Steer it first, then let it run

By Devan Borsboom

Automating too soon makes a bigger mess

People want the AI running on its own before they have worked out how the task should be done, or what a good result looks like. Then it runs poorly, and now the mess happens automatically, without anyone pressing the button.

The order I follow

1. Do the task with AI while you steer it.
2. Fix the instructions every time it gets something wrong.
3. Once it gets it right consistently, let it run on its own.

If you skip the steering, you end up cleaning up its mess instead of guiding it along the way.

Treat it like a new employee

In the beginning you watch over its shoulder. Once it has gone a while without a mistake, you let it work and review the result when it is done. That is the training-wheels approach, and it is the same one you would use with a new hire.

For higher-stakes work, check its work for a couple of weeks. You can also break the task into small pieces. Instead of asking AI to make the decision, have it find the date of each person's last visit, then find that date on the record. Each small step is easy to check, so the risk stays low.

Checking is far cheaper than doing

A process starts at maybe 80% right, then 90%, then slowly more, as specific feedback goes into the instructions. It is still important to have a person check the work. Checking a result takes a small fraction of the time it takes to do it.

People stay at the start and the end. You set the goal and describe what good looks like. AI does the middle. You review what comes out.

There is no shortcut

I spent tens of hours giving feedback on how AI handled my inbox before it got to the point of asking me only about the rare edge cases. I am not aware of a shortcut. The time goes into teaching it what good looks like for your practice.

Steer it first. Automate it once it is dialed in.