AI
The Work I Refuse to Automate, and How I Decide
I am one of the most automation-happy people you will meet. My content distribution runs across 8 channels, and I operate the whole thing myself.
And yet I keep a list of tasks I deliberately do not hand to AI. Not because it is technically impossible. It is possible. I just refuse.
That is why content that says "automate everything" makes me uncomfortable. It is advice that sounds great in a keynote and gets expensive in real life.
I am not anti-automation
Let me state my position first. Making companies leaner is an old habit of mine, long before AI became a trend: $4M+ per year from various initiatives. Process improvements, cost optimization, more efficient product decisions. Automation was only one of them.
That kind of leverage used to require seniority, an engineering team, and expensive systems. Now AI is the sharpest tool for that same habit, and it can be taught to everyone on the team.
But a sharp tool also cuts deeper when you hold it wrong. That is the part people rarely discuss.
Good managers know what not to delegate
Think about how a manager assigns work.
Weekly reports, competitor research, deck drafts, cleaning up meeting notes. All of it gets delegated without much thought. If the output misses the mark, the manager reviews and asks for a revision.
But some tasks a good manager keeps. Firing someone. Talking to the partner whose contract decides next year's revenue. Deciding what the team will build and what gets stopped.
This has nothing to do with the team being less capable. A direct report might be better with data than the manager is.
The reason is one sentence. Execution can be delegated. Responsibility cannot.
Now map that onto AI. If you have ever said "the AI is useless," check the task first. A manager who hands a decision about someone's fate to an intern, then blames the intern, would get laughed at by the whole office. If you give that kind of task to AI and then blame the AI, it is the same story. The one who does not understand the limits of delegation is you, not the model.
Four types of work I keep for myself
Here are my criteria. Not scientific. Just 20 years of leading product teams and building AI systems.
- Decisions about people. Hiring, delivering hard feedback, promoting or demoting. AI can help organize the data and the material. But the final sentence comes out of my mouth, because the person on the other side deserves to hear it from a human who owns the outcome.
- Negotiations and relationships. Negotiation is about reading the room. The pause before an answer, the shift in tone, the things deliberately left unsaid. AI can prepare scenarios and counterarguments. Sitting at the table is still my job.
- Decisions that are hard to reverse. When it goes wrong, the cost cannot be undone with one click. Here I always ask: if this is wrong and we only find out next month, who pays for it?
- Judgment about what "good" means. AI is great at chasing a target that has already been defined. Defining the target is my job. If I also hand over the definition of "good," the system will do the wrong thing with impressive efficiency.
That last point is not hypothetical. One of my posts once went live on the wrong brand account because an account-matching rule was inverted. The system ran smoothly. It was just pointed the wrong way.
My auto-reply system also once failed at scale for days because a login session had died. The system kept running, kept costing money, kept failing to send. No alarm went off. I only found out because I checked on a whim.
Both were my fault, not the AI's. I let go without supervision.
Three questions I ask before delegating to AI
Before I hand any task to AI, I ask three things.
First: if it goes wrong, who gets hurt? If the only cost is my time to redo it, let it go. If it hurts other people, trust, or money that does not come back, I keep it.
Second: can I check it quickly? I can review a draft report in minutes. A strategy decision whose impact shows up months later, I cannot. The harder it is to verify, the more certain I am to hold onto it.
Third: is there context that is not written down anywhere? Relationship history, verbal promises, the mood of the team this week. If that exists, the AI only gets half the picture. And half a picture delivered with confidence is the most dangerous thing there is.
One example from my own work. When my team built FINA, a customer support agent, we kept its cost at around $0.004 per turn. During testing, we found a cross-user data access vulnerability before it ever reached production. It was caught only because a human was still checking the part where a mistake would be most expensive.
Same pattern with support tickets. Automation went from 0 to 70% of roughly 10,000 tickets per month. The rest, the complex and sensitive ones, still went through people.
This is the pattern I keep using. AI does the heavy, repetitive part. I hold the part that requires responsibility.
And none of this is permanent. My list shifts with every new model release. Some tasks I used to keep are now delegated, after I watched the results stay consistent over time. The line moves. But it still exists.
So before this week ends, take five tasks you are thinking about automating. Run the three questions on each one. I am willing to bet you will pull at least one back.
If you want to learn to draw this line together with people figuring out the same thing, join AI Circle. It is where I share the workflows and thinking I use every day.
If you lead a team and need them to understand the limits of delegating to AI, check the corporate page.