Rules of Automation

So much of the summer and fall thus far has been thinking about automating my business.

In a strange way, it’s work that reminds me of my previous career at Apple as an automation engineer. Working on AI workflows isn’t that much different from automating UI tests for software. 

It got me thinking, what are some working rules of Automating anything, especially with AI?

1. Lock Design Before Building

It’s hard to automate tests if the design of the new process, feature or app isn’t locked in. You can’t get reliable tests because literally everything is changing underneath the automation. It could be a bug, a removed feature, or a reorganized set of steps that breaks. 

So the rule to lock in the design was meant to remove variables and only automate tests that were actionable. But locking the design is only one part.

2. Follow the human. 

The best automations follow concrete steps taken by a human. This is the clearest way to first discover what really happens during a process.

The first big mistake most people make is to jump into automating (or digitizing) parts of their process, without understanding each step. The amount of information your brain fills in automatically is more than you think. 
If you've ever tried to teach someone how to make a peanut butter and jelly sandwich by listing the exact steps, you'll discover how hard it is to clearly get each movement. So think through the steps for your self, and compare them to the following:

Get bread -> take two slices -> apply peanut butter to one side of bread -> apply jelly to other side of other bread -> combine -> profit.

What could go wrong, a short list:

1. There is no butter knife to spread the peanut butter and jelly. Some people will use their hands like savages.

2. Because each slice of bread has two sides, there is no note of the four sides we could apply PB or J to.  We assume people will know the fillings go on the inside, so the combine step could get messy.
3. We could be using olive bread, cinnamon bread, rye bread, because the type of bread is not noted. Maybe you use jalapeño jelly instead of grape. The risk: a terrible tasting sandwich because the ingredients are not well defined.

Mapping the human steps is just the start to great automation. Once these steps have been marked, we can start to test what make the process more timesaving and efficient.

3. Discrete and Repeatable.

Many people struggle with delegation. The big issue here is there are two types of actions that can be automated: discrete and non-discrete. 

Discrete actions are concrete and use little judgement to execute. Open the music app, click on the only search bar on the screen, search for 'Thee Oh Sees', click on the 'Face Stabber' album, and hit play by the song 'The Daily Heavy'.  This will start the song (which sounds like a rhythmic squeaky toy until the drums hit). This set of discrete steps will get that specific song to play each time. 

Non-discrete actions rely on judgement and thinking. The same workflow, if I tell you to open your favorite music app, search for 'Thee Oh Sees', and play the first album you see, I have no idea which song you will be playing. You will need to listen to it to understand if it's the right song, or a song that fits the current situation.

Another way to explain the idea is discrete actions have one way to do the thing, and one predictable outcome.  Predictable steps and outcomes lead to reliability.  Unpredictable outcomes cause mistrust with the process.

Which is important because reliable automation saves you time. Discrete actions are the low hanging fruit for automating a process because they save you the most time, and often require the least amount of maintenance. 

4. Automation Includes People.

Automation is never 100% autonomous, or free of people. 

Great automation reduces the need for human handholding during a process.

When automation has the right human oversight, and is reliable, then you can trust your process. This is great because outside of maintenance you won't spend time working on the tools that make automation work.

Poor automation takes up more time to maintain than time saves.

Usually, figuring out why the automated process is busted is THE sign the automation is missing a key component. 

Either the steps were never figured out (lock the design follow a human), you are trying to automate the creative (non-discrete) actions, or you are at the tail end of underlying change (Claude changes every day, the app you want to automate has a fluid set of steps and UI).

The golden rule here: Are you saving time, or adding work?  Automation should always be saving time.

The best way to make sure automation saves time is to strategically design where people come into the process. This itself is creative, and where people can stand out (cue movie voice-over voice.. ahem) "in a world where AI is everywhere."
In a world where AI is everywhere, the goal is to automate the tedious parts, and leave the design and creative parts to humans. 

Where does this leave us?

I'll leave you with one more quick point. 

These four rules, when followed, create automation that is fun and effective. The feeling of the time consuming things you don't like doing, getting done in a blink of an eye, is awe-inspiring. It's the dragon I chase in my own business, every day. 
How does this relate to AI? It's easier to violate all of these rules, and get frustrated with the tool.  AI can predict what you might want, but won't know until you are angry with it. AI wants to take over the creative parts of your work and life, even if those are the points that keep the work fun and keep the process working. 

Ultimately, if you automate things with AI, make sure it's saving you time, and not tricking you into feeling you are productive while slowing you down.

Otherwise, you are just adding more work to your plate.