing, analysis, coding, brainstorming, planning and strategy.

I use AI for all of these constantly, but I never just give it a task and accept what comes back. I bring the underlying idea and context, then work back and forth: challenge assumptions, fill in what’s missing, follow interesting threads, and keep pushing until we get somewhere good.

I’ve called this collaborative interrogation. Without it, you can get slop pretty quickly. For now, humans still bring deep expertise and intuition that are invaluable in steering the work. In effect, you’re managing the AI, and ultimately you have to stand behind every word, number and conclusion it produces.

I expect this category to change dramatically. “Recursive self-improvement” is on the way. As AI gets better memory, context and safe access to our work, it should need less direction from us. Maybe eventually we really can just send it off. We’re not there yet.

2. HUMAN OWNS

AI may do most of the work, but a human still needs to own the decision.

Hiring or firing someone. Making a large financial commitment. Signing a contract. Making a clinical decision. Changing a production system. Sending a consequential communication.

And increasingly, deciding what AI itself can access and do. Can it access this database? Send this email? Change this record? Deploy this code? Move this money?

AI can help make those decisions, and increasingly execute them. But someone still needs to be responsible for the outcome.

3. KEEP HUMAN

AI may be capable of doing it, but sometimes the human is part of the value.

Comforting someone who’s scared. Supporting someone through illness.
Mentoring someone. Having a difficult conversation. Celebrating an accomplishment. Building trust. Being there for someone who’s grieving.

AI will probably become very good at these things, and may even help humans do them better. But sometimes the fact that another human chose to show up is part of the value.

I suspect we’ll spend the next decade constantly moving these boundaries. That’s why we’re keeping the list, and I’d love to hear what’s on yours.

The question won’t just be whether AI can do something. We’ll increasingly have to decide whether it should, and who owns the decision when it does.