Someone on a tactical gear maker’s team sent in a question before an AI lunch-and-learn I gave for their company: “What am I doing manually that AI could do, besides email and formatting?”
It’s the hardest kind of question to answer about yourself. Repetitive work is invisible to the person doing it. It’s automatic, and that’s exactly what makes it repetitive.
So instead of handing them a list, I told them about my YouTube problem.
I don’t read books on AI anymore. By the time one is printed, it’s out of date. So I learn from YouTube. Most of the videos worth watching run about thirty minutes, and I only need the cliff notes.
Ten to twenty videos a day. Thirty minutes each.
That’s five to ten hours.
That’s a full-time job, some days more. I already have one.
So I built an agent that runs every night. It grabs the new videos from the channels I follow, pulls each transcript, and sends me a summary of five to seven bullets per video. If one deserves more of my time, I go watch it. Most don’t.
Then I took it one step further. I’m an auditory learner, so the agent turns each summary into audio with OpenAI’s text-to-speech and pushes it to my iPhone as a private podcast. I listen on the drive to work. I’m the only subscriber.
I should say how unlikely that is. I studied computer engineering, and every time I sat down to write code, I fell asleep. I never became a programmer. I have developers for that. AI let me build something I didn’t have the skill, or the time, to build before.
Then it went off the rails
The summaries are written by an AI model, and the script that made the audio expected a divider line between each video. One day the model stopped writing the dividers.
The script read the entire day as one video. Instead of ten short episodes, it produced a single episode around forty minutes long, titled after the summary file instead of anything in it.
I got tired of it being wrong. I turned the audio off for about four months and went back to reading the summaries.
That sounds like a failure. It’s the best argument I have for picking the right first project. When the podcast broke, nothing bad happened. A badly titled episode costs nothing. No customer saw it, and no money moved. I went back to reading summaries, which still beat five to ten hours of watching.
And the job never stopped being a machine’s job. The tools did. In my experience they’ve gotten much smarter, and the small things I used to tinker with come up far less often. A few days before that lunch-and-learn, I turned the podcast back on. The fix was to split the summary on its headings instead of trusting the dividers. It’s been back on since.
Three tests for your own week
You don’t want to automate everything. As I told that room, automate what takes the most of your time and what you do over and over. Someone said, “Like an action I run every morning.” Exactly.
Three tests, in order:
- It happens more than once a week. Once a quarter isn’t worth automating.
- It has a right answer. Copying numbers from one system into another has a right answer. Deciding whether to drop a vendor doesn’t.
- The embarrassment test. If you’d be a little embarrassed describing the task to a friend, because it sounds like something a computer should do, a computer should do it. “I spend five to ten hours a day watching YouTube to keep up” passes easily.
If nothing comes to mind, that’s the invisibility problem, not proof there’s nothing there. Start with these:
- Copying data from one system into another by hand
- Downloading the same attachment and filing it in the same place
- Reading a long document to find one number
- Rebuilding the same report or chart every week
- Writing the same message with three details changed
- Checking one list against another
- Checking whether something changed since last time
What to hand off, and what to keep
Safe to hand off: lookups, transcription, reformatting, checking a list against a list, first drafts of anything repetitive.
Hand off with a human check: anything that touches a customer, anything with a dollar amount, anything that becomes a commitment to someone outside the company.
Keep: the decision, the relationship, and the thing you’re accountable for.
Everything in the first list is looking, copying and checking. Everything in the last is deciding.
That’s why my podcast was a good first project. It sits entirely in the first list. When it broke, the cost was a badly titled episode. Start where a mistake is cheap, and move toward the second list once you trust what you built.
The question to take back to your desk
At the end of this week, look back and ask one question: what did I do this week that a machine should be doing?
Write down the first answer, even if it seems too small to bother with. Mine was watching YouTube.
Nobody hired me to watch YouTube for hours every day.
I hired myself, and never noticed.
