When should you not automate something?

Do not automate something when it is still changing, when it happens rarely, when it depends on relationships, when you do not understand it yet, or when the cost of a mistake is high and hard to reverse. In each of these cases automation locks in a bad process, removes a human who was adding value, or creates risk that a person would have caught.

Why does this question matter?

Because automation is fashionable and founders feel behind. The pressure is to automate everything, quickly. But a system built on the wrong process makes the wrong thing happen faster and more consistently, and it is harder to undo than a habit.

The founders who get the most from AI are not the ones who automate the most. They are the ones who choose well.

When is it too early to automate?

It is too early when the process is still changing. If you have done something fewer than ten times, or if the way you do it changed last month, you do not yet know what the system should do. Automating now means rebuilding it in six weeks.

Do it by hand a few more times. Write down what you did each time. When three consecutive rounds look the same, it is ready.

When does the volume not justify it?

Something that happens once a quarter and takes an hour is four hours a year. A system to handle it might take a day to build and need maintenance. Leave it alone and put the effort into the things that happen every day.

The rough rule: automate what happens weekly or more often. Write a checklist for what happens monthly. Just do what happens quarterly.

When should a human stay in the loop?

Whenever the task is really a relationship. A client renewal conversation, a difficult piece of feedback, a first reply to a high value lead, a complaint. AI can prepare all of these, and it should. But the moment itself should have a person in it, because the client can tell the difference, and the difference is what they are paying for.

Automate the preparation, the reminders and the follow up. Keep the conversation.

When is the risk too high?

When a mistake is expensive and hard to reverse. Sending money. Deleting data. Publishing to a large audience. Making a commitment on the business's behalf. These can be automated, but only with an approval step, a log and a way to undo. If you cannot build those three things, do not automate it yet.

A useful question: if this system did the wrong thing at 2am on a Saturday, how bad would Monday be?

What is the test to run before building anything?

Ask five questions.

1. Has this process been stable for at least a month?
2. Does it happen at least weekly?
3. Can I write the rules on one page?
4. Would a mistake be cheap and easy to fix?
5. Is a human adding real value here, or just passing things along?

Four or five yes answers and it is a good candidate. Two or fewer and it is not, at least not yet.

What about automating just part of a task?

Often the best answer. A client renewal conversation should not be automated, but the reminder that it is due, the summary of the account and the draft of the opening message all can be. Split the task into preparation, action and follow up. Automate the first and third. Keep the person in the middle.

What if you already automated something you should not have?

Switch it off, keep the useful parts and put a person back in. Most over automation is fixed by adding an approval step or narrowing the trigger, not by scrapping the system. The reminder and the draft can stay. The sending goes back to a human.

What should you do with the things you do not automate?

Write them down, give them an owner and put them in the weekly rhythm. Not everything needs a system. Everything needs a place.

If you would like a second opinion on what to automate and what to leave, that is exactly what the first hour of a Strategy Intensive is for.